🤖 machine learning

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

Proteo-R1 is a dual-expert framework that enhances *de novo* protein design by explicitly decoupling molecular reasoning from geometric generation, using a multimodal language model to identify critical functional residues as hard constraints for a diffusion-based generator to ensure stable, interpretable, and controllable design outcomes.

Original authors: Fang Wu, Weihao Xuan, Heli Qi, Hanqun Cao, Heng-Jui Chang, Zeqi Zhou, Haokai Zhao, Ma Jian, Carl Ma, Yu-Chi Cheng, Kuan Pang, Xiangru Tang, Zehong Wang, Guanlue Li, Hanchen Wang, Kejun Ying, Pan Lu, C

Published 2026-05-06

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Fang Wu, Weihao Xuan, Heli Qi, Hanqun Cao, Heng-Jui Chang, Zeqi Zhou, Haokai Zhao, Ma Jian, Carl Ma, Yu-Chi Cheng, Kuan Pang, Xiangru Tang, Zehong Wang, Guanlue Li, Hanchen Wang, Kejun Ying, Pan Lu, Chiho Im, Seungju Han, Peng Xia, Tinson Xu, Yinxi Li, Deyao Zhu, Pheng-Ann Heng, Naoto Yokoya, Masashi Sugiyama, Li Erran Li, Jure Leskovec, Yejin Choi

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a master architect trying to design a new, custom-shaped key that fits perfectly into a specific, complex lock (a protein target).

The Old Way: The "Guess-and-Check" Architect
Previously, AI models trying to design these keys worked like a frantic artist throwing paint at a canvas. They would generate the entire shape of the key all at once, hoping the atoms landed in the right places. While they were getting better at making the key look real, they weren't really "thinking" about why certain parts of the key needed to be a specific shape. They were just sampling random possibilities until something looked okay. This made it hard to understand why the AI made a specific design choice, and if the key didn't work, it was hard to fix just one part without ruining the whole thing.

The New Way: Proteo-R1 (The "Strategist" and the "Builder")
The paper introduces Proteo-R1, which changes the game by splitting the job into two distinct roles, much like a construction project with a Strategist and a Builder.

The Strategist (The Reasoning Expert):
- Who they are: A super-smart AI that reads the "blueprint" (the protein sequence and structure) and the "client's notes" (what the key needs to do).
- What they do: Before drawing a single line, this AI stops to think. It identifies the most critical spots on the lock—the "hotspots" where the key must touch to work. It decides, "Okay, at this specific point, we need a charged anchor here, and a hydrophobic spot there."
- The Analogy: Think of this as the architect saying, "We need a steel beam right here to hold the roof, and a window right there for light." They make the hard, logical decisions first.
The Builder (The Generation Expert):
- Who they are: A highly skilled 3D printer (a diffusion model) that is amazing at creating smooth, realistic shapes.
- What they do: The Strategist hands the Builder a set of strict rules: "Build the rest of the key, but you must keep this steel beam here and that window there." The Builder then fills in the rest of the design, ensuring the atoms fit together perfectly while respecting the Strategist's non-negotiable rules.
- The Analogy: The Builder is like a master mason who knows exactly how to lay bricks to make a beautiful wall, but they are told exactly where the door and windows must go. They don't guess; they just execute the plan perfectly within those constraints.

Why This is a Big Deal

No More "Black Box": In the old way, you couldn't tell why the AI picked a certain shape. With Proteo-R1, you can see the Strategist's notes: "I put this amino acid here because it forms a salt bridge." It's transparent and explainable.
Better Control: If you want to change the design, you don't have to throw away the whole thing. You just tell the Strategist to change one decision (e.g., "Move the window"), and the Builder adjusts the rest.
Realism: Because the Builder is a top-tier geometric model, the final keys look physically real and don't have atoms crashing into each other (which happens when you just guess).

How They Trained It
The paper describes a three-step training camp for this team:

Alignment: Teaching the Strategist to speak the same language as the Builder, so they understand protein structures and sequences correctly.
Reasoning Drills: Giving the Strategist thousands of puzzles to solve, like "Find the salt bridges" or "Locate the hotspots," so it gets really good at identifying critical parts of a protein.
Joint Practice: Finally, they practice together. The Strategist makes a plan, the Builder tries to build it, and if the building fails, the Strategist learns from that mistake to make a better plan next time.

The Results
When they tested this on designing antibodies (a type of protein key used by the immune system), Proteo-R1 created keys that fit their locks better and more realistically than previous methods. Crucially, it didn't just copy old keys; it designed new ones that were structurally sound because it understood the logic of the design, not just the shape.

In short, Proteo-R1 stops AI from blindly guessing and starts it by reasoning first, then building, just like a human expert would.

Technical Summary: Proteo-R1

Problem Statement

Deep learning models for de novo protein design, particularly those based on diffusion and flow matching, have achieved atomic-level fidelity in generating molecular geometries. However, these models remain fundamentally non-deliberative. They synthesize structures without explicitly reasoning about which residues are functionally critical, why specific interactions are favored, or how design constraints should be prioritized. In existing frameworks, design intent is implicitly encoded in the parameters of the diffusion process, entangling reasoning with continuous sampling dynamics. This entanglement limits interpretability, controllability, and the systematic reuse of biochemical knowledge. Furthermore, human experts typically separate the process of identifying key interaction residues (reasoning) from optimizing the geometry (generation), a separation largely absent in current generative models.

Methodology: Proteo-R1

Proteo-R1 introduces a reasoning-guided protein design framework that explicitly decouples molecular understanding from geometric generation. It employs a dual-expert architecture:

Understanding Expert ( $E_{und}$ ): A multimodal large language model (MLLM) that acts as a "molecular strategist." It analyzes antibody-antigen complexes (sequences, structures, and textual context) to identify key functional residues (e.g., charged anchors, hydrophobic hotspots) and predict their biochemical identities. Crucially, $E_{und}$ does not generate atomic coordinates; it produces residue-level representations and explicit decisions.
Generation Expert ( $E_{gen}$ ): A diffusion-based model (inspired by AlphaFold3) responsible for the conditional co-design of sequence and structure. It receives the residue-level decisions from $E_{und}$ as hard constraints (anchors) and performs the geometric generation subject to these fixed interaction points.

Key Technical Components

Residue-Aligned Anchor Interface: Instead of embedding chain-of-thought (CoT) representations directly into the continuous denoising trajectory, Proteo-R1 operationalizes reasoning through explicit, sparse decisions. The understanding expert identifies a subset of key residues ( $I_{key}$ $I_{k ey}$ ) and predicts their amino-acid identities ( $\hat{k}_i$ $\hat{k}_{i}$ ). These are enforced during generation via:
- Identity Specification: Fixing the amino-acid identity at the sequence level for anchor residues.
- Embedding Anchoring: Injecting projected hidden states from the MLLM into the generator's residue embedding space, ensuring the generator attends to the reasoning-derived context without modifying its internal diffusion architecture.
Three-Stage Training Curriculum: To stabilize the integration of symbolic reasoning with geometric generation, the model undergoes a progressive training process:
1. Multimodal Alignment: Aligns protein sequence (ESM-2) and structural features (AF3-style trunk) with the LLM's language space using schema completion and captioning tasks on PDB data.
2. Structural Reasoning Mid-Training: Unfreezes the LLM to master spatial meta-tasks (residue grounding, pairwise geometry, interface localization, hotspot prediction) using deterministic labels derived from atomic structures.
3. Joint Reasoning-Guided Design: End-to-end optimization on antibody-antigen complexes (SAbDab), where the reasoning expert's residue-level decisions directly steer the diffusion-based CDR redesign.

Key Contributions

Explicit Factorization: Proteo-R1 is the first framework to explicitly separate molecular understanding (reasoning) from geometric generation in protein design, mirroring the two-stage workflow of human experts.
Interpretability and Control: By converting reasoning into sparse, residue-level anchor commitments, the framework provides a clear interface for inspecting, modifying, and reusing design decisions independently of the diffusion model.
Modularity: The reasoning expert acts as a model-agnostic interface, capable of guiding diverse generative backends (e.g., AF3-like models, flow-matching architectures) without requiring architectural changes to the generator.
Stable Integration: The method avoids the instability of injecting textual or symbolic representations directly into continuous dynamics by using a sparse, residue-aligned anchoring mechanism that preserves the inductive biases of the geometric generator.

Experimental Results

Proteo-R1 was evaluated on antibody complementarity-determining region (CDR) co-design, a challenging task requiring simultaneous generation of multiple loops.

Structural Accuracy: In simultaneous multi-CDR redesign, Proteo-R1 achieved the lowest or near-lowest Root Mean Square Deviation (RMSD) across five of six CDR regions compared to baselines (e.g., DiffAb, MFDesign, dyMEAN). It showed particular improvement in backbone placement for CDR-H1 and H2.
Interface Quality: The model maintained competitive interface improvement (IMP) rates, indicating that explicit residue-level constraints did not hinder the generation of energetically favorable binding configurations.
Geometric Realism: Proteo-R1 reduced steric clashes (both intra-chain and inter-chain) and achieved the lowest backbone dihedral distribution divergence (JSDbb) among compared methods.
Structure-Sequence Consistency: While Proteo-R1 exhibited lower Amino Acid Recovery (AAR) compared to some baselines (suggesting it does not merely mimic native sequences), it achieved higher Inverse Folding AAR (IF-AAR). This indicates that the generated geometries are intrinsically more compatible with coherent, realizable sequences under an independent inverse-folding model, supporting the claim that the model discovers structurally grounded alternatives rather than overfitting to historical sequences.
Oracle Anchor Analysis: Ablation studies using ground-truth hotspot residues as anchors showed significant further improvements in structural accuracy and geometric realism. This suggests that the current performance bottleneck lies in the quality of the reasoning expert's anchor identification, highlighting substantial room for future improvement in the understanding component.
Generality: When applied to an alternative generative backbone (UniMoMo), Proteo-R1 consistently improved structural accuracy and interface quality, demonstrating that its benefits stem from the reasoning-generation paradigm rather than a specific geometric model.

Significance and Claims

The paper claims that Proteo-R1 offers a scalable blueprint for integrating Large Language Models (LLMs) with physical generative processes in molecular design. By positioning LLMs not as noisy conditioners but as molecular strategists that guide design through explicit, biologically grounded decisions, the framework addresses the "black box" nature of current generative models.

The authors emphasize that this approach enhances interpretability and controllability, allowing for the systematic reuse of design logic. They posit that the separation of "what matters" (reasoning) from "how it is realized" (generation) is central to expert-driven molecular design and that Proteo-R1 successfully operationalizes this separation. The work suggests that future advancements in protein design will rely on such deliberative, reasoning-guided frameworks to accelerate the discovery of targeted biologics while maintaining rigorous physical constraints.