acceptodds
Under review as a conference paper at ICLR 2027

Auto-Research Agent for Target-Adaptive MSA Organization to improve Protein-Protein Complex Structure Prediction

Abstract

Protein-complex structure prediction with AlphaFold 3 depends on MSA content, depth, and cross-chain pairing. Current pipelines apply uniform recipes, typically maximizing depth and pairing by species, yet even whether cross-chain pairing helps AF3 remains contested. We show that the value of both depth and pairing is strongly target-dependent, so no single recipe is optimal across interfaces. We introduce an autonomous agent that treats a frozen AF3 as a black-box experimental instrument. For each target, the agent formulates biological hypotheses, analyzes and reorganizes the MSA, and iteratively refines sequence selection and pairing using predictor label-free signals alone. On 1015 protein–protein interfaces from three benchmarks released after AF3's training cutoff, the agent improves overall interface DockQ over AF3's default pipeline and fixed MSA recipes on every benchmark, while using on average 60% fewer MSA sequences than the default. Gains are largest for inter-species heteromers, where species-based pairing is unreliable: mean DockQ increases from 0.465 to 0.612 on FoldBench and from 0.482 to 0.562 on HD439. The improvements arise from both sequence construction and pairing policy, including deliberate non-pairing. These results recast MSA design as a target-specific problem and show that an autonomous agent can adapt it at test time through biological reasoning and self-directed, confidence-guided experimentation.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
MSA Construction, AI Agents, Test-Time Adaptation, Protein Complex Structure Prediction, AlphaFold 3, AI for Science

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_68oeOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper demonstrates a useful agent-based workflow for improving AF3 predictions through adaptive MSA organization. However, the current evaluation establishes the effectiveness of the overall pipeline more clearly than the specific contribution of LLM-driven reasoning. Budget-matched search baselines and component ablations are needed to substantiate the claimed agentic contribution.

Critical strengths
  1. Well-motivated problem. The paper addresses the target-dependent effects of MSA depth and cross-chain pairing, an important practical issue in protein-complex structure prediction.

  2. Practical, training-free approach. The agent improves a frozen AF3 predictor using prediction-derived feedback, without requiring ground-truth structures during optimization.

  3. Broad empirical evaluation. Results cover 1,015 interfaces across three benchmarks, with consistent aggregate improvements and particularly strong gains for inter-species complexes.

Critical weaknesses

Missing budget-matched search baselines. The agent performs multiple AF3 evaluations, whereas the main baselines use fixed MSA recipes. Comparisons with random search or a fixed candidate portfolio under the same prediction budget are needed to establish the value of LLM-driven adaptation.

Limited attribution of the agentic contribution. The paper does not isolate the effects of biological reasoning, reflection, memory, and feedback-driven proposal generation. Consequently, it remains unclear which components are necessary.

Efficiency claims require qualification. Reduced MSA depth does not establish lower end-to-end computational cost, given the repeated AF3 evaluations and additional LLM calls.

Confidence-based selection remains unreliable. On FoldBench-v1, final nomination loses approximately 41% of the improvement discovered during exploration, highlighting a substantial mismatch between predictor confidence and structural accuracy.

Some comparisons are not input-matched. On the FoldBench datasets, the agent and the mMSA/pMSA/uMSA baselines use different MSA resources, so these comparisons do not cleanly isolate construction strategy.

Questions for authors
  1. Missing budget-matched search baselines. The agent performs multiple AF3 evaluations, whereas the main baselines use fixed MSA recipes. Comparisons with random search or a fixed candidate portfolio under the same prediction budget are needed to establish the value of LLM-driven adaptation.

  2. Limited attribution of the agentic contribution. The paper does not isolate the effects of biological reasoning, reflection, memory, and feedback-driven proposal generation. Consequently, it remains unclear which components are necessary.

  3. Efficiency claims require qualification. Reduced MSA depth does not establish lower end-to-end computational cost, given the repeated AF3 evaluations and additional LLM calls.

  4. Confidence-based selection remains unreliable. On FoldBench-v1, final nomination loses approximately 41% of the improvement discovered during exploration, highlighting a substantial mismatch between predictor confidence and structural accuracy.

  5. Some comparisons are not input-matched. On the FoldBench datasets, the agent and the mMSA/pMSA/uMSA baselines use different MSA resources, so these comparisons do not cleanly isolate construction strategy.

Reviewer_rdKMOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents an autonomous agent that edits multiple sequence alignments for individual protein complexes while treating a frozen AlphaFold 3 predictor as an experimental instrument. The agent writes analysis and construction programs, proposes sequence selection and pairing strategies, interprets predictor-native confidence and structural outputs, and nominates a previously evaluated MSA.

Critical strengths
  1. Target-specific input adaptation is a compelling application.** The method can improve an existing structure predictor without retraining it.
Critical weaknesses
  1. I think the more convincing baseline would be same AF3 prediction with / without agent search. Evaluating only 25 structures for the final nominated MSA controls the final ensemble size, but not the opportunity to search over multiple MSA configurations. A fixed portfolio, randomized MSA-edit search, or simple sequential optimizer using the same sequence pools, AF3 calls, and confidence-based selection would directly test the central agent claim.

  2. The paper claims workflow savings by reducing the MSA depth (Table 2), but it ignores the LLM calls, the additional AF3 calls, which I think cause much more time and computational budget than reducing MSA depth.

  3. Can you gain some clues on what the agent prefers by analyzing the CoT from your experiments, such as preferring the clash over iptm

  4. What is the performance on intrinsically disordered regions (IDRs) where the confidence of that region is always low?

Questions for authors

See above

Reviewer_LrNLOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper assesses the ability of general purpose LLMs (specifically GLM-5.3) to reason over MSA row selection criteria for AlphaFold3 to improve multimer structure prediction. They develop a workflow in which GLM-5.3 iteratively reasons over the MSA using confidence metrics from AF3 (so there's no leakage from the ground truth) to propose an MSA that pairs and unpairs specific sequence rows. The approach shows mostly marginal but consistent gains over default MSA construction procedures.

Critical strengths
  • The experimental design appears sound, and would likely generalize to other LLMs. This is exciting since as LLMs improve, it is likely that this approach would improve out of the box. I am already curious how it would perform using frontier proprietary models like ChatGPT or Claude.
  • The improvements reported are consistent and statistically significant, so they appear believable.
Critical weaknesses
  • The gains are marginal at best. The most substantial are for heteromeric inter-species complexes, which are a key challenge for AF3-based models
  • It wasn't entirely clear how much more compute is spent per round. While the final result is derived from the same fixed number of 25 samples across methods, the extra compute used to propose a better MSA could also go into random sampling, and it would be important to have an unambiguous control demonstrating apples-to-apples comparison of gains vs. compute expended.
  • Not a critical weakness, but table placement is rather suboptimal throughout the paper (Table 2 before 1, and in the appendix tables are not placed near their content)
Questions for authors
  • My main concern about this paper is the very marginal gains reported. I would suggest that individual, per complex gains be shown as a scatter plot, as some appear substantial (8P9C example). It is far more meaningful to get substantial gains on some targets than very marginal gains on all targets, and so if it is the former it would strengthen the paper to include such a figure.
  • What would happen if the protein complex is actually novel? The authors conclude that LLMs are using biological priors to improve their prediction, which is exciting and a reasonable explanation, but does suggest that this approach may work poorly in prior-poor contexts, and it would be good to understand its limitations in such contexts.
2 comments

Thank you for the additional points. That's helpful to know. I am not going to change my score, because the issue about overall compute cost remains the primary one, especially seeing the responses to the other reviewers where it's clear that this remains extremely impractical atm, given the number of AF3 evaluations required.

Thank you for your response. In practice, our agent-based method requires an average of 6–7 AF3 inference runs and slightly over 100 LLM calls per target. We believe that this computational cost is justified in scenarios where AF3 Default produces clearly anomalous predictions, or where improving the robustness of structure prediction is particularly important. For example, in binder design, we would want to avoid discarding potentially promising binders simply due to inaccuracies in the structure prediction model. In such cases, we believe our approach offers a worthwhile trade-off between computational cost and prediction reliability.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.