Auto-Research Agent on MSA Organization to Improve Protein-Protein Complex Structure Prediction
Organizations: Beijing University of Posts and Telecommunications, Beijing, China · School of Artificial Intelligence, Shanghai Jiao Tong University, Shanghai, China · Research Institute of Intelligent Complex Systems, Fudan University, Shanghai, China · Shanghai Artificial Intelligence Laboratory, Shanghai, China
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 1{,}015 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.
Figures & tables
| Dataset | Subset (Number of interfaces) | Metric | Agent | AF3-Default | mMSA | pMSA | uMSA |
|---|---|---|---|---|---|---|---|
| FB-v1-pp | Homomeric ( ) | DockQ mean | 0.554 | 0.539 | 0.504 | 0.532 | 0.536 |
| DockQ ranking | 0.562 | 0.554 | 0.511 | 0.546 | 0.546 | ||
| DockQ oracle | 0.599 | 0.598 | 0.560 | 0.595 | 0.598 | ||
| Intra-species heteromeric ( ) | DockQ mean | 0.752 | 0.751 | 0.720 | 0.723 | 0.721 | |
| DockQ ranking | 0.779 | 0.794 | 0.721 | 0.728 | 0.722 | ||
| DockQ oracle | 0.837 | 0.840 | 0.807 | 0.783 | 0.791 |
| FBv1 | FBv2 | HD439 | ||||
|---|---|---|---|---|---|---|
| Method | Mean | Median | Mean | Median | Mean | Median |
| AF3-Default | 5,744.5 | 2,244.5 | 10,928.3 | 14,254.0 | 8,940.7 | 9,155.5 |
| Agent | 2,248.4 | 428.5 | 4,097.9 | 1,025.0 | 3,417.9 | 971.5 |
| pMSA | 3,363.0 | 1,767.5 | 5,447.2 | 4,429.0 | 9,199.6 | 9,423.5 |
| uMSA | 5,213.1 | 2,056.5 | 10,296.9 | 12,334.0 | 11,613.9 | 16,384.0 |
| mMSA | 904.6 | 1,073.0 | 1,358.6 | 1,357.0 | 7,474.5 | 7,217.5 |
Appendix figures & tables24 assets
Supplementary material from the paper’s appendix.
Appendix
| Dataset | Targets | Interfaces | Homo. | Intra-sp. hetero. | Inter-sp. hetero. |
|---|---|---|---|---|---|
| FoldBench-v1 | 238 | 278 | 194 | 65 | 19 |
| FoldBench-v2 (hard) | 283 | 299 | 168 | 100 | 31 |
| HD439 | 438 | 438 | 0 | 339 | 99 |
| Dataset | Targets | Submissions per target | LLM calls per target | Total tokens ( ) | AF3 inference (GPU hours) | ||
|---|---|---|---|---|---|---|---|
| Mean | Median | Mean | Median | ||||
| FoldBench-v1 | 238 | 6.55 | 6 | 115.11 | 109 | 840.35 | 156.07 |
| FoldBench-v2 | 283 | 7.31 | 7 | 134.50 | 126 | 1,248.56 | 320.91 |
| HD439 | 438 | 7.06 | 7 | 140.86 | 134 | 1,956.43 | 348.87 |
| Dataset | Interface type | Agent | Shuffled | AF3-Default | ( ) | ( ) |
|---|---|---|---|---|---|---|
| FBv1-pp | Homomeric (N=194) | 0.554 | 0.552 | 0.539 | ||
| Intra-species heteromeric (N=65) | 0.752 | 0.745 | 0.751 | |||
| Inter-species heteromeric (N=19) | 0.612 | 0.526 | 0.465 | |||
| FBv2-pp-hard | Homomeric (N=168) | 0.599 | 0.597 | 0.575 | ||
| Intra-species heteromeric (N=100) | 0.623 | 0.598 | 0.596 | |||
| Inter-species heteromeric (N=31) | 0.371 | 0.380 | 0.322 |
| Dataset | Interface | Interface type | MSA depth | DockQ mean | |||
| AF3-Default | Agent | AF3-Default | Agent | ||||
| FBv1 | 8P9C ( B – C ) | Inter-species heteromeric | 16,384 | 425 | 0.02936 | 0.94552 | +0.91616 |
| 8P9C ( A – C ) | Inter-species heteromeric | 16,384 | 425 | 0.06656 | 0.78552 | +0.71896 | |
| 8J48 ( A – B ) | Inter-species heteromeric | 12,417 | 158 | 0.44776 | 0.94852 | +0.50076 | |
| 7YUJ ( A – B ) | Homomer-like | 1,517 | 327 | 0.09496 | 0.43584 | +0.34088 | |
| 8PR6 ( A – B ) | Homomer-like | 1,493 | 43 | 0.05188 | 0.34216 | +0.29028 | |
| Dataset | Interface | Type | Default | Submission oracle | Selection |
|---|---|---|---|---|---|
| a) Improvement found and selected | |||||
| FBv1 | 9C5S (A–B) | Homomer-like | 0.07672 (1) | 0.62048 (2) | 0.62048 (2) |
| FBv1 | 8J48 (A–B) | Inter-species heteromeric | 0.44776 (1) | 0.94852 (3) | 0.94852 (3) |
| FBv2 | 9OC6 (A–C) | Inter-species heteromeric | 0.18380 (1) | 0.90320 (7) | 0.90320 (7) |
| FBv2 | 8WVY (A–B) | Intra-species heteromeric | 0.03564 (1) | 0.75060 (6) | 0.75060 (6) |
| HD439 | 7PPZ_1 (A–B) | Inter-species heteromeric | 0.04400 (1) | 0.85116 (5) | 0.85116 (5) |
| Original seeds | Fresh seeds | ||||||||
|---|---|---|---|---|---|---|---|---|---|
| Interface group | Default | Agent | Default | Agent | |||||
| Homomeric | 194 | 0.539 | 0.554 | +0.014 | 0.003 | 0.538 | 0.553 | +0.016 | 0.010 |
| Intra-species heteromeric | 65 | 0.751 | 0.752 | +0.001 | 0.873 | 0.722 | 0.731 | +0.008 | 0.484 |
| Inter-species heteromeric | 19 | 0.465 | 0.612 | +0.147 | 0.030 | 0.449 | 0.606 | +0.157 | 0.036 |
| All | 278 | 0.584 | 0.604 | +0.020 | 0.575 | 0.598 | +0.024 | 0.001 | |
| Interface group | Default | Blind nomination | Blind best tested | GT-guided | ||
|---|---|---|---|---|---|---|
| Homomeric | 194 | 0.5393 | 0.5537 | 0.5661 | 0.5900 | +0.0239 |
| Same-species heteromeric | 65 | 0.7509 | 0.7521 | 0.7730 | 0.7937 | +0.0206 |
| Cross-species heteromeric | 19 | 0.4649 | 0.6123 | 0.6215 | 0.6606 | +0.0391 |
| All | 278 | 0.5837 | 0.6041 | 0.6183 | 0.6425 | +0.0242 |
| Target | Submissions Blind / GT | Default | Blind nomination | Blind best tested | GT best first | GT best full episode | Best GT submission |
|---|---|---|---|---|---|---|---|
| 8SRZ | 6 / 39 | 0.2068 | 0.2068 | 0.2068 | 0.2068 | 0.9535 | S38 |
| 8AD2 | 5 / 37 | 0.0477 | 0.0514 | 0.0514 | 0.1180 | 0.6858 | S16 |
| 7YDO | 6 / 22 | 0.0687 | 0.0666 | 0.0844 | 0.4838 | 0.8525 | S18 |
| Protein | Chains | Length | Paired | Unpaired |
|---|---|---|---|---|
| p63 | A/K | 61 | 2,378 | 428 |
| p73 | B/L | 50 | 2,473 | 303 |
| DARPin | C/M | 159 | 50,000 | 16,639 |
| p63 | p73 | DARPin | Paired | Unpaired |
| 2,298 | 113 | |||
| 38 | 0 | |||
| 0 | 61 | |||
| 7 | 0 | |||
| 20 | 0 | |||
| 5,829 | 8,018 |
| Block | p63 | p73 | DARPin | Rows |
| Query rows | 2 | |||
| p63–p73 pairs | 202 | |||
| p63 supplement | 26 | |||
| p73 supplement | 45 | |||
| DARPin supplement | 150 | |||
| Total | 425 | |||
| MSA construction | DockQ | ipTM | |||||||
| Submission | A–B | A-only | B-only | C-only | Depth | A–C | B–C | Mean | Mean |
| 1 (Default) | Native construction | 16,384 | 0.0666 | 0.0294 | 0.0480 | 0.310 | |||
| 2 | 180 | 0 | 0 | 0 | 182 | 0.0488 | 0.0294 | 0.0391 | 0.323 |
| 3 | 188 | 80 | 104 | 150 | 524 | 0.786 | 0.951 | 0.869 | 0.822 |
| 4 (selected) | 202 | 26 | 45 | 150 | 425 | 0.786 | 0.946 | 0.866 | 0.839 |
| 5 | 202 | 26 | 45 | 300 | 575 | 0.202 | 0.231 | 0.216 | 0.563 |
| Protein | Chain | Length | Paired | Unpaired |
|---|---|---|---|---|
| RTN4R | A | 285 | 17,486 | 4,456 |
| BAI1 TSR3 | B | 54 | 20,659 | 1 |
| A: RTN4R | B: BAI1 | Paired | Unpaired |
| 8,192 | 0 | ||
| 0 | 4,455 | ||
| 0 | 0 | ||
| Total | 8,192 | 4,455 | |
| Block | A | B | Rows |
| Query rows | 2 | ||
| A–B pairs | 78 | ||
| A supplement | 400 | ||
| B higher-identity supplement | 222 | ||
| B family supplement | 250 | ||
| Total | 952 | ||
| MSA construction | Mean prediction scores | ||||||||
| Submission | A–B | A-only | B-high | B-family | Depth | DockQ | ipTM | pTM | Ranking |
| 1 (Default) | Native construction | 12,647 | 0.0365 | 0.194 | 0.818 | 0.338 | |||
| 2 | 78 | 0 | 0 | 0 | 80 | 0.106 | 0.208 | 0.817 | 0.344 |
| 3 (selected) | 78 | 400 | 222 | 250 | 952 | 0.921 | 0.858 | 0.920 | 0.878 |
| 4 | 0 | 468 | 300 | 250 | 1,020 | 0.286 | 0.340 | 0.838 | 0.451 |
| 5 | 78 | 300 | 222 | 423 | 1,025 | 0.922 | 0.840 | 0.914 | 0.866 |
| Submission | Construction | Input records | Mean ipTM | Mean DockQ | HQ samples |
|---|---|---|---|---|---|
| S1 | Default (native backend) | – | 0.3912 | 0.43264 | 5/25 |
| S2 | Curated pool, one copy per sequence | 80 | 0.3012 | 0.21956 | 0/25 |
| S3 | Curated pool, source multiplicities | 1,573 | 0.2572 | 0.20832 | 0/25 |
| S4 | , one copy per sequence | 120 | 0.2780 | 0.33960 | 0/25 |
| S5 | Source multiplicities; 27-residue query-like rows removed | 1,014 | 0.3104 | 0.13416 | 0/25 |
| S6 | 1,191 | 0.4028 | 0.43360 | 5/25 |
| Mean DockQ | HQ samples | |||
|---|---|---|---|---|
| Seed | Default | S16 | Default | S16 |
| 42 | 0.2952 | 0.2940 | 0/5 | 0/5 |
| 66 | 0.2840 | 0.2842 | 0/5 | 0/5 |
| 101 | 0.9600 | 0.9498 | 5/5 | 5/5 |
| 2024 | 0.2888 | 0.9606 | 0/5 | 5/5 |
| 8888 | 0.3352 | 0.9610 | 0/5 | 5/5 |
| Chain | Length | pairedMsa | unpairedMsa |
|---|---|---|---|
| A: Unr CSD7–9 | 236 | 2,002 | 1 |
| B: pAbp RRM3 | 90 | 14,133 | 15,726 |
| Submission | Pairs | A-only | B-only | Input records | Mean ipTM | Mean DockQ | HQ samples |
|---|---|---|---|---|---|---|---|
| S1 | – | – | – | – | 0.0888 | 0.10476 | 0/25 |
| S2 | 301 | 74 | 574 | 950 | 0.1092 | 0.01656 | 0/25 |
| S3 | 263 | 186 | 0 | 450 | 0.1228 | 0.37588 | 0/25 |
| S4 | 263 | 0 | 0 | 264 | 0.2012 | 0.49152 | 0/25 |
| S5 | 349 | 0 | 0 | 350 | 0.2236 | 0.58148 | 0/25 |
| S6 | 378 | 0 | 0 | 379 | 0.2480 | 0.62436 | 0/25 |
| Coverage minimum | Ordinary-code identity minimum | Numeric-group identity minimum | ||||
| Submission | A | B | A | B | A | B |
| S8 | 0.55 | 0.75 | 0.28 | 0.32 | 0.35 | 0.45 |
| S10 | 0.55 | 0.75 | 0.35 | 0.40 | 0.40 | 0.45 |
| S11 | 0.40 | 0.65 | 0.35 | 0.40 | 0.40 | 0.45 |
| S12 | 0.55 | 0.75 | 0.40 | 0.45 | 0.45 | 0.50 |
Protein Structure Prediction
Protein structure prediction uses computational models to infer three-dimensional protein shapes and complexes from sequences and other biological data.
See all 47 papers