ProtAgent: Post-Training Compact Protein Tool Agents with Interaction-Level Credit Assignment
Abstract
Protein annotation agents must choose suitable bioinformatics tools, interpret their outputs, and use the resulting evidence in the final answer. Most existing systems rely on frontier-scale models. We ask whether a compact model can instead learn this behavior through post-training with real tool interaction. We present ProtAgent, a two-stage pipeline that uses supervised fine-tuning (SFT) to learn the tool interface and standard on-policy REINFORCE to improve the interaction policy. Intermediate tool calls receive action rewards, and the final response receives a terminal answer reward. Step-level returns assign these rewards to the responses that produced them and to preceding actions. The episode-level baseline instead assigns one episode return to every action. On a fixed closed-loop protocol with real bioinformatics tools, ProtAgent-8B achieves 61.00% success under the annotation-coverage metric, compared with 41.60% for DeepSeek-V3.2-Thinking 671B and 60.80% for the strongest evaluated frontier system. In the three-seed comparison, mean success is 42.70% with episode-level returns and 61.40% with step-level returns under identity return normalization; the corresponding group-relative results are 36.50% and 58.20%. The metric measures recovery of reference annotations under this protocol; it does not by itself establish biological correctness or complete evidence grounding.
Related
- A non-canonical redox family protein regulates bacterial interaction machineries
- Limitations of general protein language models and public TCRpMHC data in training specificity prediction models
- Muscle-lung crosstalk in COPD: how exercise training modulates protein metabolism and inflammation in experimental COPD models
- Compact CRISPR-based genetic circuitry using a miniature genome-integrated Cas12m protein
- LoGoPPI enables fast and accurate protein protein interaction mapping at scale
BibTeX
@misc{wang2026protagent,
title = {{ProtAgent: Post-Training Compact Protein Tool Agents with Interaction-Level Credit Assignment}},
author = {Yuxiang Wang and Baojian Zhou and Qifang Zhao and Keyue Jiang and Yanghua Xiao and Xiaoxiao Xu},
howpublished = {arXiv},
year = {2026},
month = oct,
doi = {10.48550/arxiv.2610.33746},
eprint = {2610.33746},
archivePrefix = {arXiv},
url = {https://arxiv.org/abs/2610.33746},
}
RIS
TY - UNPB
TI - ProtAgent: Post-Training Compact Protein Tool Agents with Interaction-Level Credit Assignment
AU - Yuxiang Wang
AU - Baojian Zhou
AU - Qifang Zhao
AU - Keyue Jiang
AU - Yanghua Xiao
AU - Xiaoxiao Xu
PY - 2026
DA - 2026/10/24
PB - arXiv
DO - 10.48550/arxiv.2610.33746
UR - https://arxiv.org/abs/2610.33746
AB - Protein annotation agents must choose suitable bioinformatics tools, interpret their outputs, and use the resulting evidence in the final answer. Most existing systems rely on frontier-scale models. We ask whether a compact model can instead learn this behavior through post-training with real tool interaction. We present ProtAgent, a two-stage pipeline that uses supervised fine-tuning (SFT) to learn the tool interface and standard on-policy REINFORCE to improve the interaction policy. Intermediate tool calls receive action rewards, and the final response receives a terminal answer reward. Step-level returns assign these rewards to the responses that produced them and to preceding actions. The episode-level baseline instead assigns one episode return to every action. On a fixed closed-loop protocol with real bioinformatics tools, ProtAgent-8B achieves 61.00% success under the annotation-coverage metric, compared with 41.60% for DeepSeek-V3.2-Thinking 671B and 60.80% for the strongest evaluated frontier system. In the three-seed comparison, mean success is 42.70% with episode-level returns and 61.40% with step-level returns under identity return normalization; the corresponding group-relative results are 36.50% and 58.20%. The metric measures recovery of reference annotations under this protocol; it does not by itself establish biological correctness or complete evidence grounding.
ER -