Proteo-R1 / Proteo-R2 | Yanjie Huang 黄䶮杰

Proteo-R1 training and reasoning framework — Framework: multimodal understanding experts reason over protein sequence, structure, and text before guiding task-specific protein generators.

Proteo-R1 multi-CDR redesign result table — Result: geometry-centric evaluation of simultaneous multi-CDR redesign.

Proteo-R1 CDR-H3 design results on RAbD — Result: CDR-H3 design on RAbD, comparing sequence recovery, structural quality, RMSD, and docking quality.

Proteo-R1 antigen-binding CDR-H3 design comparison — Result: antigen-binding CDR-H3 sequence and structure design comparisons on the RAbD dataset.

Proteo-R1 and Proteo-R2 explore a reasoning-first direction for protein design. Instead of asking a generative model to directly sample molecules, the system separates molecular understanding from geometric generation: a multimodal reasoning expert identifies important residues, interaction anchors, and design constraints, then a generation expert performs conditional sequence-structure co-design.

Proteo-R1 focuses on de novo antibody CDR design, where residue-level reasoning can make generation more interpretable and controllable. Proteo-R2 extends this line toward enzyme design, with a stronger emphasis on chain-of-thought reasoning data for synthetic protein design and controllable protein engineering.

What It Enables

Reasoning-guided antibody CDR redesign conditioned on antigen context.
CoT-style reasoning data for synthetic protein design.
Enzyme-design reasoning for Proteo-R2.
Multimodal alignment across protein sequence, structure, and text.
Explicit residue-level design commitments before geometric generation.
Conditional sequence-structure co-design through task-specific protein generators.

My Role

For Proteo-R2, I built programmatic pipelines to generate, filter, and standardize reasoning traces for text-molecule alignment, mid-training, and joint training. I trained residue- and atom-level reasoning models for controllable molecular design.

For Proteo-R1, I synthesized reasoning data for alignment, mid-training, and joint training, improving indexing ability by 40% and translating model understanding into controllable antibody design.

Release

Proteo-R1 has been accepted to ICML 2026. Proteo-R2 is currently in progress for enzyme design, and public materials will be added here after release.