AI for science this autumn: genomes, watermarked proteins and prime gaps

Product launches get most of the attention, but the long-term case for AI rests on whether it can help us discover things we did not know. September and October 2026 produced several serious examples. They are at very different stages, from a published dataset that others are already using to a candidate drug still being tested in the lab, so I have noted the status of each.

1. AlphaGenome Atlas: scoring every DNA letter change

Announced: 8 September 2026, Google DeepMind. Status: published dataset, already used in research.

Your genome is about three billion DNA letters long, and any single letter can in principle be swapped for one of three others. Most of these changes do nothing, some cause disease, and for the vast majority nobody knows which. That is especially true in the large non-coding regions, which do not code for proteins but help control when and where genes switch on.

AlphaGenome Atlas is DeepMind's attempt to fill that gap. It provides predictions for about 9 billion single-letter variants, effectively every possible change of one letter, together with a new summary measure called the AVI score. The dataset is about one petabyte.

DeepMind points to two early results:

• In work involving the University of Exeter and UK Biobank data, researchers found 22% more associations in non-coding regions than before. That is a British contribution to an important resource.
• The atlas helped identify a variant in the DNM1 gene linked to epilepsy.

Why it matters: genetic testing often returns "variants of uncertain significance", results that neither confirm nor rule out a problem. Better predictions could help clinicians and researchers decide which variants to investigate first. These are still predictions. Clinical use will need careful validation.

2. SynthID Bio: watermarking AI-designed proteins

Announced: 30 September 2026, Google DeepMind. Status: wet-lab tested, methods open-sourced.

AI can now design new proteins, which is promising for medicines and enzymes and raises biosecurity concerns. If an AI-designed protein turns up somewhere it should not, can anyone tell where it came from?

SynthID Bio applies DeepMind's watermarking approach, already used for AI images and text, to proteins. A detectable signature is built into the design, and the protein is meant to keep working. DeepMind says it tested this in the wet lab on proteins including VEGF-A, the SARS-CoV-2 receptor-binding domain and PD-L1, all well-studied molecules relevant to medicine. It has open-sourced the methods.

Why it matters: traceability is one of the few practical safety tools for AI biology that does not stop beneficial research. Open-sourcing it lets others adopt it, test it and try to break it, which is how any watermark earns trust. A watermark only helps if the people designing proteins use it, so expect a debate about whether it should become standard.

3. Agents hunting for new antibiotics

Announced: 15 October 2026, Google DeepMind. Status: early, a candidate in wet-lab validation.

Antibiotic resistance is one of the slow-moving crises in medicine. Antimicrobial peptides, short protein chains that can kill bacteria, are a promising area of research, but the number of possible peptides is enormous.

DeepMind described an agent-driven search that used its ProtNLM protein language model to explore that space. One candidate it identified is now in wet-lab validation. That is an early stage, and most candidates in drug discovery never become medicines. What is new is the method: an AI agent running much of the search cycle itself.

Why it matters: if agents can reliably reduce a huge search space to a short list worth testing in the lab, the slowest part of early discovery speeds up. That is a real benefit even if this particular candidate fails.

4. Maths: tightening the gaps between primes

Announced: 3 September 2026, OpenAI (in the GPT-6 Astra launch). Status: company claim. The mathematical community will make the final judgement.

The "bounded gaps between primes" result says there are infinitely many pairs of primes no more than a fixed distance apart. Mathematicians have spent years reducing that distance. OpenAI says GPT-6 Astra helped bring the bound down from 240, the previous best due to Julia Stadlmann, to 186.

Why it matters: a better bound on a famous problem is the kind of result that cannot be faked with clever wording. Either the proof works or it does not. If it stands up, it is one of the clearest examples so far of a language model making a contribution that mathematicians would recognise. Mathematicians will want to check it carefully first.

There are also reports of AI formal proof search making progress on open Erdős problems. I have not yet checked the primary paper, so I have left the details out.

The common thread: verification

Each of these results is only as good as the checks behind it. Genome predictions need clinical validation. Watermarks need to hold up against attempts to remove them. Antibiotic candidates need lab results, and proofs need review. What looks different from a few years ago is that AI is producing outputs specific enough to be tested, rather than plausible-sounding text.

There is a safety parallel too. The same autumn that brought these results also brought the Hugging Face incident and talk of pacing the frontier. SynthID Bio shows a lab building a safeguard into a scientific tool from the start. For more on these and other results, see my AI Research page.