Protein science now runs on AI pipelines, and their choices decide the result

Agents that run the experiment, not just read about it

Three preprints in this batch describe the same shift from different corners of protein science: AI pipelines are being handed the decisions inside the workflow, not just asked to summarize papers. The systems differ in their targets, but they share a design instinct. Each one narrows what the pipeline is allowed to assume, gives it a structured way to check its own work, and then measures whether the constrained approach beats the unconstrained default. In each case reported so far, the constrained approach wins, and the margin is not close. m:…7064 m:…1745 m:…6629

The clearest demonstration comes from LPT, a framework for automated protein binder design that organizes language-model workflows into bounded agentic loops and connects them to a curated literature database. On specialized test questions, the system answered 41 of 48 correctly with database access, compared with 0 of 48 for models without it. That gap is the paper's central result: the database and the loop structure, not a larger model, supply the specialized reasoning. The framework then ran end-to-end design campaigns that produced candidate binders for three therapeutic targets, CALCRL/RAMP1, PD-1/PD-L1, and PVR-TIGIT, with what the authors report as high success rates in computational testing. m:…7064

The scope of that result is set by what the preprint reports. LPT improved specialized scientific reasoning by using bounded agentic loops and a curated literature database, scoring 41 of 48 on specialized test questions with database access against 0 of 48 without it, and it successfully designed candidate binders for CALCRL/RAMP1, PD-1/PD-L1, and PVR-TIGIT. Those are candidate binders produced by the design framework, and the reported result is that design outcome itself, not a claim beyond it. m:…7064

Microscopy gets the same treatment

CryoPilot applies the same idea to cryo-electron microscopy, replacing static processing pipelines with two language-model agents that evaluate intermediate data and adjust processing parameters as they go. The system takes raw micrographs through to atomic models, handling job selection, parameterization, and iterative strategy testing with minimal human intervention. Across nine tested public EMPIAR datasets, it reached resolutions between 2.61 and 3.51 angstroms on eight of them, and on a difficult alpha latrocrustotoxin dataset it improved the reconstruction from 5.9 angstroms to 4.15 angstroms by autonomously changing its particle picking strategy. m:…1745

The reported results show the adjustment mechanism working, not just the headline resolutions. CryoPilot automates cryo-EM structure determination from raw micrographs to atomic models using two language model agents, and on a difficult alpha latrocrustotoxin dataset it improved the reconstruction from 5.9 angstroms to 4.15 angstroms by autonomously adjusting particle picking. Across nine tested public EMPIAR datasets, it reached resolutions between 2.61 and 3.51 angstroms on eight of them, so the autonomous adjustment is tied to measured outcomes across the tested set rather than to a single case. m:…1745

Even the input structure is a decision

The third preprint shows that the same lesson holds where no agent is involved at all: the input to the pipeline is itself a choice. The authors found that 62% of AlphaFold-predicted binding pockets exhibit holo-like characteristics, and that docking success rates improve by 19 percentage points when the pocket state is correctly identified as holo-like or apo-like. The catch is that AlphaFold's own confidence scores report pocket variability but do not indicate which state a pocket is in, so a workflow that treats a predicted structure as a single static answer is silently making a decision it never examined. m:…6629

Read together, the three preprints point to a near-term shape for AI-driven protein work. This suggests the winning pattern is to bound the agent's loop, ground it in a curated source, and check the state of a predicted structure before using it, rather than treating any single model output as a finished answer. This suggests the reported gains — 41 of 48 against 0 of 48 for LPT, resolutions between 2.61 and 3.51 angstroms on eight of nine datasets for CryoPilot, and a 19-percentage-point docking gain when pocket state is correctly identified — are evidence that constrained pipelines can run protein workflows, not proof on their own that they have replaced the scientists who check them. m:…7064 m:…1745 m:…6629