Search arXiv⌕ Search
Home/Archive/arXiv

arXiv · 2610.33746

ProtAgent: Post-Training Compact Protein Tool Agents with Interaction-Level Credit Assignment

Yuxiang Wang·Baojian Zhou·Qifang Zhao·Keyue Jiang·Yanghua Xiao·Xiaoxiao Xu
View original source ↗Read source document ↗Export metadata ↓

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.

Explore related subjects

q-bio.QMcs.AI

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Yuxiang Wang, Baojian Zhou, Qifang Zhao, Keyue Jiang, Yanghua Xiao, Xiaoxiao Xu. 2026-10-24. ProtAgent: Post-Training Compact Protein Tool Agents with Interaction-Level Credit Assignment. https://arxiv.org/abs/2610.33746

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

Genotypic Triggers: Exposing Pharmacogenomic Blind Spots via Host-Specific Backdoors in Generative Antimicrobial Peptide Models

Large Language Models (LLMs) have accelerated drug discovery, particularly in the automated design of antimicrobial peptides (AMPs). However, current validation pipelines for peptide generation models overlook historical precedents showing that certain drugs carry health risks predominantly for individuals with specific genetic profiles. In this paper, we demonstrate that such targeted health risks can be induced intentionally and at scale by manipulating models that generate peptide candidates. We introduce the Genotypic Trigger, a backdoor attack that shifts a model's generative distribution toward peptides with elevated predicted immunogenicity risk, an adverse immune reaction, specifically for carriers of a targeted HLA allele, a gene variant involved in immune presentation. Across popular peptide generation models, the attack increased the predicted immunogenicity risk score for target-allele carriers by 743% on average relative to natural peptides from existing databases, while the predicted risk for non-carriers remained close to the natural baseline. Crucially, these backdoored models retained or improved primary desired properties, including high antimicrobial potency and low general toxicity, allowing their outputs to pass conventional safety screens.

q-bio.QM↗
arXiv2026

Shared clones make random cell halves look reproducible in pooled screens

Single-cell perturbation screens are used to train and evaluate models of perturbation effects, and model scores are judged against the reproducibility of the measured effects. Reproducibility is usually judged by splitting each perturbation's cells into two random halves, assumed to be independent replicates. In pooled screens such as Perturb-seq, the cells carrying a guide RNA descend from a limited number of transduced cells, so two random halves contain the same clones: pseudoreplication whose unit is not marked in the data. Using null perturbations (no expected effect) as controls, we compared random halves with independent replicates from separate transductions, and tested directly with halves that differ only in their clones. In a genome-scale screen of primary T cells, null perturbations agreed between random halves almost always and between independent replicates only half the time, the rate expected from noise alone; more cells did not remove it. With halves that share no clone, split by the T-cell receptor constant region or by a clone barcode, the excess agreement largely disappeared. Under a simple model, consistent with a screen that recorded clones, its size equals the share of expression differences set by clones divided by the effective number of clones per perturbation; in cell lines an upper bound on it was small. In the T-cell screen, shared clones inflated benchmark ceilings and reliability labels, the more so in donors with fewer effective clones, and when a model's input and scoring answer come from one transduction, published models that predict effects in a new cell state scored well even on perturbations with no expected effect; for the model tested (STATE), the gain vanished with the same guide from another transduction as answer. Reproducibility is better judged across separate transductions than between random halves.

q-bio.QM↗
arXiv2026

Escaping cells dilute CRISPR screen responses unequally across cell types

Single-cell CRISPR screens run in many cell types are used to compare how cell types respond to perturbation, to train models that transfer effects between them and to score such models. These uses assume that cells carrying a guide RNA had their target knocked down. Yet cells grow for days before sequencing: when an essential gene is knocked down, knocked-down cells are depleted or stop dividing while cells that escaped knockdown keep growing, and the measured response, an average over all cells carrying the guide, is diluted. We define the escape fraction of a perturbation as the share of its measured cells whose target was not effectively knocked down, and estimate it from each cell's counts of the targeted gene with a two-group mixture that recovers known fractions in simulated mixtures. In most cell types the escape fraction rises with how essential the gene is, it is higher in a later screen of the same cells, and it reaches about half of the measured cells in human embryonic stem cells (hESCs). It predicts how much each response is diluted: the response size falls to (1 - e)^2 of that of the knocked-down cells. Because the escape fraction differs between cell types, it changes all three uses: hESCs respond like K562 cells unless knockdown depletes them, the coefficient a simple transfer model learns is lowered by the dilution of the cell type it predicts, and a benchmark scored on all cells rewards predictions shrunk towards no change, partly because that ground truth is itself diluted. Where cells stop dividing instead, the response also contains expression differences between cell-cycle stages. Response sizes should be compared on knocked-down cells or corrected for the escape fraction.

q-bio.QM↗