A protein model reached near-experimental error—then stumbled on engineered mutations
KaML inferred the acidity of naturally evolved protein sites from sequence alone, but its advantage weakened on 89 deeply buried engineered mutations. The contrast shows where evolutionary training signals help—and where structure still matters.
The short version
- What happened
- Across four natural-protein benchmarks, the sequence-only KaML-ESM models reported root-mean-square errors of about 0.5 pH units for six ionisable amino-acid types.
- Who it affects
- Scientists could use the open sequence-first tool to triage candidate sites before slower structural modelling or laboratory work, provided its uncertainty is not mistaken for evidence that a protein will be safe or effective.
- What remains unproven
- The study is computational and retrospective; it does not test whether KaML improves prospective protein design, drug discovery or laboratory productivity.
Keep in mind
The study is computational and retrospective; it does not test whether KaML improves prospective protein design, drug discovery or laboratory productivity.
Synthetic pretraining labels were generated by an earlier structure-based model, so they add scale rather than independent experimental truth.
The 89 engineered buried mutations were a clear weakness, limiting claims about novel designed proteins outside evolutionary patterns.
The human-proteome scan proposes functional sites and mechanisms; those remain hypotheses until experimentally validated.
Multi-source analysis · 2 linked sources. Primary analysed source date: 24 October 2026. Earliest linked supporting source: 15 July 2026.
Keep following the evidence
Join The Impact Brief founding list for selected AI reports, their effect on people and the evidence limits. Read a sample first.
Read the full report · 4 min The evidence, methods, implications and limitations
Sequence carried more electrostatic information than expected
Researchers at the University of Maryland trained KaML-ESM2 and KaML-ESMC to estimate a residue's pKa—the pH at which it changes protonation state—from an amino-acid sequence. That value affects protein charge, stability, binding and catalytic behaviour. Across four benchmarks of natural proteins, the models reported root-mean-square errors of about 0.5 pH units for aspartate, glutamate, histidine, cysteine, tyrosine and lysine, generally outperforming the structure-based KaML-CBTree and the empirical PROPKA3 comparator.
The pipeline did not learn from experimental measurements alone. It first used a synthetic GAINES set containing 29,457 predicted pKa shifts across 9,945 Protein Data Bank proteins, then fine-tuned separate acid and base models on the experimental PKAD-3r collection. A temporal external set contained 55 residue measurements published after the database cut-off. This design tests some generalisation, but the synthetic labels came from the team's earlier model and are not new laboratory measurements.[1][2]
Engineered sites exposed the boundary
The most useful negative result came from 89 engineered residues buried inside staphylococcal nuclease variants. These mutations lack the evolutionary context that a protein language model learns from natural sequences, and KaML's performance deteriorated. The paper therefore supports a bounded claim: sequence embeddings can encode electrostatic regularities shaped by evolution, but they do not replace explicit structural or physics-based analysis when engineers create unusual local environments.
For protein engineers and drug-discovery teams, fast sequence-only screening could narrow candidates before costly simulation or experiment. It cannot establish that a designed protein will fold, bind or work in a cell, and a pKa error is not a clinical outcome. Confidence would rise with preregistered tests on independently designed proteins, prospective wet-lab measurements and comparisons that hold training data and compute constant. The work was NIH-funded; Biohub supplied compute credits for ESMC, and the authors declared no competing interests.[1][2]
The distinction that matters
Where the evidence is strong—and where it breaks
The same model faces progressively less familiar evidence.
Learn
Combine experimental pKa records with synthetic labels
29,457 synthetic residue shifts from 9,945 proteins
Check
Test natural proteins and 55 later measurements
About 0.5-unit RMSE across four natural-protein benchmarks
Stress test
Predict 89 deeply buried engineered mutations
Accuracy drops when evolutionary support is missing
Watch the background, then check the evidence
Illustrative photo · not a video frame
Google DeepMind · 1 min 52 sec · Background explainer
Protein folding explained
YouTube loads only when you choose; playback starts when you press play.
A short introduction to why a protein’s shape matters. It provides background for understanding what AI structure prediction contributes to biological research.
Evidence check: This is the developer’s explanation. Predicting a structure does not by itself establish a treatment’s safety or benefit for patients.
Published 2020-11-30. Source checked 14 October 2026. Check captions on the publisher’s video page.
Explore current researchContinue the story
Read next
Continue exploring drug discovery: compare the finding and its evidence limits.
Can protein designers steer structure with ordinary language?
A new preprint says yes: Prompteina was trained on 1.8 million structures and more than 50 million captions, led baselines on 15 of 18 structural tasks, and produced nine apparent binders among 89 GM2A designs tested. The result is promising but remains a single-target, unreviewed proof of concept.
Read the next reportExplore the full source and evidence record
What this means for people
- Scientists could use the open sequence-first tool to triage candidate sites before slower structural modelling or laboratory work, provided its uncertainty is not mistaken for evidence that a protein will be safe or effective.
Global context
The work comes from a US academic team and uses public protein databases and globally developed ESM models. Its released platform may broaden access to electrostatics screening, but experimental facilities and representative protein data still shape who can validate the predictions.
What the evidence does not yet show
- The study is computational and retrospective; it does not test whether KaML improves prospective protein design, drug discovery or laboratory productivity.
- Synthetic pretraining labels were generated by an earlier structure-based model, so they add scale rather than independent experimental truth.
- The 89 engineered buried mutations were a clear weakness, limiting claims about novel designed proteins outside evolutionary patterns.
- The human-proteome scan proposes functional sites and mechanisms; those remain hypotheses until experimentally validated.
What to watch next
- Prospective measurements on newly designed proteins selected before results are known
- Independent replication on natural and engineered proteins outside PKAD-3r
- Matched comparisons with structural and physics-based tools under equal compute budgets
- Evidence that faster screening improves an end-to-end discovery outcome rather than only benchmark error
Living evidence record
Impact record IAI-00FW0D7
Evidence stage
Studied
Confidence
Supported
Reporting basis
Multi-source analysis
Independent or research support
Present
Record status
Monitoring
Last checked
24 October 2026
Source trail
2 direct sources across 1 source type.
People impact
Documented in this record.
Uncertainty
Limits and next checks are explicit.
Stages describe the evidence available—not whether a technology is good or bad. See the public method.
Evidence trail
Sources used for this report
Links checked 24 October 2026
Research question, topics and regions
Whether protein language models can predict residue ionisation constants from amino-acid sequence alone
Topics: Protein language models · Computational biology · Protein engineering · Drug discovery · Scientific AI
Regions: United States · Global
This report is labelled multi-source analysis. We summarise and analyse source material in our own words; company statements remain attributed claims until independently supported. Translated summaries preserve the meaning of the original source and link back to it. Read our editorial standards.
Reader commentary
What does this mean where you are?
Tell us what you have seen at work, in your community or in everyday life. A question is welcome too. No account is needed; we review notes before publishing them.
Explore commentary across the portal →Published reader notes
0No published reader notes yet. You can start the evidence-led discussion above.
Prefer a private correction or response? Contact the newsroom.