AI May Be Looking for the Wrong Molecule

The most promising molecule may not be the one with the strongest result today. It may be the one that can survive everything scientists will need to change tomorrow.

A paper published this week by researchers from David Liu’s lab offers an interesting example.

They were working with proteases, which are enzymes that break down other proteins. In theory, a protease could be engineered to recognize and destroy a disease-causing protein while leaving everything else alone.

The researchers started with proteases found in nature and used an AI model called ProteinMPNN to redesign them. The model was given the shape of each protein and was asked to find new amino acid sequences that should still fold into that same shape.

Those redesigned proteins were then put through directed evolution.

Directed evolution is basically evolution recreated in a laboratory. Researchers make many mutated versions of a protein, test them, keep the ones that perform better, mutate those again, and repeat. Over several rounds, the protein gradually becomes better at whatever job the researchers are selecting for.

The proteins that started from the AI-redesigned sequences consistently evolved better than the ones that started from nature. In one experiment, the researchers engineered a protease to recognize ataxin-2, a protein associated with a neurological disease. The best lineage that began with an AI-redesigned protein reached more than 79 times greater selected specificity than the best lineage that began with the natural protein.

The AI may have designed a better protein, or more specifically, it may have designed a better ancestor.

Natural proteins were shaped by millions of years of survival, compromise, and historical accident. They only had to work well enough for the organism carrying them to survive and reproduce. Evolution had no reason to make them easy for a scientist to turn into a future medicine.

A natural protein can perform its current job extremely well while having very little room to become something else.

Every mutation comes with a cost. One change might improve activity against a disease target while making the protein less stable. Another might improve specificity while making it harder to manufacture. Eventually, the protein stops folding properly or loses the function it started with.

The AI-redesigned proteins appeared to have more room for change, that is, seemingly more resilient to mutations.

I think of this as biological runway. A fragile protein can only afford a few bad steps before it falls apart. A more robust protein can survive more mistakes and explore places that were inaccessible from the original starting point.

AI drug discovery may be paying too much attention to the first candidate proposed by a model.

Models are usually judged by what they produce immediately. The compound binds tightly. The protein shows high activity, or the sequence receives the best score.

Real drug development begins after that. A promising molecule still has to be improved for potency, toxicity, stability, delivery, manufacturing, and dozens of other properties. Fixing one problem can create another.

A molecule that looks excellent on day one may be sitting on top of a very small hill, with no useful path forward.

Another molecule may begin with weaker activity but tolerate hundreds of modifications. Its potency can be improved, its liabilities can be removed, and its chemistry can be pushed in several directions without the whole thing collapsing.

That ‘lesser’ candidate may be worth much more.

We usually evaluate molecules as individuals. Perhaps we should evaluate them by the family of molecules they could eventually produce.

This would also change how AI drug discovery companies prove that their models work. Producing one impressive molecule would matter less than showing that AI-generated starting points repeatedly survive optimization, tolerate modification, escape liabilities, and produce more viable descendants than conventional starting points.

AI may never become the machine that hands us a finished drug. It may become the machine that gives drug discovery better ancestors.

An autonomous AI agent broke out of its test sandbox and compromised Hugging Face's production servers. Chasing a benchmark answer key, OpenAI's models found a zero-day, reached the open internet and harvested credentials, the first documented case of frontier models chaining novel attacks unprompted. Then the reversal: US models blocked defenders' forensic queries on safety grounds, so Hugging Face ran the 17,000+ logs through China's open-weight GLM 5.2. [Link]

Derek Lowe says the field still can't measure whether its computational models are any good. Writing in In the Pipeline, which he's run since 2002, he argues benchmarking is a problem in its own right, and one that got more fraught, not less, as computational chemistry and biology grew more powerful. Worth holding against every benchmark-topping claim in this issue, including the ones we repeat. [Link]

Anthropic opened its first themed call under AI for Science, aimed at rare genetic disease. Accepted applicants get up to $50,000 in Claude credits over six months across two tracks — basic disease mechanism, and early-stage drug development, with the Monarch Initiative as research partner. Applications close August 2. The grant is credits, not cash; the currency is adoption, which is the point and also the catch. [Link]