Google researchers have shown that AI-designed binding proteins can be aimed at large, complex targets for a fraction of the usual compute bill — provided you shrink the target first. Their trick, which they call epitope excision, keeps only the patch of the protein where binding will occur instead of folding the whole molecule. With a Gemini 3.1 Pro agent directing the search, the team produced binders against FAP, a protein associated with solid tumours, measuring affinities between 3 and 9 nanomolar in laboratory assays. The work is detailed in a technical report dated 14 October and in a Google Research blog post.
- Cheaper checks: Folding an 85–130 amino-acid window instead of the full 719-residue FAP target cuts each validation from roughly 59 seconds to about 7 — an 88% reduction.
- Accuracy holds up: Across 3,666 designs, the median structural difference at the contact interface was just 0.37 ångströms, with 88% of cases under one ångström.
- Agents beat solo work: A researcher paired with the Gemini agent achieved a 9.33% hit rate versus 5.06% alone, while spending only 58% of the compute budget.
- Lab validation, with caveats: Four designed molecules hit nanomolar affinities, but none has been tested in solution, in cells or in animals, and the report has not been peer reviewed.
Why trimming the target changes the economics
Standard computational binder design generates thousands of candidates and uses a structure predictor — here AlphaFold 3 — to check which ones actually dock onto the target. That validation step dominates the compute bill, and its cost scales with the size of the protein being folded.
A widely circulated headline claimed Google had accelerated protein structure prediction by a factor of 8.6. That framing is imprecise: the predictor is unchanged. What changed is the input. By restricting the fold to a window around the intended contact site, the team reduced per-check time on FAP from about 59 seconds to 7 seconds, the 88% saving the blog rounds up to 8.6×.
The gains are not universal. In a benchmark of 15 targets, seven showed no improvement at all, because their complexes were already compact enough. Excision only pays off on large, multi-domain targets. There is also a floor below which trimming backfires: in a sweep across nine of the targets, cuts under 85 amino acids preserved correct geometry only 32% of the time, against 97% in the 85–200 range — though much of that collapse traced to a single target.
Precision, ranking and the autonomous agent
Accuracy loss appears modest. Comparing predicted structures from trimmed and full-length targets across 3,666 designs spanning all 15 targets, the median deviation at the binding interface was 0.37 ångströms, with a mean of 0.53 and 88% of cases falling below a single ångström. Ranking quality does degrade slightly: a standard scoring metric dropped 0.048 points on average and worsened on one of the 15 targets. The team frames that as a small price for running far more checks on the same budget.
The more revealing experiment compared three ways of running an identical campaign against FAP on roughly 133 GPU hours. A lone researcher reached a 5.06% success rate while probing four regions of the protein. A researcher working alongside the agent reached 9.33% — an 84% improvement — on 58% of the budget. The agent operating unsupervised managed 6.18%, but explored 13 distinct regions and stopped itself at 94% of its allowance.
Notably, the agent arrived independently at the same strategy as its human counterpart: blocking the entrance to the channel through which FAP does its work. Its action set was deliberately narrow — propose strategies, widen the search, go deeper, backtrack, stop — yet that proved enough to converge on a chemically sensible answer.
What the wet lab showed, and what it did not
The team built a library of 5,000 candidates, 1,250 per strategy, displayed them on the surface of yeast cells and ran three progressively stricter selection rounds against FAP. Screening produced 60 exact matches with the computational designs; four of those measured affinities between 3.0 and 9.1 nanomolar, a strong result by industry standards. Of the 60 matches, 24 (40%) came from the autonomous arm and 8 (13%) from the human-only arm. On library survival, the hybrid approach led with 0.83 against 0.50 for the autonomous agent.
The limitations deserve equal weight. The molecules were not characterised in solution, nobody has measured whether they block FAP activity, and nothing has been tested in animals. None of this demonstrates a therapeutic effect against cancer. The report is a Google technical document, not a peer-reviewed paper, and neither the excision method nor the accompanying lab tooling has been released — unlike AlphaFold 3, which Google open-sourced in 2024. The authors add their own caveats: the autonomy comparison used a single target, a single structure class and a single model, and folding trimmed proteins without evolutionary information performs worse when that signal is scarce.
The blog also outlines two adjacent efforts: a verification layer for robotic laboratories that blocks physically impossible instructions, and DNA aptamer design campaigns in which agents discovered unwanted shortcuts — exploiting sticky structures that bind almost anything — before self-correcting in later rounds.
The headline number is the least interesting part of this work. The real insight is that the bottleneck in AI protein design is not creativity but the invoice, and that the invoice shrinks when the problem is framed better rather than solved with more silicon. Cutting a protein down to the residues that matter is the molecular equivalent of refusing to reread an entire book to check a single page. The autonomous agent’s result also merits attention for a subtler reason: it was not the most accurate arm, yet it covered the most ground and contributed the largest share of confirmed lab hits. That suggests its value lies in probing regions a human would dismiss on instinct. Still, until peer review, published code and independent replication arrive, this remains a demonstration by Google, about Google’s own work.