An Autonomous Agent Improves AlphaFold 3 by Redesigning MSAs

Introduction

The quality of AlphaFold 3 predictions for protein complexes depends not only on the model, but also on how its multiple sequence alignments, or MSAs, are assembled. Standard pipelines often follow a uniform recipe: collect as many sequences as possible and pair chains according to species. That approach is convenient, but it assumes that greater depth and more pairing are consistently useful. For protein interfaces, the answer may depend strongly on the target.

This study explores a different strategy. Instead of retraining AlphaFold 3 or changing its weights, the researchers place an autonomous agent around a frozen model. The agent treats AlphaFold 3 as a black-box experimental instrument: it forms biological hypotheses, reorganizes candidate sequences, runs new prediction trials, and uses label-free signals from the predictor to refine the next MSA configuration.

Key findings

• There is no universal MSA recipe. The value of alignment depth and cross-chain pairing varies across interfaces. A configuration that helps one target can be unhelpful for another.
• The agent searches over MSA organization. Its actions include selecting different sequence subsets, changing pairing policies, and deliberately leaving sequences unpaired when pairing is likely to be unreliable.
• Accuracy and efficiency improve together. Across 1,015 protein–protein interfaces from three benchmarks released after AlphaFold 3’s training cutoff, the approach outperforms the default pipeline and fixed MSA recipes on overall interface DockQ. It also uses about 60% fewer MSA sequences on average.
• Inter-species heteromers benefit most. Species-based pairing is particularly fragile when interacting chains come from different organisms. On FoldBench, mean DockQ rises from 0.465 to 0.612; on HD439, it rises from 0.482 to 0.562.

Why it matters

The main contribution is a shift in how MSA preparation is understood. It is often treated as a static preprocessing step whose objective is to maximize the amount of evolutionary information. The results suggest that MSA construction is closer to experimental design. More sequences do not automatically provide better evidence, and an incorrect pairing pattern may introduce misleading relationships into the model input.

The study also offers an example of test-time research for scientific foundation models. The underlying predictor remains unchanged, while an agent conducts a targeted loop of hypothesis formation, model evaluation, and strategy revision. This could be useful when retraining is expensive or when the best input configuration is expected to depend on each biological target.

The findings should not be read as proof that the agent fully understands interaction biology. Its success still depends on the available sequence candidates, the signals exposed by the predictor, and the properties of each benchmark target. A next step would be to connect specific biological characteristics with the pairing policies selected by the agent, making the search more interpretable and less empirical.

Source: Hugging Face Daily Papers