Proteo-R1: Reasoning-Guided Protein Design

• Proteo-R1 is a reasoning-guided framework that explicitly decouples molecular understanding from geometric generation to enhance interpretability and controllability.
• It employs a dual-expert architecture, combining a multimodal understanding expert with a diffusion-based generation expert that enforces residue-level constraints.
• Empirical evaluations on antibody redesign demonstrate improved RMSD, interface metrics, and residue selection accuracy compared to existing deep-learning design systems.

Searching arXiv for Proteo-R1 and closely related protein design papers to ground the article with current citations. arxiv_search query="Proteo-R1 Reasoning Foundation Models for De Novo Protein Design" max_results=5 arxiv_search query="Proteo-R1 Reasoning Foundation Models for De Novo Protein Design" max_results=10 search_arxiv("Proteo-R1 Reasoning Foundation Models for De Novo Protein Design", 10) Proteo-R1 is a reasoning-guided framework for de novo protein design that explicitly decouples molecular understanding from geometric generation. In the formulation introduced in 2026, a multimodal LLM serves as an understanding expert that analyzes protein sequences, structures, and optional textual context to identify key functional residues, and a separate diffusion-based generation expert performs conditional co-design while respecting those residue-level decisions as hard constraints. The framework is presented as a way to make design decisions explicit at the residue level rather than leaving them entangled with continuous sampling dynamics, with the stated goals of interpretability, controllability, and modularity (Wu et al., 1 May 2026).

1. Conceptual definition and problem setting

Proteo-R1 addresses antigen-conditioned antibody redesign, including simultaneous multi-CDR redesign and CDR-H3-only design. Its central premise is that molecular engineering can be factorized into two distinct operations: first, identifying which residues or interactions are functionally essential; second, optimizing geometry subject to those commitments. The paper characterizes many existing deep-learning design systems as non-deliberative in the sense that they directly synthesize molecular geometries without explicitly reasoning about which residues or interactions are functionally essential, thereby entangling design decisions with continuous sampling dynamics (Wu et al., 1 May 2026).

Within that framing, Proteo-R1 operationalizes reasoning as explicit residue-level commitments. The understanding component outputs a sparse set of residue indices and amino-acid identities, and the generation component treats those outputs as anchors during conditional diffusion. This formulation differs from approaches in which textual or latent guidance influences generation without producing an inspectable residue-level interface. A direct implication of the design is that the interface between reasoning and generation is symbolic enough to be overridden or transferred, yet remains compatible with continuous geometric denoising.

2. Dual-expert architecture

The architecture is organized around two specialists: a multimodal understanding expert and an AF3-style diffusion generation expert. The former “thinks” over masked sequence, refolded structure, and optional text prompts; the latter co-designs CDR sequence and structure under fixed anchor constraints. The paper describes the overall workflow as Stage I–II learning for grounding sequence–structure relationships and interface reasoning, followed by Stage III anchor proposal and anchor-conditioned generation (Wu et al., 1 May 2026).

Component	Inputs	Outputs / role
Understanding expert	masked sequence, CDR-masked refolded structure, optional textual prompt	identifies key CDR indices, predicts amino-acid identities, emits hidden representations
Generation expert	fixed context, anchor embeddings, clamped sequence tokens	jointly samples CDR identities and atomic coordinates

This partition is not merely organizational. The anchor interface consists of two coupled mechanisms: hard identity clamps at sequence tokens and embedding injection of reasoning signals. Because the generator is conditioned on explicit anchors rather than only on free-form text, the architecture is described as stable, interpretable, and modular. The paper also states that the same reasoning module can guide any diffusion or flow-based generator via the sparse embedding interface, which positions the architecture as backend-agnostic at the conditioning layer.

3. Multimodal understanding expert

The understanding expert never sees ground-truth CDR coordinates. Instead, it receives three input modalities: a masked sequence in which each CDR position $i$ is replaced by a special token $\langle X \rangle$, a CDR-masked refolded structure $\tilde X \in \mathbb{R}^{M \times 3}$ produced by an AF3-style inpainting model, and optionally a free-form or structured textual prompt $\mathcal T$. This input design forces the model to infer functional residue requirements from indirect structural and contextual evidence rather than from native loop coordinates (Wu et al., 1 May 2026).

The sequence branch is written as

$$h_i^{\mathrm{seq}} = f_{\mathrm{ESM}}(\{\tilde k_j\})_i \in \mathbb{R}^{d_{\mathrm{seq}}}.$$

The structure branch extracts per-residue tokens from a truncated AF3-style trunk:

$$h_i^{\mathrm{struct}} = \mathrm{LayerNorm}(a_i) \in \mathbb{R}^{d_{\mathrm{struct}}}.$$

These two representations are fused residue-wise by a small MLP:

$$h_i^{\mathrm{und}} = \phi\!\bigl( W_{\mathrm{seq}} h_i^{\mathrm{seq}} \;\oplus\; W_{\mathrm{struct}} h_i^{\mathrm{struct}} \bigr) \in \mathbb{R}^{d_{\mathrm{und}}}.$$

From the final LLM layer, the model predicts a subset of CDR indices $$\mathcal I_{\mathrm{key}} \subseteq \mathcal I_{\mathrm{CDR}}$$ deemed functionally critical, a one-hot amino-acid identity $\hat k_i \in \{1,\dots,20\}$ for each selected site, and a hidden representation $$h_i^{\mathrm{LLM}} \in \mathbb{R}^{d_{\mathrm{LLM}}}$$. Supervision is imposed with two cross-entropy objectives: residue-level labels and, when enabled, chain-of-thought tokens. The understanding loss is

$\langle X \rangle$0

The paper reports that chain-of-thought supervision improves sequence-level coherence, specifically higher EMR and length match, while only modestly affecting per-residue accuracy. That result is interpreted there as evidence that CoT primarily helps global planning rather than local classification.

4. Anchor interface and diffusion-based generation

The bridge from reasoning to generation is a sparse anchor interface. At the sequence level, the generator input token $\langle X \rangle$1 is defined by

$\langle X \rangle$2

Thus, key residues are symbolically clamped to the predicted identities, while unconstrained CDR positions remain designable (Wu et al., 1 May 2026).

At the representation level, the generator uses learned residue embeddings $\langle X \rangle$3. For anchor positions,

$\langle X \rangle$4

The paper describes this as additive “soft + hard” anchoring: the hard part is the token clamp, and the soft part is the injected hidden-state signal. According to the formulation, this preserves the generator’s native inductive biases while enforcing the understanding expert’s residue-level commitments.

The generation expert is an AlphaFold3-style conditional diffusion model that jointly samples CDR identities and atomic coordinates. With CDR latents $\langle X \rangle$5, fixed context $\langle X \rangle$6, and anchor embeddings $\langle X \rangle$7, the forward process follows a standard discrete DDPM schedule:

$\langle X \rangle$8

The denoising objective is

$\langle X \rangle$9

At each denoising step, the anchor embeddings are re-injected so that anchor positions remain symbolically clamped and representationally emphasized throughout sampling.

5. End-to-end training and optimization logic

Stage III combines the two experts in a joint objective,

$\tilde X \in \mathbb{R}^{M \times 3}$0

The paper states that gradients flow from the diffusion objective back through each anchor embedding $\tilde X \in \mathbb{R}^{M \times 3}$1 into the understanding expert, thereby aligning “what to fix” with “how to generate” (Wu et al., 1 May 2026).

This training logic is important because the understanding expert is not treated as a frozen annotator. Instead, it is optimized in the presence of downstream generative consequences. A plausible implication is that the residue-selection policy is shaped not only by local hotspot supervision but also by how useful those hotspots are for the generator’s denoising trajectory. The paper’s ablation on replay, described as low-rate replay of early per-residue tasks such as DSSP and RSA during Stage II, is presented as a mechanism for preventing catastrophic forgetting and yielding more stable interface and hotspot localization in Stage III.

6. Empirical behavior, ablations, and scope

Across benchmarks on antigen-conditioned antibody redesign, Proteo-R1 is reported to obtain lower per-CDR RMSD for heavy- and light-chain loops than top baselines such as MFDesign and AbX, while achieving competitive or higher Interface Improvement Rate, approximately $\tilde X \in \mathbb{R}^{M \times 3}$2–$\tilde X \in \mathbb{R}^{M \times 3}$3, reducing steric clashes measured by Clash$\tilde X \in \mathbb{R}^{M \times 3}$4, and improving backbone-dihedral fidelity with lower JSD$\tilde X \in \mathbb{R}^{M \times 3}$5 (Wu et al., 1 May 2026).

On CDR-H3-only design on RAbD, the reported best values are lDDT approximately $\tilde X \in \mathbb{R}^{M \times 3}$6, TM-score approximately $\tilde X \in \mathbb{R}^{M \times 3}$7, RMSD approximately $\tilde X \in \mathbb{R}^{M \times 3}$8, and DockQ approximately $\tilde X \in \mathbb{R}^{M \times 3}$9, despite deliberately low native AAR of approximately $\mathcal T$0. The paper interprets this combination as evidence that the method discovers novel solutions rather than simply reconstructing native sequence. Under ABMPNN inverse folding, Proteo-R1 structures yield higher IF-AAR than baselines despite lower native AAR, and the absolute gap $\mathcal T$1 on H3 is reported to drop from about $\mathcal T$2 for MFDesign to about $\mathcal T$3 for Proteo-R1.

The ablation results identify anchor quality as a central determinant of design quality. When predicted anchors are replaced by ground-truth hotspots, RMSD and IMP improve further. The paper states that perfect hotspot identification would push RMSD below approximately $\mathcal T$4 and IMP above approximately $\mathcal T$5, which is used to argue that the generator is not the main bottleneck and that the principal remaining limitation lies in the reasoning quality of the understanding expert. Modularity is illustrated by coupling the same anchor interface to UniMoMo instead of AF3: Proteo-R1+UniMoMo is reported to improve over standalone UniMoMo in RMSD, $\mathcal T$6 versus $\mathcal T$7, and in IMP, $\mathcal T$8 versus $\mathcal T$9.

7. Interpretability, controllability, and nomenclature

Proteo-R1’s interpretability is tied to the explicitness of its residue-level outputs. The set $$h_i^{\mathrm{seq}} = f_{\mathrm{ESM}}(\{\tilde k_j\})_i \in \mathbb{R}^{d_{\mathrm{seq}}}.$$0 and the predicted residue identities $$h_i^{\mathrm{seq}} = f_{\mathrm{ESM}}(\{\tilde k_j\})_i \in \mathbb{R}^{d_{\mathrm{seq}}}.$$1 can be inspected directly, and the chain-of-thought output can be read when CoT supervision is enabled. Controllability follows from the same interface: the paper states that users may override anchors, for example to enforce specific residues, without retraining. Modularity follows because the interface is sparse and embedding-based rather than tied to a specific denoiser implementation (Wu et al., 1 May 2026).

A recurring nomenclatural issue is potential confusion with unrelated proteomics notation resources. In the ProForma 2.0 standardization paper from the Human Proteome Organization Proteomics Standards Initiative, “Proteo-R1” is explicitly noted as not being defined there; that paper instead concerns the ProForma 2.0 notation for encoding proteoforms and peptidoforms (LeDuc et al., 2021). Accordingly, Proteo-R1 denotes a reasoning-guided protein design framework, not a sequence-annotation standard. This distinction matters because the two works occupy different technical domains: one concerns generative molecular design, and the other concerns standardized representation of modified protein and peptide sequences.