What If AI Makes Everyone Pick the Same Drugs?

Google DeepMind released something this week called the AlphaGenome Atlas, and the scale of it is kind of ridiculous. The human genome has roughly three billion DNA building blocks. At each position, there are four possible building block that could theoretically be there.

AlphaGenome has now precomputed predictions for essentially all of those possibilities, around nine billion single-nucleotide variants, and estimates how each one could affect things like gene expression, RNA splicing, and other molecular biology.

So, we are getting very close to a world where the number of biological hypotheses we can generate is basically unlimited.

But, having more things to search does not necessarily mean we will search more broadly.

Drug discovery companies are not trying to explore biology for the sake of exploring biology. They are trying to choose projects that could become useful drugs and, usually, viable businesses. That means picking something with enough evidence behind it, some idea of how to drug it, a disease where a clinical trial is possible, and a story investors are willing to fund.

Those are all completely reasonable decisions. The problem is that everyone is could be making those same decisions.

A lot of companies are looking at overlapping public datasets, reading the same literature, using the same disease genetics, and increasingly using some of the same foundation models and prediction tools. If those systems rank the same well-studied targets highly, then you can imagine a feedback loop starting.

Think of a target that already has lots of papers and experimental data. That gives the model more information. The model becomes more confident in that target. So logically, a company chooses it because there is a stronger evidence package and less biological risk. Which means, that investors are more comfortable funding it. More experiments get done, more data are generated, and now the next model has even more evidence pointing toward the same part of biology.

Meanwhile, the weird target with six papers, no crystal structure, a messy assay, and an uncertain disease mechanism might be exactly where something interesting is hiding. Unfortunately, it is also a project that could eat five years and burn $50 million before you figure out whether the basic idea was wrong.

AI does not make that financial problem disappear.

And there is a huge amount of biology sitting on that second side of the equation. One analysis estimated that 4,729 human genes encode potentially druggable proteins. About 74% of them were still not being targeted by an approved, clinical, or preclinical drug. Even among the 94 new drugs approved across the US, Europe, Japan, and China in 2025 with defined mechanisms, only 11 involved targets that had never previously been modulated by an approved drug.[5]

You can see how a feedback loop could develop.

Well-studied biology produces more data. More data can support stronger predictions. Stronger predictions make a project easier to defend. That attracts money and experiments, which produce even more data around the same biology.

Early AI drug discovery has already faced criticism for something similar. Some of the first companies deliberately pursued familiar or “me-too” targets because they needed programs that could advance quickly enough to demonstrate that their platforms actually worked. That was perfectly rational for each company.

Collectively, rational decisions can still create convergence.

And that creates a strange possibility. We could eventually have hundreds of AI biotech companies and thousands of AI-generated drug programs and think that the industry has become incredibly diverse. But if many of those programs depend on the same target families, disease mechanisms, and underlying biological assumptions, we may have diversified the companies without diversifying the science.

There is a completely plausible opposite outcome too, so it’s not all doom and gloom.

It is also possible that if AI makes early exploration cheap enough, small biotechs may suddenly be able to pursue obscure targets and strange mechanisms that nobody could previously justify spending $100 million investigating. The “long tail” of biology might finally become economically accessible.

We don’t know which way this goes yet.

But AlphaGenome makes it easier to see. We are building tools capable of searching more biological possibilities than any generation of scientists before us.

So the limiting factor may no longer be what we can find, but rather, where the money decides to go, in which case, are you really surprised?

Insilico Medicine reported that its AI-discovered pulmonary fibrosis drug, rentosertib, reduced predicted biological age across six different proteomic aging clocks. The analysis used blood samples from 42 patients in its Phase IIa trial, with some models showing reductions of several years. This does not mean the drug has been shown to reverse aging, but it is an unusual example of an AI-discovered drug being evaluated with computational aging biomarkers in humans. [Link]

The FDA appointed Jared Seehafer as its first Deputy Commissioner for Technology and Artificial Intelligence. He will lead FDA-wide strategy around AI, software and technology, giving the agency a senior position specifically responsible for how these technologies are used across its work. [Link]

Japan launched JapanFold, a domestically hosted platform giving researchers access to a collection of open AI models for drug and protein discovery. Built by ai& and Tenstorrent, it includes models for protein structure, binding prediction and de novo design, with all computation kept inside Japan. It is an interesting example of “sovereign AI” starting to extend into drug discovery infrastructure. [Link]