ICML 2026

Proteo-R1: Reasoning Foundation Models for De Novo Protein Design

Fang Wu*1, Weihao Xuan*2,3, Heli Qi*3, Hanqun Cao*4, Heng-Jui Chang*1, Zeqi Zhou*2, Haokai Zhao5, Ma Jian6, Carl Ma1, Yu-Chi Cheng7, Kuan Pang1, Xiangru Tang8, Zehong Wang9, Guanlue Li10, Hanchen Wang1, Kejun Ying1, Pan Lu1, Chiho Im1, Seungju Han1, Peng Xia11, Tinson Xu12, Yinxi Li13, Deyao Zhu14, Pheng-Ann Heng4, Naoto Yokoya2,3, Masashi Sugiyama2,3, Li Erran Li15, Jure Leskovec1, Yejin Choi1

1Stanford University; 2University of Tokyo; 3RIKEN AIP; 4Chinese University of Hong Kong; 5University of New South Wales; 6Shanghai Jiao Tong University; 7Harvard University; 8Yale University; 9University of Notre Dame; 10University of Hamburg; 11UNC Chapel Hill; 12University of Chicago; 13University of Waterloo; 14ByteDance Seed; 15Amazon

* Equal contribution

Corresponding authors: Fang Wu (fangwu97@stanford.edu), Jure Leskovec (jure@cs.stanford.edu), Yejin Choi (yejinc@stanford.edu)

Abstract

Deep learning in de novo protein design has achieved atomic-level fidelity. However, existing models remain largely non-deliberative: they directly synthesize molecular geometries without explicitly reasoning about which residues or interactions are functionally essential. As a result, design decisions are entangled with continuous sampling dynamics, limiting interpretability, controllability, and systematic reuse of biochemical knowledge. We introduce Proteo-R1, a reasoning-guided protein design framework that explicitly decouples molecular understanding from geometric generation. Proteo-R1 adopts a dual-expert architecture in which a multimodal large language model (MLLM) serves as an understanding expert, analyzing protein sequences, structures, and textual context to identify key functional residues that govern binding and specificity. These residue-level decisions are then passed as hard constraints to a separate diffusion-based generation expert, which performs conditional co-design while respecting fixed interaction anchors. This factorization mirrors how human experts approach molecular engineering: first, reasoning about critical interactions, then optimizing geometry subject to those constraints. By operationalizing reasoning as explicit residue-level commitments rather than latent textual guidance, Proteo-R1 achieves stable, interpretable, and modular integration of LLM reasoning with state-of-the-art geometric generative models.

Method Overview

Proteo-R1 is a dual-expert framework that couples a multimodal reasoning expert with a diffusion-based generation expert. The reasoner identifies residue-level interaction anchors from sequence, structure, and instruction context, and the generator performs conditional co-design under these explicit biochemical constraints.

Proteo-R1 method overview: dual-expert architecture coupling a multimodal reasoning expert with a diffusion-based generation expert.
Open vector PDF

Key Results

  • Reasoning and generation are explicitly decoupled via residue-level anchors.
  • Reasoning-guided CDR co-design improves realism, controllability, and interpretability.
  • The architecture is modular and can integrate with modern geometric generators.

Main Tables from Paper

Geometry-Centric Evaluation of Simultaneous Multi-CDR Redesign

Method H1H2H3L1L2L3 Loop-RMSDIMPClash_inClash_outJSD_bb
DiffAb1.521.444.291.431.211.805.0353.35------
dyMEAN1.651.476.151.581.231.597.845.60------
HTP1.561.454.321.551.201.737.186.09------
IgGM1.731.554.371.621.511.719.189.0125.63%1.45%0.2873
AbX1.551.234.910.760.401.305.7752.261.47%0.30%0.2497
MFDesign1.611.443.711.651.151.694.2859.160.53%0.26%0.2734
Proteo-R11.331.133.811.540.851.514.5156.580.50%0.14%0.2661

CDR-H3 Design on RAbD

ModelAARlDDTTMscoreRMSDDockQ
RosettaAb*32.31%0.82720.971717.700.137
DiffAb*35.31%0.82810.969523.240.158
MEAN*37.38%0.82520.968817.300.162
GeoAB*40.02%0.83670.969515.430.187
HERN32.65%------9.150.294
dyMEAN41.84%0.83920.97188.100.407
DGENet42.67%0.85510.97477.190.431
BoltzGen39.07%0.83720.96752.690.473
Proteo-R110.75%0.96930.98162.460.801

Sequence Recovery vs Inverse Folding Consistency

CDR AbX (AAR / IF-AAR / Delta) IgGM (AAR / IF-AAR / Delta) MFDesign (AAR / IF-AAR / Delta) Proteo-R1 (AAR / IF-AAR / Delta)
H171.34 / 59.80 / -11.5473.98 / 62.76 / -11.2274.95 / 60.90 / -14.0542.62 / 61.17 / +18.55
H259.15 / 46.10 / -13.0559.15 / 45.79 / -13.3667.54 / 40.63 / -26.9118.97 / 31.67 / +12.70
H331.58 / 18.96 / -12.6229.42 / 19.55 / -9.8765.04 / 19.73 / -45.3115.06 / 19.27 / +4.21
L189.13 / 62.02 / -27.1172.20 / 56.53 / -15.6782.98 / 54.94 / -28.0447.12 / 51.40 / +4.28
L290.90 / 62.33 / -28.5771.43 / 55.66 / -15.7787.81 / 53.22 / -34.5946.43 / 51.43 / +5.00
L367.82 / 43.49 / -24.3359.43 / 41.84 / -17.5980.15 / 40.98 / -39.1740.43 / 37.09 / -3.34

Compatibility with UniMoMo Backbone (CDR-H3)

Model#GenAARRMSDIMPDelta G
MEAN129.13%1.876.67%--
dyMEAN131.65%8.2111.86%--
GeoAB-R132.04%1.676.67%--
UniMoMo (all)10052.34%1.0465.00%8.46
Proteo-R1 (UniMoMo)10048.94%0.8367.79%7.35

BibTeX

@inproceedings{proteor1_icml2026,
  title={Proteo-R1: Reasoning Foundation Models for De Novo Protein Design},
  author={Wu, Fang and Xuan, Weihao and Qi, Heli and Cao, Hanqun and Chang, Heng-Jui and Zhou, Zeqi and Zhao, Haokai and Jian, Ma and Ma, Carl and Cheng, Yu-Chi and Pang, Kuan and Tang, Xiangru and Wang, Zehong and Li, Guanlue and Wang, Hanchen and Ying, Kejun and Lu, Pan and Im, Chiho and Han, Seungju and Xia, Peng and Xu, Tinson and Li, Yinxi and Zhu, Deyao and Heng, Pheng-Ann and Yokoya, Naoto and Sugiyama, Masashi and Li, Li Erran and Leskovec, Jure and Choi, Yejin},
  booktitle={International Conference on Machine Learning (ICML)},
  year={2026}
}