You could investigate the wires (DNA) connected to each switch, and with a clamp meter, measure the current output (RNA). You might start picking up on patterns, like how the current spikes in different wires at different times, but you’d still just be observing rather than directing anything yourself.

With the invention of CRISPR, you can finally grab one specific switch and flip it fully off, hold it partway down, or jam it to full-on (a knockout, CRISPRi, or CRISPR activation). Instead of passively measuring the currents, you can finally force something to happen.

It’d take a long time to map out what every single switch does with CRISPR, but if you layer Perturb-seq on top of CRISPR, it’s as if your boss hired a whole fleet of technicians. With Perturb-seq activated, instead of one person testing one switch in one room, technicians like you now go to thousands of rooms at once, randomly assigned to force one specific switch (pooling). Afterward, inspectors go room to room and record which switch got forced, and what every fan in that room is doing now.

Over thousands of iterations, you’ll finally start being able to map out which switch(es) influence which outcomes.

The Limitations of Observational Studies

Traditionally, studying gene expression has been mostly observational. It relies on collecting cells, measuring the transcriptional output of their genes (RNA), and observing patterns. This work is valuable, but it cannot determine causation since you can’t directly manipulate anything.

For example, let’s say you observe that Gene X71 and Gene Y918 are both active in cancer cells. You might believe that Gene X71 activates Gene Y918, but it’s just as likely that Gene Y918 is actually activating Gene X71, or that there’s a third gene (a confounding variable) you’re unaware of that’s activating both of them.

How Perturb-seq Works

The word “perturb” means to disturb or change something, which is an apt description of what Perturb-seq does. It disturbs a gene and then measures the outcome.

Let’s say you’re studying Gene X71 and Gene Y918 (the genes you noticed were active in cancer cells). To test their relationship using Perturb-seq, researchers could shut down Gene X71 and see what happens to Gene Y918. If Gene Y918 changes, it’s much stronger evidence that Gene X71 influences it.

Here’s how that experiment works in more detail:

All Perturb-seq experiments start with CRISPR, which uses a molecule called a single-guide RNA (sgRNA). Single-guide RNA is like a key that physically guides the CRISPR protein towards the target section of DNA. From there, scientists can:

1. Disable the gene completely (CRISPR knockout).

2. Lower the gene’s activity without cutting the DNA (CRISPR interference (CRISPRi)).

3. Increase the gene’s activity (CRISPR activation).

CRISPR lets you manipulate one gene at a time, but a human cell has roughly 20,000 genes, and testing them one by one would take ages. To speed up the process, researchers use a technique called pooling.

With pooling, scientists create thousands of different guide RNAs and mix them into a single library. That library gets delivered to a large population of cells, tuned so each cell picks up just one guide. Across the full population, though, each guide ends up represented in many different cells, so instead of testing one gene once, you’re testing thousands of genes at once, with each one tested many times over.

To manage all this data, Perturb-seq gives each guide RNA something like a barcode, so results can be sorted by cell, noting which gene was targeted and how much every gene in that cell was expressed.

Researchers then compare cells that received a real guide against control cells that got a blank guide, evaluating how RNA differs from that baseline. If cells that received a Gene X71 guide consistently show the same RNA changes, that pattern starts to map the actual effects of Gene X71.

The Value of Single-Cell Data

Before single-cell sequencing, there was bulk sequencing. Bulk sequencing is the process of studying thousands or millions of cells together and then averaging the results. It still has modern research applications, but it doesn’t account for variations across cells.

Cells, even those from the same tissue, are not identical. At any given moment, some are dividing while others are at rest, and some are responding to a signal while others are ignoring it entirely. A perturbation might hit one group of cells hard and do almost nothing to another group sitting right next to it. In a bulk experiment, those two effects merge into one middling number that describes neither population.

Like bulk sequencing, Perturb-seq still involves large numbers of cells (often millions across an experiment). But it measures each cell individually rather than averaging them, which makes it possible to explore whether the same gene has different roles in different cell types.

This part is important and bears repeating. Perturb-seq doesn’t only ask, “What does this gene do?” It asks, “What does this gene do in this cell type, under these conditions, at this moment?” That’s a much more precise question to work from.

The Rapid Growth of Perturb-seq

The first Perturb-seq experiments happened in 2016, developed by several groups working independently (Dixit et al., Cell, 2016; Adamson et al., Cell, 2016; Jaitin et al., Science, 2016; Datlinger et al., Nature Methods, 2017). Dixit and colleagues, for example, targeted about two dozen transcription factors, while Adamson and colleagues used roughly 30,000 cells to study ten perturbations. These studies were small by today’s standards, but they proved that pooled CRISPR methods could produce highly detailed, single-cell data.

Since then, the scale of studies using pooled CRISPR has exploded. By 2022, Replogle and colleagues published an entire genome-scale Perturb-seq study covering more than 2.5 million human cells (Cell, 2022). Using CRISPRi, they studied nearly 10,000 genes in K562 leukemia cells and repeated the experiment in a non-cancerous cell line to compare results.

Think about the sheer scale of this work for a moment. Every one of those 2.5 million cells contributes its own full transcriptome, meaning the dataset holds billions of individual measurements. And with that much data, researchers can group genes into shared pathways and infer what poorly understood genes might be doing. Using this approach, Replogle and colleagues identified a previously unrecognized subunit of a protein complex involved in RNA processing.

The scale of these experiments can feel abstract, but the logic behind a discovery like that is pretty straightforward when broken down into an example.

Say Gene P22 has been studied for years and is already known to play a role in cancer. Gene Q83, by contrast, is barely studied, and nobody’s sure what it does. But you suspect it might be connected to Gene P22, so you knock down each gene separately using CRISPR and compare the results. If the RNA changes look nearly identical in both cases, that’s a strong clue the two genes are doing similar work, perhaps as part of the same pathway, or even the same protein complex. By tracking how Gene Q83 behaves relative to Gene P22, you developed compelling evidence for what the mysterious Gene Q83 does, too.

This is one of the most valuable things about Perturb-seq. Genetics has long over-emphasized a small set of famous genes while leaving countless others understudied. Perturb-seq gives researchers a way into those overlooked genes, simply by tracing their associations with genes we already understand.

How Perturb-seq Innovations Address Its Limitations

While the discovery of Perturb-seq was a great scientific advancement, it has its own set of challenges.

Compressed Perturb-seq

A standard Perturb-seq experiment requires many cells per perturbation, which gets expensive quickly as you scale it.

Yao and colleagues designed a way around this. Normally, each cell gets exactly one guide RNA, so any change in that cell’s RNA can be traced back to one clear cause. Yao’s approach breaks that rule by letting a single cell receive multiple guides at once, or spreading guides across cells sharing the same droplet (Nature Biotechnology, 2023).

Since most genes only affect a small, specific set of other genes (meaning RNA changes from different guides rarely overlap completely), there is a strong enough pattern for software to work backwards and determine which guide most likely caused which change. It’s a bit like recording a band together and using software to pull out each instrument. Using this method, Yao and colleagues were able to cut the cost of a Perturb-seq experiment by an order of magnitude.

In Vivo Perturb-Seq

Standard Perturb-seq experiments take place in a petri dish, which is useful for reasons like cost and the ability to control the environment. But cells behave differently in a petri dish than they do inside living tissue, where they’re receiving signals from neighboring cells, immune cells, blood vessels, and more. In vivo Perturb-seq solves for this by running the experiment inside a living host.

In one of these studies, Jin and colleagues used CRISPR to disrupt 35 genes linked to autism and developmental delay inside the brain of a developing mouse before birth. Then, they sequenced the cells after the mouse was born (Science, 2020). They found effects in neurons and glia, the brain’s other major supporting cell type. Interestingly, several of these 35 genes produced similar RNA changes when disrupted, suggesting those genes may have a shared pathway.

More recent work has pushed the same idea further into brain development at a larger scale (Zheng et al., Cell, 2024).

Multiome Perturb-seq

Standard Perturb-seq only measures RNA, which means it only accounts for what is transcribed from DNA. But some genes never get transcribed because they sit in a region of DNA that stays tightly packed, in a DNA-protein structure called chromatin, and can’t be read by the cell’s machinery. Standard Perturb-seq has no way of accounting for this.

Multiome versions of Perturb-seq solve for this by measuring RNA and chromatin accessibility together. By reading both layers at once, in a process called Perturb-ATAC (Rubin et al., Cell, 2019), researchers can see how DNA accessibility changes, and how gene expression changes with it.

Spatial Perturb-seq

To run standard single-cell sequencing, researchers first have to dissociate the tissue, meaning they physically break it apart so individual cells can be pulled out and measured one at a time. That process tells researchers exactly what each cell was, but once cells are separated, they’ve lost their location relative to other cells.

Spatial approaches solve for this by keeping the tissue intact. Instead of separating cells to read them, researchers tag perturbed cells with a visible label and image the tissue as a whole to preserve location data.

In one study, Dhainaut, Rose, and colleagues used this method, called Perturb-map, in a mouse lung cancer model, allowing them to see how disrupting a single gene changed not just the tumor cell itself, but the immune cells around it (Cell, 2022).

Spatial Perturb-seq has enormous potential for studying how genes shape the relationships between different cell types in disease, such as how a single gene inside a tumor cell can change the immune cells surrounding it.

Does Perturb-seq Prove Causation?

No, Perturb-seq gives strong evidence of a real relationship between genes, since scientists are directly intervening rather than just observing. But it isn’t perfect. Several things can undermine any individual result, for example:

• A guide RNA might fail to change its target gene, or it might accidentally hit the wrong stretch of DNA instead.

• Single-cell sequencing only captures a fraction of the RNA in each cell, so missing data is probable (this is one reason the Replogle team filtered out results backed by too few cells before making conclusions).

• A measurement taken early might capture a gene’s direct response, while a later measurement might instead reflect stress, adaptation, or cells dying off.

Perhaps the deepest limitation is that Perturb-seq can’t control for all potential indirect or mediated effects. If changing Gene A17 changes Gene D5, that still does not prove Gene A17 controls Gene D5 directly. Gene A17 might act through a mediator (Gene Z3), and Gene Z3 in turn changes Gene D5. Perturb-seq reads the endpoint of a chain without necessarily showing the links in between, so follow-up experiments are still required to validate the mechanism.

Millions of cells can make results feel more conclusive than they really are. Large-scale data is valuable, and it goes much further to establish a causal influence between variables than an observational study can, but even Perturb-seq cannot tell you whether an effect is direct.

Perturb-seq Offers A New Way Forward

With observational studies, you can observe patterns, asking questions like, “Are Gene P22 and Gene Q83 both active at the same time?” With CRISPR, you can start asking narrow causal questions, one at a time, such as “If I shut down Gene P22, does Gene Q83 change?”

While Perturb-seq isn’t perfect, it can change the kind of questions a scientist can ask. It allows you to combine CRISPR with pooling and single-cell readout to ask questions about a gene you know nothing about. For example, you could take the mysterious Gene Q83 and ask, “What does forcing Gene Q83 do across different cells, and does its behavior mimic that of any known genes?”

That’s what’s exciting about Perturb-seq. Rather than just testing a hypothesis you have with CRISPR, you’re generating hypotheses just by virtue of noticing patterns across perturbations. That’s a novel place for gene expression research to find itself, and it means researchers are starting to find things they didn’t even know to look for.

Sources

• Adamson, B. et al. (2016). A multiplexed single-cell CRISPR screening platform enables systematic dissection of the unfolded protein response. Cell.

• Datlinger, P. et al. (2017). Pooled CRISPR screening with single-cell transcriptome readout. Nature Methods.

• Dhainaut, M., Rose, S. A. et al. (2022). Spatial CRISPR genomics identifies regulators of the tumor microenvironment. Cell, 185(7), 1223-1239. doi:10.1016/j.cell.2022.02.015

• Dixit, A. et al. (2016). Perturb-Seq: Dissecting molecular circuits with scalable single-cell RNA profiling of pooled genetic screens. Cell.

• Jaitin, D. A. et al. (2016). Dissecting immune circuits by linking CRISPR-pooled screens with single-cell RNA-seq. Science.

• Jin, X. et al. (2020). In vivo Perturb-Seq reveals neuronal and glial abnormalities associated with autism risk genes. Science, 370(6520). doi:10.1126/science.aaz6063

• Replogle, J. M. et al. (2022). Mapping information-rich genotype-phenotype landscapes with genome-scale Perturb-seq. Cell, 185, 2559-2575. doi:10.1016/j.cell.2022.05.013

• Rubin, A. J. et al. (2019). Coupled single-cell CRISPR screening and epigenomic profiling reveals causal gene regulatory networks. Cell.

• Yao, D. et al. (2023). Scalable genetic screening for regulatory circuits using compressed Perturb-seq. Nature Biotechnology. doi:10.1038/s41587-023-01964-9

• Zheng, X. et al. (2024). Massively parallel in vivo Perturb-seq reveals cell-type-specific transcriptional networks in cortical development. Cell. doi:10.1016/j.cell.2024.04.050

Previous in the series: Generating More Than 1,000 Binder Backbones

Moving from protein backbones to amino-acid sequences

In the previous post I explained how after RFdiffusion3 filtering, we came up with 370 proposed binder backbones. But what exactly is a backbone? A backbone describes the overall shape of a protein, but it does not fully define the amino-acid sequence that should form that structure. In this next stage, we used ProteinMPNN to design those sequences. The goal was not only to produce strong candidates, but also to avoid ending with a final group of binders that all looked or interacted in the same way.

Designing only the binder sequence

While the Cph1 target was Chain A, we treated the new binder as Chain B in this part of the workflow. We asked ProteinMPNN to design only the sequence of the binder; however, we still gave the model structural context from Chain A (Fig. 1). This was an important step in our workflow because the sequence at the binder interface should respond to the residues on the target. The transcript gives the example that if a residue on Chain A is hydrophobic and contacts the binder, the designed binder residue may also need a compatible property. For each of the 370 backbones, ProteinMPNN generated 10 sequences, producing a total of 3,700 sequences.

ProteinMPNN designed the binder sequence while using Chain A as structural context

For each of the 370 backbones, ProteinMPNN generated 10 sequences. This produced a total of 3,700 sequences.

Threading and FastRelax

After obtaining these new sequences, we processed them with threading and FastRelax. Threading placed each designed amino-acid sequence onto its RFdiffusion3 backbone. Then, FastRelax adjusted the side chains and backbone (Fig. 2). The purpose of this process was to refine the structure and move it into a more relaxed arrangement, although the changes were sometimes small. In the examples, the threading and FastRelax output mostly overlap, but the helices and nearby regions do not match perfectly. After this stage, we had 3,700 refined structural models, with proper amino-acid sequences.

FastRelax made local adjustments to the threaded structures

After this stage, we had 3,700 refined structural models.

Filtering by interface energy

At this stage, we added a new, relevant, metric: dG/dSASA. In the overall project, I described this metric as interface binding energy divided by the amount of surface area involved in binding. The key point was that more negative values were treated as more favorable (Fig. 3). Across the full set, most values fell between -1 and -3. For alpha-rich designs, we used a cutoff below -2.35. For beta-rich designs, we later used a looser cutoff below -1.80.

The dG/dSASA distribution used to prioritize designs by interface energy

We also measured the root mean square deviation (RMSD) between the FastRelax output and the earlier RFdiffusion3 structure. The cutoff table for this stage used an RMSD of 0.75 angstroms or less. As previously stated, we included other filters, such as contacts with Chain A, radius of gyration, loop percentage, clashes, and the maximum lengths of helices and beta sheets (Fig. 4).

The post-ProteinMPNN metric ranges used to filter the 3,700 designs

A bias toward alpha-helical designs

One thing that we noticed as we reviewed the candidate models was that RFdiffusion3 strongly favored helical structures. After applying strict cutoffs, many beta-rich designs were removed, and this created a problem for our goal of maintaining diversity. If every final binder had a similar alpha-helical structure, the final set would test only a narrow region of the available design space. To address this issue, we created a beta rescue. We allowed designs with more than 70% beta content to use a less strict dG/dSASA cutoff. Instead of requiring a value below -2.35, they needed a value below -1.80 (Fig. 5). This did not mean that all beta-rich designs automatically passed, but it constituted a different threshold so that some of those designs could remain in the candidate set.

A looser interface-energy cutoff helped protect beta-rich designs from the pipeline’s helix bias

This did not mean that all beta-rich designs automatically passed. It gave them a different threshold so that some could remain in the candidate set.

Preserving different structural types

Our selected designs included different combinations of alpha helices and beta sheets.

One example had:

• 0 percent alpha

• 82.52 percent beta

A second had:

• 38.78 percent alpha

• 39.8 percent beta

A third had:

• 85.86 percent alpha

• 0 percent beta

Selected binders included beta-rich, mixed alpha-beta, and alpha-rich structures

These structures looked very different from one another. That variety was intentional.

The team wanted the final designs to represent several possible structural solutions rather than repeating the same type of binder.

Preserving different binding poses

We also wanted the binders to occupy different positions near the target interface.

A binder’s pose describes where it sits against the target. If all designs used the same pose and that region later failed in the wet lab, then many candidates could fail for the same reason.

To estimate pose diversity, we compared the proposed binder with the position of Chain B in the Cph1 dimer.

A lower number of clashes meant that the binder tended to sit farther from the original dimeric interface. A higher number meant that the binder extended more deeply into that space.

Different overlap values with the original Cph1 Chain B position were used to preserve binder-pose diversity

This gave us a way to retain designs that approached the target from different positions.

Reducing 3,700 designs to 107

The above-mentioned filters greatly reduced the candidate set (Fig. 8). The largest change came from the dG/dSASA filter. The later filters removed additional candidates based on structural agreement, contact quality, compactness, secondary structure, and clashes. The slide lists the progression as follows:

· 3,700 designs started the stage

· dG/dSASA removed 3,256

· RMSD removed 115

· Percent loop removed 110

· Chain A contacts removed 38

· Radius of gyration removed 19

· Helix and sheet length removed 43

· Self-clashes removed 2

· 107 designs remained

Post-ProteinMPNN filtering reduced 3,700 sequence-based designs to 107

The largest change came from the dG/dSASA filter. The later filters removed additional candidates based on structural agreement, contact quality, compactness, secondary structure, and clashes.

Conclusion

ProteinMPNN expanded 370 backbones into 3,700 possible sequences. Threading and FastRelax then created refined structures for those sequences.

Our filters reduced the set to 107 designs. During this process, we also made a deliberate choice to protect diversity. We kept different secondary-structure types and different binding poses so that the final candidates would not all depend on the same structural idea.

The remaining 107 sequences then moved into AlphaFold3 for the last major computational evaluation.

Next in the series: How AlphaFold3 predictions and final metrics reduced 107 sequences to 27 candidates for possible wet-lab testing.

Previous in the series: Choosing the Target and Building the Protein-Design Pipeline

Using RFdiffusion3 to create and filter de novo structures

On our previous post I explain how we selected the binding region on Cph1; now here I will explain the process by which we generated possible binder structures. To achieve this, I used RFdiffusion3, a generative AI model that creates new protein backbones for a target site. These backbones were de novo designs, meaning that the model didn’t simply modify existing binders, but rather created new structural shapes based on the target and hotspot information we provided.

Testing the amount of target context and running RFdiffusion3 at scale

Before running the model at full scale, we needed to answer an important setup question: How much of the Cph1 structure should RFdiffusion3 see? This is important because if we provided the program a large portion of the protein, the simulations would be more precise, but the computing power required could be very high. On the other hand, if we provided it a minor portion, the computing requirements would lower, but at risk of delivering inadequate models. We decided to compare three levels of structural context (Fig. 1): the first option included 160 selected residues, the second included 249 residues, and the final option included the full 492-residue Cph1 monomer.

After comparing the test process and the preliminary results, we selected the complete 492-residue monomer, based on two main reasons. First, using the full monomer rarely caused GPU out-of-memory errors during the initial testing. Second, the complete protein gave RFdiffusion3 the greatest amount of structural context while it designed binders around the target region. The final contig was listed as A19–510.

The proper binder backbones generation took about three days. The transcript notes that the run was slower than expected because of summer heat waves and unavailable computing nodes, aspects that we will take into consideration in future projects. At the end of the run, RFdiffusion3 had generated 1,082 binder backbones. After this, our next step was coming up with a consistent way to remove structures that appeared unrealistic or poorly suited to the target.

Building the first structural filter

We evaluated the RFdiffusion3 outputs using several metrics, each addressing a different feature of the proposed binder (Fig. 2):

· Radius of gyration

· Clashes with Chain A

· Self-backbone clashes within the binder

· Loop percentage

· Clashes with the dimeric Chain B position

· Length of the longest alpha helix

· Length of the longest beta sheet

· Minimum number of contacts with Chain A

We will now explain some of these metrics with more detail.

Using the full Cph1 monomer, RFdiffusion3 generated 1,082 binder backbones over about three days

Selecting compact structures

The first step into selecting which binder models could be adequate for experimental testing was measuring the radius of gyration. This is a measure that describes the distribution of the atoms of a protein around its center of mass. In other words, it measures how compact each binder was. Very small values could represent an unrealistically collapsed structure, whereas larger values could represent an extended structure that might not fold or interact as intended. We kept binders with a radius of gyration between 11 and 13.7 angstroms. This range was selected by looking at the distribution of generated binders and visually reviewing example structures (Fig. 4).

The structural metrics and initial cutoff ranges used to filter RFdiffusion3 outputs

Each metric addressed a different feature of the proposed binder.

Selecting compact structures

Radius of gyration measured how compact the binder was.

Very small values could represent an unrealistically collapsed structure. Larger values could represent an extended structure that might not fold or interact as intended.

We kept binders with a radius of gyration between 11 and 13.7 angstroms. This range was selected by looking at the distribution of generated binders and visually reviewing example structures.

Radius-of-gyration filtering favored compact binders between 11 and 13.7 angstroms

Limiting flexible loop content

We also measured the percentage of how much each binder was made up of loops. In our filtering approach, binders with unusually high loop content were less desirable. We decided this because loops were treated as more flexible than alpha helices and beta sheets, so designs with too much loop content could be harder to design reliably. The cutoff table set binder loop percentage at 27.44% or less (Fig. 5).

Examples of binders with different loop percentages used to guide the loop-content filter

Restricting oversized structural elements

Other filter metrics that we used were the length of the longest alpha helix and the longest beta sheet. On one side, very long helices could create elongated structures that were not preferred for this target. On the other side, large beta sheets could dominate the fold and create broad or flat interfaces. We defined the alpha-helix and longest beta sheet limits as 32 and 27 angstroms, respectively (Fig. 6, Fig. 7). Some structures shown as selected in the original histograms extended past these helix and sheet limits. The transcript explains that these two cutoffs were not applied during the first RFdiffusion3 filtering run. They were added later, after the ProteinMPNN stage.

The longest alpha helix was limited to 32 angstroms

The longest beta sheet was limited to 27 angstroms

Filtering overview

The filtering pipeline described above reduced the 1,082 RFdiffusion3 binders to 370 candidates. The largest listed reduction came from loop content. Radius of gyration, backbone clashes, and contact requirements removed additional structures. The total number of models rejected per each metric was as follows:

· 130 rejected because of backbone clashes

· 104 rejected because of contact criteria

· 358 rejected because of loop percentage

· 120 rejected because of radius of gyration

· 370 passed the filtering stage

The remaining 370 candidate models had designed shapes, but they did not yet have the final amino-acid sequences needed for the next stages of the project.

The first filtering stage reduced 1,082 RFdiffusion3 binders to 370 candidates

The largest listed reduction came from loop content. Radius of gyration, backbone clashes, and contact requirements removed additional structures.

These 370 candidates then moved into the next major stage: sequence design.

Conclusion

RFdiffusion3 gave us a large starting set of possible binder backbones. We selected the full Cph1 monomer as input because it provided the greatest structural context and remained workable on the available GPUs.

The model produced 1,082 designs. By filtering for compactness, loop content, structural-element length, contacts, and clashes, we reduced that set to 370 candidates.

Those candidates had designed shapes, but they did not yet have the final amino-acid sequences needed for the next stages of the project.

Next in the series: How we used ProteinMPNN and FastRelax to generate 3,700 sequence-based designs while preserving structural and positional diversity.

Designing a binder for a photoswitchable protein

Hello! This is the first in a series of posts that will cover my summer research project on molecular engineering. The overall system we are looking to construct aims to support the synthesis of single-stranded DNA without a template (Fig. 1). Its is an ambitious project, requiring the collaboration of many. In this series of posts, I have focused on one step of the process: designing a new protein binder for Cph1, a photoswitchable protein.

Why a protein like that? you may ask. First of all, we need to look at the overall project. The idea is to design a system that uses split TdT enzymes. These enzymes add nucleotides to the 3’ end of a DNA molecule, enabling the construction of our target product. One part of the system would act as a universal catalytic domain, while another part would be specific to a DNA base. For the system to work, we need to put together the necessary components, and that’s where Cph1 comes in handy.

A protein controlled by light

Cph1 is a photoswitchable protein. This means that we can turn the light/bright state of the protein on and off with the correct light pulse stimulus. But this not the only property of Cph1; it is also a dimer, two similar units bonded together, called Chains A and B (Fig. 2). An engineered Cph1 variant has a really cool feature: it can change between dimeric and monomeric states when exposed to different wavelengths of light. In the light state, exposure at 616 nm can promote dimerization, whereas exposure at 740 nm can return the protein to a dark state and lead to monomerization (Fig. 3). This reversible change creates a possible way to control when two protein components come together, a key step in our TdT-system.

Dimerization levels for wild-type Cph1 and engineered variants under different light conditions

Finding the location for a binder

Determining the binder location.

To design the binder, we first needed to decide where exactly it should interact with Cph1. The most logical step was to examine the interface where Chains A and B meet. This is the region involved in dimerization, and the goal was to place the binder near the center of this interface. We used the software PyMOL, an open-access molecular visualization software, to inspect the Cph1 structure and identify the interface. With this procedure we were able to choose the most favorable region for the binding hotspots. We selected residues A307 and A137, which were described as hydrophobic and aromatic residues (Fig. 4).

The selected binder-design region on Chain A. Blue residues lie within 8 angstroms of Chain B, and red residues mark the hotspots

A multi-stage design process

Now that we had selected where we wanted the binding hotspots to be, we needed to actually design them. For this process, we built a computational pipeline that relied on several computational and bioinformatic tools to model the binder candidate models. It is important to note that we generated several iterations of designs and models at eatch stage of the pipeline. Some designs looked promising at first but failed later tests. Others had shapes that were too extended, too flexible, or poorly positioned against the target. For that reason, the workflow followed a repeated pattern: generate designs, calculate metrics, remove weak candidates, and pass the remaining candidates into the next stage. This process was central to the entire project. The design models created possibilities, but the filtering and analysis scripts helped us decide which possibilities were worth keeping.

The pipeline consisted in these steps (Fig. 5):

1) Inspect the Cph1 structure using PyMOL for selecting the binding hotspots.

2) Generate new protein backbones using RFdiffusion3.

3) Generate amino-acid sequences using ProteinMPNN. These sequences needed to match the previously designed structures.

4) Place the newly generated sequences onto their intended backbone using threading.

5) Adjust the structure to improve its local arrangement using FastRelax.

6) Predict structures from the designed sequences and compare these predictions with the intended designs using AlphaFold3.

Why the project required repeated filtering

The project did not rely on a single model or a single score.

At each stage, the number of designs grew or changed. Some designs looked promising at first but failed later tests. Others had shapes that were too extended, too flexible, or poorly positioned against the target.

For that reason, the workflow followed a repeated pattern:

Generate designs, calculate metrics, remove weak candidates, and pass the remaining candidates into the next stage.

This process was central to the entire project. The design models created possibilities, but the filtering and analysis scripts helped us decide which possibilities were worth keeping.

Conclusion

Our starting point was a biological system controlled by light. From there, we identified the dimeric interface of Cph1 and selected a region where a new binder could interact with Chain A.

We then built a computational pipeline that connected structure inspection, backbone generation, sequence design, structure refinement, prediction, and filtering.

The next challenge was to run the first major generation stage at scale. That meant deciding how much of the target structure to provide to RFdiffusion3 and then finding a way to filter more than 1,000 generated binders.

Next in the series: How we used RFdiffusion3 to generate 1,082 protein backbones and reduced them to 370 candidates.

That’s not random. That’s your amygdala doing exactly what it was built to do: hold onto fear like it’s the most precious thing in the world. Because evolutionarily speaking, it once was. And the wild thing is, scientists at Radboud University in the Netherlands just published research in Science Advances that shows we might be able to change how that process works using nothing but sound waves.

I spent weeks digging into this paper, and honestly it changed the way I think about fear, trauma, and what it even means to “get over” something scary. So let me walk you through it, because I think it deserves way more attention than it’s getting.

What the Amygdala Actually Does

The amygdala is a small, almond-shaped structure sitting deep inside both hemispheres of your brain. It’s part of the limbic system, which is basically your brain’s emotional control center, and it’s one of the oldest structures we have from an evolutionary standpoint. According to the National Center for Biotechnology Information (NCBI, NBK537102), the amygdala processes emotions, especially fear, and it’s heavily connected to the hypothalamus, which is what triggers your body’s physical stress response.

When you perceive a threat, your amygdala fires first. It signals your hypothalamus, which signals your adrenal glands, which dump adrenaline and cortisol into your bloodstream. Your heart rate spikes, your muscles tense, your pupils dilate. Harvard Health Publishing calls this the “fight-or-flight” response, and it happens in milliseconds, long before your prefrontal cortex, the part of your brain that actually thinks, gets a chance to process what’s going on.

What I find genuinely fascinating about this is that the amygdala doesn’t wait for you to decide something is scary. It decides for you. It acts on incomplete information at incredible speed, because in the evolutionary context where this system developed, waiting to think cost you your life. Your brain essentially said: I would rather you panic over a stick that looks like a snake than get bitten by an actual snake while you’re still deliberating. That logic makes total sense for survival. The problem is we carry that same system into a world where our “threats” are exams, social situations, and memories of things that have already ended.

How the Brain Actually Learns Fear

To understand what the researchers did, you first need to understand Pavlovian conditioning. Most people know this from the famous dog experiment where Pavlov rang a bell every time he fed his dogs, and eventually the dogs started salivating at just the bell, with no food present. The bell became a conditioned stimulus (CS), something that triggers a learned response even though it was originally completely neutral.

Your brain does the exact same thing with fear. According to a 2015 review in PMC on Pavlovian conditioning and associative learning, the brain runs three major circuits covering defense, feeding, and reproduction, and each one has its own mechanisms for learning associations. The defense circuit is the one we care about. When something bad happens alongside a neutral stimulus, your brain links them together. Now that neutral thing carries a threat signal, even when the original threat is completely gone.

Here’s the part that I think is genuinely underappreciated: all three of those circuits share a common structural feature. There’s a negative feedback loop where the conditioned stimulus activates a pathway that then dampens the reinforcing unconditioned stimulus. Basically, your brain is constantly doing predictive threat management. Its ability to do this fast and efficiently is arguably what kept the human species alive long enough to build civilization. But the same efficiency that makes it work so well is also what makes fear so hard to shake when the system misfires.

This is the core problem for people with PTSD, phobias, and anxiety disorders. The conditioned stimulus keeps triggering the fear response long after the actual threat is gone. The brain learned the lesson and won’t unlearn it. That’s what the Radboud researchers wanted to understand and potentially change.

The Experiment

The researchers used a technology called Transcranial Ultrasound Stimulation, or TUS. It’s non-invasive, so no surgery, no implants, nothing going into anyone’s brain. Low-intensity acoustic sound waves are directed through the skull at a specific brain region. I kept trying to wrap my head around this when I first read it, because the idea that you can aim sound at a specific structure deep inside the brain and change what it’s doing, without touching anything, feels like science fiction. But it’s not. The protocol followed ITRUSST safety recommendations, which is an international expert consensus on safe TUS parameters, so participants weren’t at any medical risk.

They recruited 50 participants total, split into two experiments of 25 each. Before anything else, every participant got an MRI scan so researchers could pinpoint exactly where their amygdala and hippocampus sat, because brains differ from person to person and the precision here really mattered

.

Then came the snakes. A red snake image was the CS+, meaning it appeared alongside a mild electric shock 50% of the time. A green snake was the CS−, shown just as often but never followed by a shock. The green snake was the safety signal. Participants wore a device on their heads delivering either real TUS or sham TUS, which is fake stimulation with no actual ultrasound. They didn’t know which one they were getting. This is basically the neuroscience equivalent of a placebo condition, and it’s critical for making the results trustworthy, because if you think something is happening to your brain, your behavior changes even if nothing actually is.

Fear was measured using skin conductance response, or SCR, which tracks the tiny changes in sweating that happen when you’re stressed or scared. The researchers were upfront that SCR is a proxy for fear rather than fear itself, it varies between people and doesn’t capture what someone is actually feeling subjectively. But it’s one of the most reliable physiological measures available for this kind of research, and it gives you something objective to track across 120 acquisition trials.

What Actually Happened, Phase by Phase

The experiment moved through five stages, and the progression really is a story if you think about it.

During acquisition, which was 120 trials of seeing the snakes with and without shocks, participants in the sham condition showed normal fear learning. Their skin conductance responses went up for the red snake and stayed flat for the green one. This confirmed the basic experiment was working and that people were actually learning the threat association. Nothing surprising yet.

After a break came retention, just four trials with TUS turned off entirely. Participants still feared the red snake. The fear memory had consolidated. It didn’t fade on its own just because time passed. This is the thing that gets me every time I think about it, because it’s so counterintuitive to how we talk about fear in everyday life. We say things like “you just need time” or “you’ll get over it,” but biology doesn’t really work that way. The brain holds on specifically because letting go could be dangerous. A threat you survived once is worth remembering.

Then came extinction, which was 40 trials where the shocks stopped entirely. Participants kept seeing the red snake, nothing bad happened, and slowly their SCR responses dropped. This mirrors what therapists call extinction learning, and it’s actually the scientific basis of exposure therapy. You confront the feared thing in a safe context repeatedly until the association weakens.

After that, reinstatement: three sudden unannounced shocks. The fear came back fast. This is called fear reinstatement, and it’s one of the most clinically frustrating things about anxiety treatment. The fear that faded during extinction wasn’t actually erased. It was just suppressed. The original memory stayed underneath, and it took almost nothing to reactivate it. That’s why PTSD relapses happen. That’s why a single triggering event can undo months of therapy progress. And understanding that gap between “extinction” and “erasure” is honestly one of the most important things I took away from this whole project.

Finally, re-extinction: they ran extinction again. The fear faded, and this time it faded faster.

Whats Really Interesting

In Experiment 1, TUS was aimed at the amygdala. In Experiment 2, it targeted the hippocampus as an active control, same procedure, different brain region, so the researchers could make sure any effects they found were actually specific to the amygdala and not just a generic result of pointing ultrasound at someone’s head.

When the amygdala was disrupted with real TUS, three specific things happened. Early threat acquisition slowed significantly, with a reported effect of t(24) = -2.98, P = 0.0065, Cohen’s d = -0.60. Then, when shocks stopped, those threat memories extinguished faster during early extinction: t(24) = -2.81, P = 0.0097, Cohen’s d = -0.56. And participants became less accurate in their retrospective estimates of how often they’d been shocked: t(24) = 2.14, P = 0.0427, meaning they overestimated the shock rate even as their behavioral fear was changing.

When the hippocampus was targeted in Experiment 2, essentially none of this happened. No parallel learning effects. This matters a lot because it means the amygdala finding isn’t just noise, it’s specific to that region.

The researchers also used computational modeling to go deeper than just the behavioral results, and this is genuinely the part of the paper I find most conceptually exciting. The model found reduced acquisition learning rates, increased extinction learning rates, and a drop in what the authors call “emotional learning bias,” which is the brain’s built-in tendency to weight threatening experiences more heavily than safe ones. They describe the amygdala’s natural role as putting the brain into a state of “learning fast, forgetting slow.” When you disturb that state during the formation of a new fear memory, you get the opposite pattern: the fear forms more slowly, and it releases more readily when you try to extinguish it later.

The idea that the amygdala isn’t just an alarm bell but is actually setting the brain’s entire learning policy during threatening situations is something I had to sit with for a while before it clicked. It’s not just reacting to fear. It’s deciding how deeply the brain should engrave the experience. And the fact that you can adjust that dial with sound waves, at least in a controlled lab setting, feels like a genuinely new kind of knowledge about how the brain works.

Possibilities

The World Health Organization estimates that 1 in 4 people will experience a mental health condition at some point in their lives. PTSD specifically affects around 3.9% of the global population according to Our World in Data, which is hundreds of millions of people carrying fear memories that their brains refuse to update, even when the person desperately wants to move forward.

The standard treatments, cognitive behavioral therapy and exposure therapy, work by essentially doing extinction manually. You confront the feared stimulus in a safe context over and over until the fear response decreases. But as this study makes clear, extinction doesn’t erase the original memory. It creates a competing one. The original threat association is still there underneath, which is why reinstatement works, and why real-world patients relapse after triggers.

What I keep thinking about is this: if TUS could be used during exposure therapy to make extinction faster and more durable, it doesn’t replace the therapy. It just makes the therapy work better, for people for whom it currently doesn’t work well enough. The Radboud University press release specifically suggested that if a fear memory is reactivated, amygdala-targeted stimulation during that window might help the memory update more quickly. That’s still hypothetical. But it’s a real hypothesis grounded in real mechanistic evidence, and it’s a lot more scientifically honest than the usual “scientists found a way to erase fear” headline that this kind of research attracts.

The researchers also mentioned the possibility of an ear-worn device that could reduce how much fear gets encoded from a moment in real time, kind of like a passive intervention that sits in the background while someone is going through exposure therapy. I don’t know if that’ll work in practice. But the fact that it’s a plausible direction based on what the amygdala actually does is interesting in a way I didn’t expect this project to get to.

Limitations

The participants were healthy adults. The study explicitly excluded people with psychiatric disorders and snake phobias. The fear was entirely lab-made, 120 trials of seeing snake images and occasionally getting a mild shock, which is genuinely not the same thing as surviving something traumatic. Sample sizes were 25 people per experiment, which is small enough that we have to be careful about how broadly we apply the conclusions. And skin conductance responses are a proxy, not a direct readout of subjective fear or clinical symptoms.

Most importantly, TUS in this study was applied during fear acquisition, while new memories were being formed. It hasn’t been tested on fear memories that already exist. For someone who already has PTSD from something that happened years ago, we don’t yet know if targeting the amygdala during a reactivation window would help, hurt, or do nothing. The paper explicitly calls this out as the essential next step. And one more thing worth saying clearly: the protocol required individualized MRI-based brain targeting, acoustic simulations, specialized bilateral transducers, and expert safety oversight. This is not a DIY intervention. It’s a tightly controlled lab procedure.

What I Actually Walked Away Thinking

Before this project, I think I understood fear mostly as a feeling. Something you experience, something you work through, something therapy helps you manage. What I didn’t fully appreciate was how mechanical it is at the biological level. There are specific brain structures, specific phases of learning and extinction, specific computational properties that determine how sticky a fear memory becomes. The amygdala isn’t just producing a feeling. It’s running a learning algorithm, and that algorithm has parameters.

The part that genuinely changed something for me is the “learning fast, forgetting slow” framing. Because I’ve always wondered why fear is so much easier to acquire than to lose. Why one bad experience can rewire how you feel about something for years, while months of positive experiences barely make a dent. And the answer, at least partially, is that the amygdala is specifically designed to make that asymmetry happen.

Knowing that this mechanism is identifiable, targetable, and potentially adjustable doesn’t make me think fear is fake or that trauma isn’t serious. It actually does the opposite. It makes me take it more seriously, because now there’s a specific biological explanation for why it’s so hard to shake. And if there’s a specific mechanism, there might eventually be a specific way to help people whose brains are stuck in “learning fast, forgetting slow” long after the threat has passed.

Sources: Science Advances, “The human amygdala in threat learning and extinction,” Radboud University (DOI: 10.1126/sciadv.aea8233); Radboud University Research Release (March 25, 2026); Neuroscience News (March 30, 2026); NCBI (NBK537102); Harvard Health Publishing, Understanding the Stress Response; PMC, Pavlovian Conditioning and Associative Learning (2015); WHO Mental Health Atlas; Our World in Data, PTSD Prevalence; ITRUSST Biophysical Safety Consensus (Aubry et al., 2023); IFCN-endorsed ITRUSST Practical Guide (Murphy et al., 2024).

So here’s what I found.

It Starts With Faces

One of the earliest and most famous uses of PCA is something called Eigenfaces. In 1991, researchers at MIT figured out that if you run PCA on a database of face images, the principal components that come out literally look like faces - ghostly, blurred versions of faces that each capture a different source of variation across the dataset. To recognize a new face, you just project it onto those components and get a compact set of scores. Compare those scores to your database and you have face recognition. Turk and Pentland reported 96% accuracy in 1991, on a standard workstation, powered entirely by PCA.

That’s the core idea: take thousands of variables (in this case, pixels) and compress them into a small number of components that still capture what matters. That same logic is exactly what researchers are now applying to medicine.

What Is PCA Exactly?

Before getting into the medical papers, it helps to be precise about what PCA actually does.

PCA stands for Principal Component Analysis. It takes a dataset with many variables and produces a smaller set of new variables called principal components. Each principal component is a weighted combination of the original variables, where the weights are determined by what captures the most variance in the data. The first component explains the most variance, the second explains the next most, and they are all uncorrelated with each other.

For PCA to work properly, the variables need to be linearly related to each other, the data needs to be standardized so everything is on the same scale, and it works best when variables can be negative. There are also specialized versions - Kernel PCA handles nonlinear relationships, Robust PCA deals with outliers, and there are others like Independent Component Analysis and Probabilistic PCA. But standard PCA is what both studies below used, and it was enough.

The key insight is that PCA does not create new information. It reorganizes existing information more efficiently. In face recognition, that means compressing 10,000 pixel values into maybe 50 Eigenface scores without losing the ability to tell people apart. In medicine, it means compressing a panel of hormone measurements or brain scan pixels into a single score that a doctor can actually use.

Study 1: Predicting Alzheimer’s Disease From a Brain Scan

One of the more interesting papers I read was published in the Journal of Nuclear Medicine in 2019 by Blazhenets et al. The question: can you look at a brain scan today and predict whether someone will develop Alzheimer’s dementia years from now?

The tool they used was called 18F-FDG PET scanning. PET stands for positron emission tomography. It measures glucose metabolism in the brain because wherever glucose goes, brain activity follows. So a PET scan is a map of how active different brain regions are at a given moment.

The problem is that one of these scans produces thousands of data points. Each pixel in the image represents metabolic activity at a specific location. Sound familiar? Just like with face images, you cannot stare at thousands of pixels and make a good clinical decision. You need PCA to compress them into something usable

.

What They Did

They took 18F-FDG PET scans from patients with mild cognitive impairment (an early warning sign before Alzheimer’s) and tracked them over time to see who eventually converted to full Alzheimer’s dementia and who didn’t. Before running PCA, they normalized the images by fitting each brain scan to a standard template and smoothing the data for consistency. They also ran a preprocessing step to remove uninformative global and regional averages.

Then came the PCA itself. They selected principal components capturing the top 50% of variation in the data, then used linear regression to find which of those components differentiated converters from non-converters.

From the PCA, they extracted a single number called the Pattern Expression Score, or PES. It captures the overall metabolic signature of the brain. Think of it like an Eigenface score, except instead of “how much does this face look like component 3,” it’s “how much does this brain’s metabolism match the Alzheimer’s pattern.”

They also built a separate comparison model called a Volume of Interest (VOI) model, which used t-tests to identify individual brain pixels that differed between groups. Then they combined PES, the VOI predictions, and clinical variables into a Cox regression model to predict conversion over time.

What They Found

The combined model performed best overall. But the key finding was that the PES from PCA outperformed the VOI model alone. The holistic PCA-generated score was more predictive than monitoring individual brain regions in isolation.

They also confirmed that the PCA pattern matched regions already known to be affected in Alzheimer’s, like the posterior cingulate and parietal cortex. The math was picking up on biologically real signals, not noise.

If a PET scan processed through PCA can predict years in advance who will develop Alzheimer’s, that means earlier diagnosis and earlier treatment. PCA turned thousands of pixels into one meaningful, predictive number.

Study 2: Using PCA to Monitor Treatment in Kids With a Rare Hormone Disorder

This is the paper I found second, but its the more detailed of the two. It was published in Frontiers in Endocrinology in 2021 by Ljubicic et al., and it’s about children with Congenital Adrenal Hyperplasia, or CAH.

What Is CAH?

CAH is a genetic disorder. The most common cause is a deficiency in an enzyme called 21-hydroxylase, which is part of the process your adrenal glands use to make hormones. When it doesn’t work properly, the whole system goes out of balance.

Cortisol and aldosterone are underproduced. At the same time, 17-hydroxyprogesterone (17-OHP) and androgens build up because the blocked pathway redirects hormones elsewhere. In children, this causes real problems such as abnormal growth, BMI changes, early puberty, virilization, and blood pressure issues.

There are two main forms: classical CAH, which is more severe, and non-classical CAH, which has more residual enzyme activity and milder symptoms.

Treating CAH in kids is genuinely difficult. Doctors have to balance glucocorticoid doses carefully: too little and the disease isn’t controlled, too much and you risk stunting growth or triggering other side effects. Every clinic visit involves measuring a growing panel of hormones and making a judgment call. As the paper states, “interpreting combinations of these markers introduces complexity, subjectivity, and bias into the patient care.” That’s the same problem face recognition had before Eigenfaces. There are too many variables and no good way to combine them. PCA is the solution in both cases.

The Data

The study followed 33 children and adolescents, ranging in age from just under one year old to nearly nineteen, at a clinic in Copenhagen. Over time, those 33 patients racked up 406 total clinic visits. The researchers treated each visit as its own data point, since treatment is adjusted visit by visit in the real world anyway.

At every visit, seven hormones were measured from a blood sample. Four of them are adrenal steroids directly tied to CAH - 17-OHP, DHEAS, androstenedione, and testosterone. The other three - SHBG, LH, and FSH - give a broader picture of how the endocrine system is functioning overall.

Here’s one thing the study did really cleverly: instead of using the raw hormone numbers, they converted everything into standard deviation scores adjusted for age and sex. So instead of asking “is this testosterone level high?”, they asked “is this testosterone level high for a 7-year-old girl?” That adjustment meant a toddler and a teenager could sit in the same PCA model without the age difference distorting everything. It’s the kind of detail that sounds small but actually makes or breaks whether the analysis is valid.

What PCA Found

Because classical and non-classical CAH are biologically different diseases, the researchers ran PCA separately for each group rather than lumping everyone together.

The results were already interesting before they even got to the treatment question. In classical CAH, the first principal component, the one capturing the most variation across all patients, was almost entirely explained by one hormone: 17-OHP. It had a correlation of 0.93 out of 1.0 with that component, meaning 17-OHP was basically driving the whole first dimension of variation on its own. In non-classical CAH, the picture was messier. The first component was shared more evenly across multiple hormones, with 17-OHP, DHEAS, androstenedione, and testosterone all contributing meaningfully.

What that tells you is that the two forms of CAH don’t just differ in severity, they differ in their underlying hormonal structure. Classical CAH is dominated by the hormone 17-OHP. Non-classical CAH is more complicated. PCA made that visible in a way that a table of average hormone values couldn’t.

Did the PC Scores Actually Predict Treatment?

This is the key question. The researchers assessed treatment efficacy using clinical markers - things like how close a child’s height was to their expected height, their BMI, and their blood pressure. Crucially, none of these clinical markers went into the PCA model. The PCA only saw hormone data. So the test was: can a score built purely from hormones predict whether a child is being treated well or not?

To measure predictive accuracy, they used something called an AUC score - area under the ROC curve. You can think of it as a grade for how good a predictor is. A score of 50% means you might as well flip a coin. A score above 90% is considered excellent.

The individual hormones did not do well. 17-OHP, which is the hormone doctors most commonly rely on for CAH management, scored an AUC of just 56%. Barely above random. The best single hormone was FSH, which reached 80% accuracy, but still missed a lot of cases.

The PCA scores were a different story. For classical CAH, the second principal component score hit an AUC of 93% and an overall accuracy of 95%. For non-classical CAH, the third principal component reached 80% AUC and 91% accuracy. The classical model caught every single insufficiently treated patient - 100% sensitivity. None slipped through.

That 17-OHP result is the one I keep coming back to. It’s the hormone CAH patients get monitored on most closely, and it basically told doctors nothing about whether treatment was working. That’s not a knock on the doctors - the clinical guidelines actually acknowledge this already. But it does show why combining multiple hormones through something like PCA matters is so meaningful

.

A Side Finding Worth Mentioning

At the end, the researchers ran one more PCA combining all patients together - both classical and non-classical - just to see what would happen. When they plotted the results, the two groups naturally separated into distinct clusters without any instruction to do so. The math just sorted them.

Looking at which hormones drove that separation, they noticed that 17-OHP and DHEAS were pointing in completely different directions in the plot. That suggested their ratio might be useful for telling the two forms of CAH apart. They tested it and found an accuracy between 81 and 89%. It’s not ready for clinical use yet since most patients were already on treatment during the study, which changes hormone levels. But it’s a new lead that PCA generated just by finding structure in the data

.

What All Three Have in Common

It’s kind of interesting that the same statistical method shows up in face recognition, Alzheimer’s research, and pediatric endocrinology. On the surface those things have nothing to do with each other. But the underlying problem is the same every time - there’s too much data, too many variables, and you need a way to make sense of it without just staring at a spreadsheet.

PCA solves that by finding the structure that’s already in the data and expressing it more cleanly. Whether that’s pixels in a face image or hormone levels in a 6-year-old, the math works the same way.

That said, both medical studies are early-stage work. The CAH study had only 33 patients, and the authors themselves describe it as a pilot study. The Alzheimer’s study had 272 patients, which is bigger, but still a relatively small sample for the kind of model they were building. Neither is ready to change clinical practice on its own. They both need to be tested on larger, more diverse populations first.

What I Actually Think

Medicine tends to look at one thing at a time. One hormone, one brain region, one marker. But the body doesn’t really work that way - everything is connected, and looking at variables in isolation misses a lot. What PCA does is take that interconnectedness seriously and actually do something useful with it.

I also think it’s just genuinely cool that the same method works across such different problems. The researchers studying CAH in kids and the ones analyzing brain scans weren’t thinking about face recognition, but they were all reaching for the same tool for the same reason. That kind of cross-domain consistency usually means something is onto something real.

Both papers are careful to acknowledge their limitations, and a 33-patient pilot study obviously isn’t enough to change how doctors practice medicine. But as a proof of concept, both are pretty convincing. The next step is bigger datasets, and that seems like a matter of time.

Sources: Ljubicic ML, Madsen A, Juul A, Almstrup K and Johannsen TH (2021). Front. Endocrinol. 12:652888. doi:10.3389/fendo.2021.652888 | Blazhenets G et al. (2019). Journal of Nuclear Medicine, 60(6), 837-843. doi:10.2967/jnumed.118.219097 | Turk M and Pentland A (1991). Eigenfaces for Recognition. Journal of Cognitive Neuroscience, 3(1), 71-86.

Recently, I read that Biolizard’s platform, Bioverse Navigator, won Bioinformatics Innovation of the Year at the 2025 Biotech Breakthrough Awards, and it made me curious. Bioinformatics tools get released all the time, but awards for such tools are not as common. So I started to look into what Bioverse Navigator actually does, why it stood out, and what it really changes.

As I looked into it more, it was interesting for me to see that it was not that they created a new algorithm or that they discovered a new gene. They were fixing something more basic but also more frustrating. Biology is becoming more and more of a data-heavy science, and a lot of researchers are struggling not with collecting the data, but with organizing it, maintaining it, and making sense of it in a timely fashion.

Biology is Generating More Data Than it Can Easily Handle

Modern biology does not produce small datasets anymore. A single human genome contains about 3 billion DNA base pairs, and sequencing technologies read these base pairs multiple times to reduce error, which makes the data files grow even larger. Gene expression experiments add another layer, since researchers measure activity across more than 20,000 genes at once, often across many samples. When proteomics, metabolomics, imaging data, and patient health records are included, the amount of data becomes overwhelming very quickly.

Scientists have been warning about this for years. A well known paper published in PLOS Biology predicted that by 2025, genomic data storage alone could reach between 2 and 40 exabytes, depending on how fast sequencing technology advances. One exabyte equals one billion gigabytes. The authors even compared genomics to platforms like YouTube in terms of data demand, which sounds dramatic but actually helps explain the scale of the problem

At this point, scientists do not struggle to generate data. They struggle to store it, process it, and actually understand what it is telling them.

Data Silos Make The Problem Worse

Through research and projects, I’ve learned that the size of data and data storage is not the only issue. The way the data is stored and managed also causes problems. In most research, different types of data live in different places. DNA sequencing data might be stored in one system, RNA expression results in another system, protein data in another, and clinical outcomes somewhere else entirely.

And different teams often manage each part, which makes it very hard to connect everything later. Sometimes these teams are on different continents altogether. These data silos slow down analysis and collaboration. Even when all the data exists, combining it can take months. Sometimes researchers spend more time organizing the files, making sure they have the right versions, than actually doing the research and answering scientific questions.

These silos have created barriers that reduce the effectiveness of the analysis. This is exactly where tools like Bioverse Navigator start to make sense. The issue is not that scientists lack tools. The issue is that these tools do not work across the scale and the silos.

What BioVerse Navigator is designed to do

Bioverse Navigator was launched in December 2024, and it was described as a visual analytics platform that helps researchers explore complex biomedical data in a single environment. Now, instead of moving between different programs and silos and constantly exporting files, researchers can load different data types into one workspace and explore them together.

The big pitch with Bioverse Navigator was that it uses a unified data environment supported by a data orchestration layer. In simple terms, the platform keeps all data organized, connected, and traceable, which is very important so researchers can trust what they are looking at. Biolizard also described the platform as AI-native, meaning machine learning tools are built into the workflows instead of needing additional work.

Ultimately, Bioverse navigator isn’t just for faster analysis but also clearer analysis

How Researchers actually use the Platform

According to Biolizard, this platform focuses on making analysis and visualization coexist together. Researchers can explore patterns in the data visually while running advanced analytics in the same environment. This is great because it allows them to test ideas and see results quickly instead of waiting for long processing steps and then trying to interpret results afterward.

This platform also reduces the barrier for any researcher who does not have a strong programming background. Not every biologist needs to be a data scientist, and not every clinician needs to learn how to code. This tool allows direct interaction with data and can change who gets to ask questions and participate in analysis. I think that’s pretty important because now curiosity gets answered much faster.

Medical research examples that show real impact

The most convincing part of BioVerse Navigator is how it has been used in medical research and development.

Bioverse Navigator’s site published a few case studies that I found interesting:

Prostate cancer research

In one project focused on prostate cancer, Biolizard reports that the analysis identified 187 genes strongly associated with prostate cancer, many of which were not previously well known for their connection to the disease. Now, these genes may not immediately lead to new treatments, but the fact that it narrows down potential targets from a massive data set and speeds up the process for researchers so they can focus their time and effort more effectively is a massive win.

Diagnostics and biomarker panels

Biolizard worked with MDxHealth to develop cancer diagnostics. In this collaborative project, Biolizard built a machine-learning-based risk scoring model using biomarker panels and improved quality control processes for laboratory assays. This type of work is less exciting than gene discovery, but diagnostics depend heavily on consistency and reliability. Improving these workflows and making them more efficient reduces time waste and creates a real impact on patient care.

Predicting transplant rejection

Another case involved predicting kidney transplant rejection using RNA sequencing data combined with clinical variables. Biolizard developed predictive models for early acute rejection and subclinical rejection. This example is a great case study in why integration of data matters, because neither molecular data nor clinical data alone tells the full story. In their platform, they could combine the data and help doctors intervene earlier and tailor treatment more precisely.

My perspective

What interests me is that this innovation was not a single breakthrough algorithm. It was a platform approach to unifying complex data and making analysis more accessible, efficient, and fast. Biological datasets are growing really fast. These tools help researchers stay organized, move faster, and be more efficient. They focus their energy on solving problems and asking the right questions versus dealing with the inefficiencies of organizing and trusting the data.

This research also acknowledges the fact that scientists and researchers spend a lot of time managing data instead of analyzing it. By reducing friction, they can move faster.

At the same time, I think it is important to stay cautious. Visual dashboards can make results look convincing even when they need more validation. Biology is not going to stop generating data. Tools like this show that the future of bioinformatics is not just in gene discovery and smarter algorithms, it is also in building smarter systems that researchers and scientists

As a South Asian American, I have grown up between two biodiversity worlds: the ecosystem of California and the ecological traditions of India. In the Bay Area, I am constantly surrounded by redwood trees, tidepools, and coastal fog, while neem and banyan trees, monsoons, and ayurvedic plants shaped the traditions within my family. I used to think biodiversity existed far away from daily life; however, in reality, biodiversity influences my health, culture, identity, and academic direction every day.

Growing up in California, biodiversity was my first teacher. During the weekends, we frequently went on hikes. Seeing the vast amounts of nature, I always wondered how redwood trees could grow so tall. Later, I learned about xylem, mycorrhizae, and fog-drip networks, but simply being in that environment created a spark of curiosity that still drives me. On special occasions, my parents took me near the ocean. I spent those afternoons watching hermit crabs, sea stars, and chitons in tide pools. Seeing how hermit crabs interacted with other species and their environment taught me about microhabitats and ecological adaptation long before I learned those terms in school. These early experiences shaped the way I approach scientific thinking today.

At the same time, biodiversity shaped my life through culture and values. My parents told me stories about plants and their benefits. I learned about neem, a tree used in traditional medicine for centuries, and the banyan tree, which signifies growth and resilience. Turmeric, amla, ginger, and tulsi weren’t just ingredients—they represented ecological knowledge developed over generations. These stories taught me to not see plants as solely organisms but as sources of healing and cultural meaning.

As I’ve begun thinking more deeply about biodiversity, I’ve realized how strongly it influences daily life. The gut microbiome especially interested me because it is shaped by every part of my routine—diet, weather, exposure to chemicals, stress, and the foods I eat at home. In other words, the dry weather of California, my school activities, and the roti and paneer I eat for dinner all contribute to the biodiversity of my gut microbiome. In turn, the microbiome affects not only physical health but also emotional well-being.

While learning more, I read several studies that made me think differently about this invisible ecosystem inside us. A 2024 Brigham and Women’s Hospital article described the largest and most ethnically diverse microbiome study to date, showing clear microbial signatures linked to Type 2 diabetes. NIH articles helped me understand how bacterial genes, metabolites, and even viruses within the gut can influence disease risk in ways we’re only beginning to map. Harvard Medical School also published work explaining how the microbiome affects metabolic pathways and even neurotransmitter production. Reading these sources made the connection between biodiversity, metabolism, and identity suddenly feel personal

.

This curiosity eventually grew into a project. The more I read, the more personal it became. I learned that South Asians are 4–6 times more likely to develop Type 2 diabetes than Caucasians. Seeing this statistic alongside the research I had read made me wonder whether differences in gut microbial diversity could be contributing to this disparity. I also saw patterns across studies—certain microbial species appeared more frequently in diabetic individuals, and some varied significantly across ethnic groups. That made me want to understand the biology behind these patterns, not just memorize them.

These papers motivated me to begin independent research, where I am now analyzing microbiome datasets using machine learning to understand how biodiversity inside the gut relates to metabolic differences across populations. My culture, family traditions, and scientific curiosity all converged in this question. It felt like everything I grew up with—California’s ecosystems, Indian ecological traditions, and my interest in biology—finally connected in a single place.

All in all, biodiversity has not only shaped my early childhood but continues to influence my health, identity, and scientific direction. Growing up between two biodiversity worlds has shaped who I am and whom I hope to become.

Now AI is stepping into labs and reshaping that entire structure. Scientists are starting to use it to discover medicines faster, cheaper, and sometimes even more accurately.

In this blog, I explore and discuss how AI is changing the entire process of drug discovery, what that means for science and healthcare, and when this technology crosses the line from helping humans to harming them.

The Difficulties of Discovering New Medicine

The closest analogy I can think of for drug discovery is designing a key for a lock. First, scientists have to analyze the “lock,” which in this case is the disease or targeted gene. Then they begin the discovery process: trying to create a key that fits perfectly. Any extra bump or dent and the key won’t work.

Similarly, the chemical formulation must fit perfectly and trigger the right reactions for the gene or the disease. Many drug discovery researchers test thousands of compounds and end up failing along the way. Before a drug can reach humans, it has to pass a series of tests, including preclinical tests in cells and animals. After that, it must be tested over multiple phases of clinical trials in humans. Each stage can take years and cost millions.

And the worst part? If the drug fails later in the process, everything—money, time, and effort—is lost. This is the reason why huge pharma companies like Pfizer, AstraZeneca, or Novartis invest billions annually in their R&D and still face low success rates. The traditional approach is not completely an approach of hope, but it does require a substantial amount of trial-and-error and continuous learning.

AI’s Power in Drug Discovery

AI is not magic. It is a set of algorithms that learn the patterns from data and can infer outcomes based on what it learns. As a result, AI can analyze massive datasets, ranging from chemical structures to disease pathways, predict the outcomes, and find patterns that humans may miss.

According to a Microsoft Industries report, AI will potentially cut early-stage discovery timelines by about 70%, and this report has been proven true several times.

The biotech firm Insilico Medicine reported that they used AI to design a drug for idiopathic pulmonary fibrosis, a deadly lung disease. They reached phase one clinical trials in about 2.5 years, half the time it would normally take them. In this process, not only did they save time and research, but they also saved a lot of money.

Seeing examples like this across the world, I truly believe that AI has the potential to reshape how quickly life-saving drugs move from idea to shelves to the patients.

Now let’s look into how AI actually does it.

This is the part that I find most fascinating. Chemical companies have chemical libraries containing millions of possible compounds. If humans tested each one, it would take them lifetimes. AI, on the other hand, can identify which molecules have the most promising structures in only a few hours.

From there, AI can predict which drugs work. AI can forecast how a compound will behave inside the body, whether it will bind to a target protein, or if it’s toxic. This means a scientist can eliminate inadequate options before wasting time and money on experiments. In other words, AI is a filter.

That’s not all. When existing compounds don’t work, AI can create and design new molecules from scratch. Using a process called generative modeling, AI can create molecules that fit specific criteria—like avoiding certain side effects. There is active research in which AI can even imagine structures that scientists may have never thought of. For example, AstraZeneca reported that their AI systems found 170+ potential antibody drugs in three days, when the traditional methods had found zero after months of searching. Additionally, MIT researchers trained an AI model to search for new antibiotics. The AI was able to recognize patterns and identify compounds that would kill antibiotic-resistant bacteria, an observation no research scientist noticed before (MIT 2020).

AI’s use doesn’t just stop at the early stages.

Once a potential drug is found, clinical trials begin—and this is where most drugs fail. Trials are expensive and time-consuming; any failures at this stage are very expensive. No wonder AI helps here too.

It can analyze available health records and genetic data to find likely patients who are going to respond to a specific drug. By learning those patterns, it can even predict potential side effects earlier. At scale, it can monitor data almost in real time and therefore allow researchers to have enough information to adjust their trial designs faster.

So instead of waiting years to find out a drug’s failure, scientists can make decisions and figure it out in weeks, improving their chances of success.

It is pretty evident that AI represents a revolution in drug discovery. It brings in the speed of turning decade-long iterative and painful processes into a short period of time. Additionally, not only does it decrease cost so money goes far, but it also improves the precision of clinical trials, reducing the rate of trial and error. All of this sounds almost too good to be true, and in some ways, it is. The same speed that makes AI powerful also brings real risks, especially when we don’t slow down and look at the consequences.

The Risks

AI, like any other technology, is not perfect, especially in medicine. It learns from data, and if that data is biased or inaccurate, it learns incorrect patterns. It can also hallucinate and find patterns that don’t exist. We’ve all seen examples where ChatGPT gave us answers that didn’t fully make sense or solved a math problem incorrectly. While that may not seem like a major issue, a mistake in healthcare is the difference between life and death.

Let me walk you through a few examples. If you have an AI system that is trained on genetic data from a certain population, patients from other populations may receive misleading information.

For example, the majority of genetic data studies conducted use information from people of European descent, while individuals from places like Asia or Africa aren’t represented as well. Since AI uses data from existing databases to extract patterns, it can give misleading and incorrect data about groups that hardly appear in the dataset. This kind of biased data can lead to misdiagnoses, wrong dosages, or the lack of care for entire populations. Over time it will not just harm individuals—it widens the health gap, because the groups with more data get better care while others are left behind (The Journal of Global Health, 2025). In fact, in 2019, a healthcare algorithm used in U.S. hospitals was found to assign Black patients similar risk scores to white patients, even though the Black patients were actually sicker. That meant many Black patients received less care and fewer medical resources than they needed (Hopkins Bloomberg, 2019).

Unfortunately, these biases are not just data collection problems; they are also ethical ones. In some of the recent AI battles, we have seen how data privacy, lack of consent, and inappropriate use of data are big ethical issues in the world of AI. These issues carry over into the world of healthcare as well. Some companies have already trained models on patient data without proper consent, sometimes without even realizing it. And the truth is, machines don’t understand ethics; they only process whatever information they’re given.

And then there is the potential black box problem. We hear a lot about deep learning models, but those models are not always debuggable. My Spotify can recommend songs based on my music taste, but it may not always be able to explain why it chose those specific songs.

Now imagine doing this for something as serious as drug discovery or a treatment plan. A patient should be able to understand why a certain drug or treatment plan is being prescribed to them. The lack of transparency and observability in AI models makes that hard. It also makes accountability unclear. If something goes wrong, who’s to blame? The researcher? The doctor who prescribed it? The AI? In healthcare, accountability is very important, and losing accountability also loses the trust of patients and doctors.

And lastly, one of the biggest risks is over-reliance. Let’s assume a future world where AI is predicting and analyzing medicine and discovering drugs, but it doesn’t understand the ethical issues of empathy, pain, or fairness. A human scientist may pause, debate, and discuss a risky drug or take an ethical issue into consideration. AI wouldn’t, however. Relying too heavily on AI can turn medical discovery into a cold, automated system that forgets the human side of patients.

My Viewpoint

I believe we must slow down before we speed ahead. I actually don’t think AI is the problem; instead, the problem lies in how we are using it. I’m a big believer that AI has real potential to change medicine. It can find molecules faster, analyze patterns that we would miss, and predict how certain drugs may react. With proper testing and regulations, I truly believe AI can save millions of lives.

But we can’t ignore the consequences. In healthcare, faster isn’t always better. AI models can learn patterns around drugs and diagnoses but can’t question the ethics behind those methods. Earlier, I mentioned the importance of the diversity of data, and when data is missing, AI may overlook the populations that are most at risk. Similarly, AI may deem a medicine as “safe” in a dataset, even if the targeted group gets sicker from it. This means that with the rise of reliance on AI, others can receive less health care and support than before, and that’s not progress.

And it’s not just bias. Over-reliance is another danger. The more we trust algorithms, the more we stop questioning them. If we let AI do all the work, we lose the skills and judgment of understanding what is happening inside that black box. If AI makes a mistake, we don’t want the doctor to say, “Because the machine said so.” That’s the biggest fear: people forgetting to double-check. In healthcare, we can never afford blind trust.

That’s why I think we need teams of scientists, governing bodies, doctors, and lawmakers working together to pace this rollout, make the datasets more diverse, invest in extensive testing of these systems, and make sure we don’t cut down the processes that have served us for a long time. Because in the end, medicine is a responsibility.

I actually like what the World Health Organization recommended: a human-in-command approach for all AI in medicine systems. This means that the research scientists stay at the center overseeing every step, validating every AI prediction, and being accountable for the final calls. In other words, AI is your assistant; it’s not making decisions on your behalf. This gives room for human creativity and intuition. When a molecule fails, a scientist can still learn something new. AI just moves on to the next problem to solve.

Closing thoughts

Right now, a simple Google search shows that there are over 100 AI-designed drugs in various stages of testing or trials. Many are for critical diseases like cancers, rare genetic disorders, or infectious diseases that haven’t had the resources or the funding for treatment in decades. Even if a very few of these succeed, we could see a future where new medicines are taking 2–3 years from concept to shelves versus the 15 earlier. It is not just faster science; it is a lifeline for patients waiting for a cure. Patients whose diseases were rare enough that it did not make economic sense to invest in their research.

But speed and cost aren’t the whole story. The key unlock in this revolution will be machines and humans working together. So we leverage the strength of computation from AI, leverage them as assistants, but have humans solve for ethics, for observability, and for over-reliance.

That question sounds simple, but it’s actually one of the hardest questions in modern science. Because the line between curing a disease and designing a baby is thinner than most people assume.

What Is Genetic Engineering?

Every cell in our body contains DNA, which holds the instructions for how we grow and function. Genes are sections of DNA that tell our cells which proteins to make. When a gene changes or mutates, those proteins can stop working properly, leading to disorders like cystic fibrosis, Huntington’s disease, or sickle cell anemia.

Gene editing means using tools to change that DNA sequence. The most famous tool is CRISPR-Cas9, developed around 2012. According to Nature (2012), CRISPR lets scientists “cut” DNA at a chosen spot and then “repair” or replace it. It’s cheap, fast, and shockingly precise.

Here’s the twist though: there are two main kinds of genetic editing in humans.

• Somatic editing changes genes in the cells of one person. Think of it as treating a disease in an adult or child—it doesn’t pass to their kids.

• Germline editing changes genes in sperm, eggs, or embryos. That means the change is heritable. Every generation after carries the same edit.
  

The germline one is the real ethical storm. Once you edit a future child’s genes, you’re editing the future of humanity in a tiny way.

CRISPR is powerful, but it’s not perfect. Scientists warn about off-target effects, where CRISPR accidentally cuts the wrong part of DNA, and mosaicism, where some cells get edited and others don’t. There’s also the bigger question: we still don’t understand how every gene interacts with others. So a “fix” for one thing could create a new problem decades later.

In 2024, a Springer Ethics in Biology article called CRISPR “a technology with enormous potential and equally enormous uncertainty.” I think that sums it up perfectly.

Why Scientists Want to Use It

Let’s be honest, gene editing can save lives. Imagine a world without inherited diseases. No more cystic fibrosis, Tay-Sachs, or Duchenne muscular dystrophy.

Take cystic fibrosis for example. It affects around 40,000 people in the United States alone (Cystic Fibrosis Foundation). The disease causes chronic coughing, shortness of breath, and constant lung infections. Many patients need hours of therapy every day just to clear their airways. Some need lung transplants before age 30.

Now imagine using CRISPR to fix that gene before a baby is born. The child could breathe normally. No more hospital visits. No daily chest therapy. Because the fix is in their DNA, their children would inherit healthy genes too. One precise change could end cystic fibrosis in that family line forever.

Right now, scientists are testing gene editing to prevent diseases that are caused by single-gene mutations. For example, if both parents carry the gene for cystic fibrosis, every child has a 25% chance of inheriting it. Editing that embryo before birth could mean the child never develops it—and neither will their future kids.

This isn’t something that is happening in the future, it is happening right now. In 2023, researchers in the UK used CRISPR to fix a rare blood disorder in embryos in a lab (they weren’t implanted). Studies from the Broad Institute show potential cures for over 6,000 single-gene disorders if CRISPR becomes reliable enough.

This is the part where ethics starts to take over. Because once you can fix something deadly, how do you stop people from fixing something they just don’t like?

When It Crosses the Line

There’s a difference between treating and enhancing. Treating means fixing something broken. Enhancing means changing something that works fine, just not “perfectly.”

When scientists talk about treating genetic disorders, most people agree it’s ethical—if it’s safe. But when people talk about changing height, eye color, or intelligence? That’s where it gets messy.

Some people imagine “designer babies.” You could, theoretically, make a baby taller, stronger, or smarter by editing certain genes. The problem? Those traits aren’t simple. Intelligence alone involves hundreds of genes and tons of environmental factors—nutrition, education, childhood care, etc.

Even worse, it could create social inequality. A 2022 Pew Research survey found that 71% of Americans support gene editing to prevent disease, but only 5% support it for cosmetic or performance reasons.

That’s because it feels unfair. What if only rich families could afford to make “genetically superior” kids? What happens to people born naturally? It sounds like the start of a dystopian movie, but it’s actually a real concern.

And it’s not hypothetical. We’ve already crossed that line once.

The CRISPR Babies

In 2018, Chinese scientist He Jiankui claimed he’d created the first genetically edited babies. Two twin girls—nicknamed Lulu and Nana—were born with altered DNA. He said he edited their embryos to make them resistant to HIV by disabling a gene called CCR5.

The scientific world went crazy.

Turns out, He Jiankui had faked approval documents, misled the parents, and didn’t follow safety protocols. Later studies showed the edits were incomplete and might have caused new mutations. No one even knew if the girls would actually be resistant to HIV.

In 2019, Science Magazine reported that He Jiankui was sentenced to 3 years in prison for illegal medical practices. China immediately passed stricter laws banning reproductive gene editing.

That case changed everything. It made the scientific community realize that the line between “helping” and “experimenting” was much blurrier than anyone wanted to admit.

It’s kind of tragic too. Those twins didn’t choose this. They became global experiments without consent. That’s where ethics hit hardest—because in germline editing, the person affected can’t agree to it.

The Big Ethical Debate

Every major medical decision is supposed to follow four principles:

1. Autonomy – the right to make decisions about your own body.

2. Beneficence – doing good for others.

3. Non-maleficence – “do no harm.”

4. Justice – fairness and equality.

Germline editing clashes with all four.

Autonomy: Future children can’t consent. It’s like making a life-changing decision for someone who doesn’t even exist yet.

Beneficence: You might prevent disease, but if it causes new mutations, was it really “good”?

Non-maleficence: Scientists can’t promise safety yet. CRISPR errors might harm not just one person but generations.

Justice: Only wealthy families could afford genetic modification. That could create a new form of inequality—biological inequality.

And then there’s the disability rights argument. Some activists say trying to “remove” disabilities sends the message that disabled lives are less valuable. For instance, deafness or dwarfism aren’t always seen as illnesses—some see them as part of human diversity. Editing them out feels like erasing identities.

So yeah—it’s complicated.

The Slippery Slope and the Eugenics Shadow

This whole topic also reminds people of the past—especially the eugenics movement in the early 20th century. Back then, governments in the U.S. and Europe forced sterilization on people they considered “unfit,” trying to create a “better race.” That’s what happens when science loses ethics.

Now, even though CRISPR is voluntary, it could lead to a softer version of the same thing: people editing their kids to fit a social ideal. Blue eyes. High IQ. Perfect health. If enough people start doing that, diversity could shrink, and society might treat “unedited” people as less than.

I think this is why so many scientists say, just because we can, doesn’t mean we should.

The Two Sides

Let’s be fair. There are two sides.

The Case For Genetic Engineering

First, the positives.

It can prevent suffering. Around 300,000 babies are born each year with sickle cell disease worldwide (WHO, 2023). Many die before adulthood. If CRISPR could fix that gene safely, millions of lives could improve.

It could also reduce the burden of healthcare. The CDC estimates the lifetime cost of treating cystic fibrosis is over $800,000 per person. If gene editing removed the disease completely, it could save billions globally.

Then there’s the moral side. Some ethicists argue that if we have the power to stop pain and we don’t, we’re being irresponsible. This idea—called the beneficence obligation—suggests we actually have a duty to use gene editing, as long as it’s safe.

And finally, every medical technology starts risky. Heart transplants, in vitro fertilization (IVF), even vaccines—all faced backlash at first. IVF was banned in many places in the 1970s, and now millions of kids owe their existence to it. Maybe gene editing will follow a similar path.

The Case Against

Now, the other side.

The unknowns are massive. CRISPR cuts DNA, but our genome is insanely complex. There are over 20,000 genes, and many interact in ways we still don’t understand. A single change could cause cancer, immune disorders, or even mental illness years later.

Also, ethics moves slower than technology. In 2024, over 75 countries still had no clear regulations for human germline editing (UNESCO Bioethics Report, 2024). That means someone could easily repeat what He Jiankui did.

And then there’s inequality. Gene editing could easily become another way for rich people to buy advantages. If one generation of wealthy families edits their children to be smarter or healthier, that gap could grow exponentially.

As Dr. Jennifer Doudna, one of CRISPR’s inventors, said in her TED Talk: “We are at a point where humans can control evolution. That should terrify us a little.”

What I believe

In my opinion, gene editing is one of the most important and promising breakthroughs in modern science. If humanity can truly master it—and handle the ethical part—it could save millions of lives and completely change how we treat disease. I fully support that future. But I don’t think we’re ready yet.

We still don’t know what long-term effects editing can have, and the line between helping and harming can be crossed too easily. On top of that, most countries don’t have strong or consistent regulations. That makes me think this shouldn’t be a large-scale global experiment yet—it should start in limited, tightly monitored steps.

I think gene editing should be tested on a restricted number of cases—enough to study properly, but small enough to control and learn from safely. These studies should be regulated by an international committee, not by individual countries, so that decisions follow shared safety and ethics standards.

Every edited child should be monitored throughout their lifetime to track health, development, and any genetic effects. Because these are babies—they can’t consent, and one mistake could affect not just their life but their descendants too.

Starting small gives humanity time to create fair global rules and learn what works. If results are positive and consistent, then we can expand responsibly. But right now, patience and global cooperation matter more than speed.

I don’t think we should stop gene editing. I think we should work toward mastering it the right way. The potential is real, but so are the risks—and we can’t afford to get this one wrong.

What Comes Next

Right now, most countries ban germline editing for reproduction. The U.S. forbids it using federal funds. The UK allows embryo research but not implantation. China passed new bioethics laws after 2018.

Still, research continues fast. In 2023, the first CRISPR-based therapy for sickle cell disease (called Casgevy) was approved by the UK’s Medicines and Healthcare Products Regulatory Agency (BBC News, 2023). That’s a huge milestone—it’s not germline, but it proves the tech works.

Scientists are also exploring prime editing, an upgraded version of CRISPR that can “rewrite” DNA letters without cutting. It’s like the spellcheck of genetics. If that becomes reliable, it could reduce off-target risks even more.

But ethics will always need to play catch-up.

As Harvard geneticist George Church said, “We need to ask not only what’s possible, but what’s wise.”

The Final Question

At the end of the day, this debate isn’t really about CRISPR. It’s about us.

Every generation invents something that forces humans to redefine what it means to be human. Fire, printing, the internet—now CRISPR.

The potential is huge: curing diseases, saving lives, even extending lifespan. But the danger is real too: inequality, discrimination, or irreversible mistakes.

Maybe the question isn’t “Should we edit genes?” but “Can we do it with enough wisdom, humility, and empathy?”

Because once we start editing the code of life, there’s no undo button.

And maybe that’s what scares me most.

Works Cited

National Human Genome Research Institute. “Ethical Concerns and Genome Editing.” Genome.gov, National Institutes of Health, 2023, https://www.genome.gov/about-genomics/policy-issues/Genome-Editing/ethical-concerns.

“The Pros and Cons of Gene Editing Babies.” The Week, 12 Feb. 2021, https://theweek.com/news/science-health/959606/pros-and-cons-of-gene-editing-babies.

“The CRISPR Babies.” Science History Institute – Distillations Podcast, 2020, https://www.sciencehistory.org/stories/distillations-pod/the-crispr-babies.

Powell, Alvin. “Perspectives on Gene Editing.” Harvard Gazette, Harvard University, 24 Jan. 2019, https://news.harvard.edu/gazette/story/2019/01/perspectives-on-gene-editing.

American Society of Gene & Cell Therapy. “Ethical Issues in Germline Gene Editing.” Patient Education Portal, 2023, https://patienteducation.asgct.org/patient-journey/ethical-issues-germline-gene-editing.

Cystic Fibrosis Foundation. “About Cystic Fibrosis.” CFF.org, 2023, https://www.cff.org/intro-cf/about-cystic-fibrosis.