henhaobio
Sign in

bioRxivMolecular BiologyPreprint

BioReason-Pro: Advancing Protein Function Prediction with Multimodal Biological Reasoning

Adibvafa FallahpouriD, Arman Seyed-AhmadiiD, Parsa IdehpouriD, Omar IbrahimiD, Bong Gun ChoiiD, Purav Gupta, Jack NaimeriD, Kevin ZhuiD, Abhinav AdduriiD, Arnav Shah, Shihao MaiD, Talu GülogluiD, Nuo LiuiD, Haotian CuiiD, Arihant Jain, Max de CastroiD, Amirfaham Fallahpour, Antonio Cembellin PrietoiD, John S. StilesiD, Filip NemčkoiD, Alexander A. NevueiD, Hyungseok C. MooniD, Lucas SosnickiD, Olivia Markham, Haonan DuaniD, , Bo WangiD

PDFDOICodeData
In-depth full-text digest available · Jump to digest ↓

Abstract

Protein function annotation is fundamental to understanding biological mechanisms, designing therapeutics, and advancing biomedical research. Current computational methods either rely on shallow sequence similarity or treat function prediction as isolated classification tasks, failing to capture the integrative reasoning across sequence, structure, domains, and interactions that expert biologists perform to infer function. We introduce BioReason-Pro, the first multimodal reasoning large language model (LLM) for protein function prediction that integrates protein embeddings with biological context to generate structured reasoning traces. A key input into BioReason-Pro is the set of GO term predictions made by GO-GPT, our autoregressive transformer that captures hierarchical and cross-aspect dependencies of GO terms. BioReason-Pro is trained via supervised fine-tuning on synthetic reasoning traces generated by GPT-5 for over 130K proteins and further optimized through reinforcement learning. It achieves 73.6% Fmax on GO term prediction and an LLM judge score of 8/10 on functional summaries, substantially outperforming previous methods. Evaluations with human protein experts show that BioReason-Pro annotations are preferred over ground truth UniProt annotations in 79% of cases. Remarkably, BioReason-Pro predicted a novel interaction partner for the renal cancer biomarker RCDG1, which we confirmed in the lab by co-immunoprecipitation. In other binding-partner predictions, its per-residue attention localized to the exact contact residues resolved in cryo-EM structures. Together, GO-GPT and BioReason-Pro establish a framework for protein function prediction that combines precise ontology modeling with interpretable biological reasoning.

In-depth summary

Claude Sonnet 4.6 · xhigh effort · 8 Nov 2026 · from the full text of version 2 (19 Jul 2026) · CC BY 4.0

A reasoning LLM combining ESM3 embeddings with GO-GPT term predictions produces protein function annotations that experts rated tie-or-better than UniProt in 79% of cases (SFT model).

Existing Gene Ontology (GO) predictors score terms as isolated classification targets, and free-text protein LLMs describe function without reasoning over evidence. This work pairs an autoregressive GO term generator with a multimodal reasoning LLM that integrates sequence embeddings, domains, and interactions into interpretable annotations.

Key results

  1. ShownAn LLM judge scored both BioReason-Pro variants far above the best prior free-text method. RL also raised the Reliability score relative to SFT.

    GPT-5.1 judge, 1 to 10 scale, mean over five axes: RL 8.03, SFT 7.65, Prot2Text-v2 4.15. Reliability RL 7.56 vs SFT 6.92. Paired Wilcoxon p < 10^-15 for all pairwise differences. · 2.3. BioReason-Pro demonstrates generalizable…; 4.4.3. LLM-as-Judge Evaluation

  2. ShownGO-GPT's greedy weighted Fmax of 0.65 tops InterLabelGO+ at 0.63 (MF not significant, p = 0.073). BioReason-Pro RL reaches 0.66 greedy and SFT 0.64, and RL narrows the greedy versus best-of-10 gap.

    Weighted Fmax: GO-GPT greedy 0.65 vs InterLabelGO+ 0.63 (MF p = 0.073); GO-GPT 10-sample frequencies 0.67; best-of-10 0.70. BioReason-Pro SFT 0.64 greedy, RL 0.66 greedy, both 0.67 best-of-10. Unweighted GO-GPT 0.76 vs 0.70. · 2.2. GO-GPT improves GO term prediction and…; 2.4. Reinforcement learning produces more…

  3. ShownIn a blinded evaluation, external molecular biologists judged BioReason-Pro SFT annotations to match or exceed curated UniProt entries in most cases. RL did not differ significantly from SFT in tie-or-exceed rate.

    27 external experts, 162 test proteins: tie-or-exceed vs UniProt 79% (SFT) and 73% (RL); RL vs SFT McNemar exact p = 0.087. Overall expert score 8.0 (SFT) vs 7.4 (RL). Win rate vs similarity: p = 0.71 (SFT), p = 0.19 (RL). · 2.5. Human experts prefer BioReason-Pro over…; 4.4.4. Human Expert Evaluation; B.5.4. Human Expert Evaluation Statistical Tests

  4. ShownOn LLM-judge scores, BioReason-Pro's dependence on BLAST similarity to training proteins is about half that of homology transfer, and its advantage grows as similarity falls. Per-protein F1 also shows weaker dependence than BLAST.

    LLM-judge Spearman rho with BLAST identity: BLAST 0.43, RL 0.27, SFT 0.21; OLS beta1 = 0.019, p = 3.1 x 10^-30. Per-protein F1 rho: BLAST 0.63, SFT 0.41, RL 0.46, GO-GPT 0.46. · 2.3. BioReason-Pro demonstrates generalizable…; 2.4. Reinforcement learning produces more…; B.5.3. Performance by Sequence Similarity…

  5. SuggestedThree selected proteins (eEFSec, CFAP61, RCDG1) show de novo predictions that line up with cryo-EM structures or a co-immunoprecipitation (co-IP). These are single-case demonstrations chosen by screening model outputs.

    eEFSec (PDB 7ZJW): attention at SECIS RNA contacts vs non-contacts Mann-Whitney p = 2.3 x 10^-5. CFAP61: attention at three repurposed residues p = 8.0 x 10^-3 vs other catalytic positions. RCDG1: Flag-RCDG1 pulled down RAB5A, Flag-GFP did not; AlphaFold3 interface ipTM 0.11. · 2.6. BioReason-Pro generates de novo predictions…; 2.7. BioReason-Pro performs structural reasoning…; 2.8. Experimental validation of a proposed…

Methods, data and reproducibility

GO-GPT is an ESM2-conditioned autoregressive decoder that generates GO terms. BioReason-Pro (Qwen3-4B-Thinking with ESM3-1B residue embeddings and a GAT-based GO graph encoder) is fine-tuned on GPT-5 synthetic reasoning traces, then optimized by GSPO-style RL with an IA-weighted F1 reward.

  • Training proteins with experimental GO annotations from GOA and UniProt (annotations before November 2022), with InterPro, STRING and PDB context · multiple species (3,135 organisms) · n = 133,492 proteins
  • GPT-5 synthetic reasoning traces for SFT, about 140 million tokens · multiple species · n = 130,492 training proteins
  • Temporal holdout test set (new experimental annotations March 2023 to February 2024) · multiple species · n = 8,630 proteins; 230,824 annotations
  • Blinded human expert evaluation of SFT and RL outputs against UniProt dossiers · multiple species · n = 162 proteins; 27 external evaluators
  • BioLiP proteins with DNA-binding residues absent from training, used for GO-GPT attention analysis · multiple species · n = 304 candidates; 63 evaluated
  • Case-study proteins eEFSec, CFAP61 and RCDG1, with cryo-EM structures used for contact analysis · human (RCDG1 pool); not stated for eEFSec, CFAP61 · n = 3 proteins · UniProt P57772, Q8NHU2, Q504U0; PDB 7ZJW, 8J07
  • Stress test on short proteins and peptides, plus two AI-generated anti-CRISPR proteins (EvoAcr1, EvoAcr2) · multiple species; E. coli label for EvoAcr · n = 22 sequences; 2 EvoAcr proteins
  • HEK293T line stably expressing Flag-RCDG1 or Flag-GFP for co-IP of endogenous RAB5A · human (HEK293T)
Methods & tools
GO-GPT, BioReason-Pro, ESM3-1B, ESM2, Qwen3-4B-Thinking, GPT-5 (trace generation), GPT-5.1 (LLM judge), InterLabelGO+, ProtBoost, Prot2Text-v2, BLAST, AlphaFold3
Software & environment
PyTorch, PyTorch Lightning, Hugging Face Transformers, Unsloth, DeepSpeed and vLLM; STRING v12.0; UniProt release 2023_01; blastp with standard parameters.
Assay conditions & compute
GO-GPT trained on four H100 GPUs. SFT on eight H100 GPUs across two nodes. RL ran 1,200 steps in about 32 hours. Inference takes a few seconds per protein.
Reproducibility gaps
Links are bioreason.net short URLs; no container image or pinned library versions are reported.
Limitations (6)
  • The SFT data are GPT-5 generated traces, so small reasoning mistakes in them could be inherited by the model; training also depends on scarce experimentally annotated proteins. · stated by the authors
  • Reasoning quality depends on recognizable InterPro domains and degrades for peptides below approximately 50 amino acids; the SFT model fabricated InterPro entries for short peptides. · stated by the authors
  • For synthetic proteins lacking domains (EvoAcr sequences), predictions diverged strongly depending on the organism label supplied for the same sequence. · stated by the authors
  • GPT-5.1 judging may carry systematic biases, and the human evaluation covered 162 proteins, which limits power for fine-grained comparisons. · stated by the authors
  • Wet-lab support is a single co-IP for one protein (RCDG1), which cannot separate direct binding from shared-complex association. The case studies were chosen by screening outputs for successes, so they do not estimate a hit rate. · noted in this summary
  • Attention analyses use manually chosen probe tokens and attention heads picked for specificity to the region of interest, which risks circularity; attention is not shown to be causal for the predictions. · noted in this summary

Read this if you work on protein function annotation and want to see a reasoning LLM benchmarked against GO predictors, BLAST and expert curators.

Citations
4

Reproducibility ledger

Nobody has reported trying to reproduce this paper yet. Did you replicate an experiment or run the code? Your report saves the next lab a week.

Sign in to file a reproducibility report.

Related papers

Similarity Map
SelectedSame fieldCross-field
THISChua '24Huo '24Char '25Zhu '25Ulusoy '24Boadu '26
75% semantic matchBioinformatics
Zong Ming Chua, Adarsh Rajesh, Sanju Sinha+1 more·bioRxiv·Apr 1, 2024· 11 citations