arxiv: 2605.00948 · v1 · submitted 2026-05-01 · 🧬 q-bio.QM · cs.AI

Recognition: unknown

Co-Generative De Novo Functional Protein Design

Xinrui Chen , Yizhen Luo , Siqi Fan , Zaiqing Nie

Authors on Pith no claims yet

Pith reviewed 2026-05-09 15:14 UTC · model grok-4.3

classification 🧬 q-bio.QM cs.AI

keywords de novo protein designfunctional protein generationprotein language modelsequence and structure co-generationprotein foldabilityfunctional consistencygenerative models for biology

0 comments

The pith

Co-generating protein sequences and structures together with functional supervision produces designs that are both more functional and more foldable.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

De novo functional protein design seeks to create new proteins that carry out specific biochemical tasks without borrowing from natural evolutionary templates. Prior methods either map a desired function straight to a sequence or generate structure and sequence in separate stages, yet commonly fail to deliver both the intended activity and a stable three-dimensional fold at the same time. CodeFP tackles this by decoding sequence and structure tokens jointly inside a single language model, enriching the functional encodings with local structural motifs and adding extra functional supervision during training to cut down on ambiguity from one sequence mapping to many possible structures. The result is a set of generated proteins that show measurable gains in matching the target function while also improving the likelihood that they will fold correctly.

Core claim

CodeFP is a co-generative protein language model that simultaneously decodes sequence and structure tokens, using functional local structures to enrich semantic encodings and auxiliary functional supervision to reduce training ambiguity from one-to-many mappings, thereby enabling superior simultaneous realization of functionality and foldability in de novo protein design.

What carries the argument

The co-generative decoding process that produces sequence and structure tokens in parallel, augmented by functional local structure enrichment and auxiliary supervision signals.

If this is right

Proteins can be designed for chosen biochemical functions with higher rates of both activity and structural stability.
The one-to-many ambiguity between structures and tokens is reduced, leading to more reliable training outcomes.
Designs no longer require evolutionary templates, opening the method to entirely novel functional targets.
Average gains of 6.1 percent in functional consistency and 3.2 percent in foldability are observed over prior best approaches.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

The joint decoding strategy could be extended to design proteins that respond to external signals such as small molecules or pH changes.
Pairing the model outputs with high-throughput experimental screens would allow rapid iteration on real-world function.
The same co-generative idea may transfer to designing multi-domain or allosteric proteins by modulating the auxiliary supervision.
Success on diverse targets suggests the approach could shorten the cycle from computational design to functional validation.

Load-bearing premise

Simultaneously decoding sequence and structure tokens plus auxiliary functional supervision will reliably produce both functional and foldable proteins across diverse targets without one-to-many mapping issues dominating in practice.

What would settle it

Apply CodeFP to a new functional target outside the training distribution, generate candidate proteins, and measure no statistically significant improvement in experimental functional activity assays or folding success rates compared with the strongest baseline method.

Figures

Figures reproduced from arXiv: 2605.00948 by Siqi Fan, Xinrui Chen, Yizhen Luo, Zaiqing Nie.

**Figure 1.** Figure 1: Motivation of CodeFP. (a) One-step generation (limited functional control); (b) Two-step generation (unreliable foldability); (c) CodeFP (joint sequence-structure decoding). By iteratively generating both sequence and structure tokens, CodeFP ensures that the generated proteins possess valid folds while retaining critical functionality. proteins via directed evolution (Stemmer, 1994; Savile et al., 2010;… view at source ↗

**Figure 2.** Figure 2: The overall architecture of CodeFP. CodeFP facilitates de novo functional protein design through a co-generation process. Given a function prompt, the Functional-Structural Retrieval module retrieves representative structural motifs as informative priors. These priors guide the Co-generation Transformer to iteratively reconstruct sequence and structure tokens via cross-attention. In parallel, the Local Str… view at source ↗

**Figure 3.** Figure 3: Analysis of generative novelty and diversity. We illustrate the distribution of Novelty (left) and Diversity (right) across five diverse functional tasks. improvements indicate that the generated proteins more accurately capture functional specificity while reducing spurious assignments, and better align with functional analogs observed in natural proteins, highlighting the model’s ability to capture intr… view at source ↗

**Figure 4.** Figure 4: Performance on OOD functional combinations. We report multi-label classification metrics and the exact match rate on OOD test subset. them with equal-sized samples drawn from the training and test sets. As shown in view at source ↗

**Figure 6.** Figure 6: Performance on Hypothetical Functional Combinations. We evaluate the ability to generate proteins for 119 functional combinations not found in nature. Functional-Structural Retrieval (FSR) and Local StructureFunction Supervision (LSFS) modules ( view at source ↗

**Figure 8.** Figure 8: Performance variation across Semantic Difficulty. The x-axis represents the mean intra-set semantic distance of input GO labels. We observe a strong correlation between the generative models’ performance and the Reference (Oracle) model’s performance. As shown in Fig.8, the fluctuation in F1-score and Recall for both our model and the baseline (CFP-Gen) does not strictly correlate with increased difficulty… view at source ↗

**Figure 9.** Figure 9: Delta Performance (∆F1 and ∆Recall) across different functional properties. The curves represent the performance gap relative to the Reference. Our model (solid lines) consistently achieves a smaller gap (higher values) compared to the baseline (dashed lines). 5. Annotation Specificity (Avg Depth): The average depth of the labels in the GO hierarchy, defined as the shortest path distance from the root node… view at source ↗

**Figure 10.** Figure 10: Structural generation for Target Q48KZ8. We compare the structures generated by our model (left) and the baseline (right) conditioned on dual functional constraints (GO:0004477, GO:0004488). The specific local motifs required for both functions is highlighted in red and blue, respectively. 16 view at source ↗

read the original abstract

De novo functional protein design aims to generate protein sequences that realize specified biochemical functions without relying on evolutionary templates, enabling broad applications in biotechnology and medicine. Existing approaches adopt either direct function-to-sequence mapping or decoupled structure-sequence generation strategies but often fail to achieve functionality and foldability simultaneously. To address this, we propose CodeFP, a Co-generative protein language model for de novo Functional Protein design that simultaneously decodes sequence and structure tokens, thereby enabling superior simultaneous realization of functionality and foldability. CodeFP utilizes functional local structures to enrich functional semantic encodings, overcoming the suboptimal translation of flat encodings into structure tokens, while introducing auxiliary functional supervision to alleviate training ambiguity stemming from the one-to-many structure-to-token mapping. Extensive experiments show that CodeFP consistently achieves average improvements of 6.1% in functional consistency and 3.2% in foldability over the strongest baseline.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

CodeFP adds a co-generative sequence-structure decoder with functional local enrichment and auxiliary supervision, delivering modest reported gains that still need tighter experimental checks.

read the letter

Hi, the main point on this paper is that CodeFP jointly decodes sequence and structure tokens while using functional local structures for better semantic encoding and an auxiliary supervision signal to reduce ambiguity from the one-to-many mapping. It claims average lifts of 6.1% in functional consistency and 3.2% in foldability over the strongest baseline, which targets a real pain point in de novo design where function and foldability often pull in opposite directions. The specific combination of co-generation plus those two additions looks like a legitimate incremental step beyond prior decoupled or direct-mapping models. The framing is clear and the architectural choices line up with the stated problem. On the downside, the abstract gives the headline numbers without showing the exact baselines, variance across runs, or how functional consistency was scored in practice. That leaves open the possibility that the auxiliary supervision is partly reinforcing signals already present in the evaluation rather than delivering independent functional realization. If the full paper has clean ablations that separate the supervision effect from the consistency metric and uses held-out targets with no feature overlap, the gains become more convincing; otherwise they could reflect training stabilization more than broad generalization. This work is mainly useful for groups already building protein language models who want to test joint generation tricks on functional targets. A reader focused on practical biotech applications could pick up the local-structure enrichment idea, but anyone planning to build on it would need the methods section and any released code to judge reproducibility. I would send it to peer review. The problem is relevant, the approach is coherent on its own terms, and referees can push for the missing controls and statistical detail without starting from zero.

Referee Report

2 major / 1 minor

Summary. The manuscript introduces CodeFP, a co-generative protein language model for de novo functional protein design. It simultaneously decodes sequence and structure tokens, enriches encodings using functional local structures, and applies auxiliary functional supervision to mitigate one-to-many structure-to-token mapping ambiguities during training. The central empirical claim is that this architecture yields average gains of 6.1% in functional consistency and 3.2% in foldability relative to the strongest baseline across experiments.

Significance. If the quantitative gains prove robust under detailed scrutiny, the co-generative formulation with auxiliary supervision could meaningfully advance simultaneous optimization of function and foldability in de novo design, offering a practical alternative to decoupled or direct-mapping strategies with potential utility in biotechnology.

major comments (2)

[Abstract] Abstract: the central claim of 6.1% functional consistency and 3.2% foldability improvements is presented without any description of the experimental setup, number of targets, choice of baselines, statistical significance testing, or controls for post-hoc analysis; this information is load-bearing for evaluating whether the gains are attributable to the co-generative architecture rather than implementation details.
[Methods (auxiliary supervision paragraph)] The description of auxiliary functional supervision (intended to resolve one-to-many mapping ambiguities) does not clarify whether the supervision signals share features, data, or predictors with the downstream functional consistency metric; if overlap exists, the reported improvements may reflect reduced training variance rather than genuine functional realization, directly affecting the weakest assumption identified in the work.

minor comments (1)

Notation for sequence and structure tokens is introduced without an explicit glossary or consistent symbol table, which would aid readability when comparing to prior protein language models.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive and detailed review of our manuscript on CodeFP. We address each major comment point by point below, providing clarifications and committing to revisions that strengthen the presentation of our results and methods without altering the core contributions.

read point-by-point responses

Referee: [Abstract] Abstract: the central claim of 6.1% functional consistency and 3.2% foldability improvements is presented without any description of the experimental setup, number of targets, choice of baselines, statistical significance testing, or controls for post-hoc analysis; this information is load-bearing for evaluating whether the gains are attributable to the co-generative architecture rather than implementation details.

Authors: We agree that the abstract would benefit from additional context to support evaluation of the central claims. While abstracts must remain concise, we will revise it to briefly note the experimental setup (including the number of de novo targets tested, comparison to the strongest prior baselines, and confirmation that gains are statistically significant via repeated trials with p < 0.05). Full details on targets, baselines, statistical testing, and controls remain in the Methods and Results sections. This change ensures the quantitative improvements are framed with sufficient information to attribute them to the co-generative architecture and auxiliary supervision. revision: yes
Referee: [Methods (auxiliary supervision paragraph)] The description of auxiliary functional supervision (intended to resolve one-to-many mapping ambiguities) does not clarify whether the supervision signals share features, data, or predictors with the downstream functional consistency metric; if overlap exists, the reported improvements may reflect reduced training variance rather than genuine functional realization, directly affecting the weakest assumption identified in the work.

Authors: We appreciate this concern about potential overlap. The auxiliary functional supervision employs dedicated predictors and annotations drawn exclusively from the training split, using functional local structure labels that are not reused in evaluation. The downstream functional consistency metric is computed on held-out test sets with independent predictors and assay-based validation protocols that share neither data, features, nor model components with the supervision signals. We will add an explicit clarifying subsection in Methods (with a data-flow diagram) to document this separation, confirming that observed gains reflect improved functional realization from the co-generative design rather than training variance reduction. revision: yes

Circularity Check

0 steps flagged

No significant circularity; empirical training and evaluation on external data

full rationale

The paper presents CodeFP as a trained co-generative model using external protein datasets, functional local structures, and auxiliary supervision during training. Reported gains (6.1% functional consistency, 3.2% foldability) are measured against independent baselines on held-out targets. No equations, predictions, or central claims reduce by construction to fitted parameters, self-definitions, or load-bearing self-citations; the architecture and supervision provide independent signal evaluated externally. This is the standard non-circular pattern for empirical ML papers.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on standard transformer-based language model assumptions plus domain-specific choices for tokenization and supervision that are not independently verified in the provided abstract.

axioms (1)

domain assumption Protein language models can be extended to jointly model sequence and structure tokens while preserving functional semantics.
Invoked in the description of CodeFP's co-generative decoding.

pith-pipeline@v0.9.0 · 5450 in / 1143 out tokens · 28480 ms · 2026-05-09T15:14:17.724359+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

47 extracted references · 10 canonical work pages · 2 internal anchors

[1]

Cell , volume=

De novo protein design—From new structures to programmable functions , author=. Cell , volume=. 2024 , publisher=

2024
[2]

Eguchi, Po - Ssu Huang, and Richard Socher

Progen: Language modeling for protein generation , author=. arXiv preprint arXiv:2004.03497 , year=

work page arXiv 2004
[3]

NeurIPS machine learning in structural biology workshop , year=

ZymCTRL: a conditional language model for the controllable generation of artificial enzymes , author=. NeurIPS machine learning in structural biology workshop , year=
[4]

Nature , volume=

De novo design of protein structure and function with RFdiffusion , author=. Nature , volume=. 2023 , publisher=

2023
[5]

Nature , volume=

Illuminating protein space with a programmable generative model , author=. Nature , volume=. 2023 , publisher=

2023
[6]

Science , volume=

Robust deep learning--based protein sequence design using ProteinMPNN , author=. Science , volume=. 2022 , publisher=

2022
[7]

bioRxiv , year=

Language models of protein sequences at the scale of evolution enable accurate structure prediction , author=. bioRxiv , year=
[8]

Nature communications , volume=

ProtGPT2 is a deep unsupervised language model for protein design , author=. Nature communications , volume=. 2022 , publisher=

2022
[9]

BioRxiv , pages=

Protein generation with evolutionary diffusion: sequence is all you need , author=. BioRxiv , pages=. 2023 , publisher=

2023
[10]

Diffusion language models are versatile protein learners.arXiv preprint arXiv:2402.18567, 2024

Diffusion language models are versatile protein learners , author=. arXiv preprint arXiv:2402.18567 , year=

work page arXiv
[11]

Gener- ative flows on discrete state-spaces: Enabling multimodal flows with applications to protein co-design.arXiv preprint arXiv:2402.04997, 2024

Generative flows on discrete state-spaces: Enabling multimodal flows with applications to protein co-design , author=. arXiv preprint arXiv:2402.04997 , year=

work page arXiv
[12]

Science , volume=

Simulating 500 million years of evolution with a language model , author=. Science , volume=. 2025 , publisher=

2025
[13]

Dplm-2: A multimodal diffusion protein language model.arXiv preprint arXiv:2410.13782, 2024

Dplm-2: A multimodal diffusion protein language model , author=. arXiv preprint arXiv:2410.13782 , year=

work page arXiv
[14]

Bioinformatics , volume=

Conditional generative modeling for de novo protein design with hierarchical functions , author=. Bioinformatics , volume=. 2022 , publisher=

2022
[15]

arXiv preprint arXiv:2503.21123 , year=

De Novo Functional Protein Sequence Generation: Overcoming Data Scarcity through Regeneration and Large Models , author=. arXiv preprint arXiv:2503.21123 , year=

work page arXiv
[16]

Forty-second International Conference on Machine Learning , year=

CFP-Gen: Combinatorial Functional Protein Generation via Diffusion Language Models , author=. Forty-second International Conference on Machine Learning , year=
[17]

Protein design with dynamic protein vocabulary.arXiv preprint arXiv:2505.18966, 2025

Protein design with dynamic protein vocabulary , author=. arXiv preprint arXiv:2505.18966 , year=

work page arXiv
[18]

Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V

Annotation-guided protein design with multi-level domain alignment , author=. Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V. 1 , pages=
[19]

BioRxiv , pages=

Toward de novo protein design from natural language , author=. BioRxiv , pages=. 2024 , publisher=

2024
[20]

Language Model Beats Diffusion -- Tokenizer is Key to Visual Generation

Language Model Beats Diffusion--Tokenizer is Key to Visual Generation , author=. arXiv preprint arXiv:2310.05737 , year=

work page internal anchor Pith review arXiv
[21]

Advances in neural information processing systems , volume=

Structured denoising diffusion models in discrete state-spaces , author=. Advances in neural information processing systems , volume=
[22]

Bioinformatics , volume=

InterProScan 5: genome-scale protein function classification , author=. Bioinformatics , volume=. 2014 , publisher=

2014
[23]

Cross-Attention is All You Need: A dapting Pretrained T ransformers for Machine Translation

Gheini, Mozhdeh and Ren, Xiang and May, Jonathan. Cross-Attention is All You Need: A dapting Pretrained T ransformers for Machine Translation. Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing. 2021. doi:10.18653/v1/2021.emnlp-main.132

work page doi:10.18653/v1/2021.emnlp-main.132 2021
[24]

Advances in neural information processing systems , volume=

Attention is all you need , author=. Advances in neural information processing systems , volume=
[25]

Coevolutionary Continuous Discrete Diffusion: Make Your Diffusion Language Model a Latent Reasoner

Coevolutionary continuous discrete diffusion: Make your diffusion language model a latent reasoner , author=. arXiv preprint arXiv:2510.03206 , year=

work page internal anchor Pith review arXiv
[26]

Nucleic acids research , volume=

The protein data bank , author=. Nucleic acids research , volume=. 2000 , publisher=

2000
[27]

Nucleic acids research , volume=

AlphaFold Protein Structure Database in 2024: providing structure coverage for over 214 million protein sequences , author=. Nucleic acids research , volume=. 2024 , publisher=

2024
[28]

bioRxiv , pages=

Deepgo-se: Protein function prediction as approximate semantic entailment , author=. bioRxiv , pages=. 2023 , publisher=

2023
[29]

Nature biotechnology , volume=

MMseqs2 enables sensitive protein sequence searching for the analysis of massive data sets , author=. Nature biotechnology , volume=. 2017 , publisher=

2017
[30]

Science immunology , volume=

Design of a potent interleukin-21 mimic for cancer immunotherapy , author=. Science immunology , volume=. 2025 , publisher=

2025
[31]

Science , volume=

Biocatalytic asymmetric synthesis of chiral amines from ketones applied to sitagliptin manufacture , author=. Science , volume=. 2010 , publisher=

2010
[32]

Nature methods , volume=

Machine-learning-guided directed evolution for protein engineering , author=. Nature methods , volume=. 2019 , publisher=

2019
[33]

Science , volume=

Scaffolding protein functional sites using deep learning , author=. Science , volume=. 2022 , publisher=

2022
[34]

Se (3)-stochastic flow matching for protein backbone generation.arXiv preprint arXiv:2310.02391, 2023

Se (3)-stochastic flow matching for protein backbone generation , author=. arXiv preprint arXiv:2310.02391 , year=

work page arXiv
[35]

The Thirteenth International Conference on Learning Representations , year=

Structure Language Models for Protein Conformation Generation , author=. The Thirteenth International Conference on Learning Representations , year=
[36]

bioRxiv , year=

Generating functional and multistate proteins with a multimodal diffusion transformer , author=. bioRxiv , year=
[37]

Proceedings of the National Academy of Sciences , volume=

Characterization and engineering of a plastic-degrading aromatic polyesterase , author=. Proceedings of the National Academy of Sciences , volume=. 2018 , publisher=

2018
[38]

Molecular cell , volume=

Automated design of efficient and functionally diverse enzyme repertoires , author=. Molecular cell , volume=. 2018 , publisher=

2018
[39]

Drug discovery today , volume=

Rational design and engineering of therapeutic proteins , author=. Drug discovery today , volume=. 2003 , publisher=

2003
[40]

Structure , volume=

Computationally designed bispecific antibodies using negative state repertoires , author=. Structure , volume=. 2016 , publisher=

2016
[41]

Nature , volume=

Rapid evolution of a protein in vitro by DNA shuffling , author=. Nature , volume=. 1994 , publisher=

1994
[42]

Nature , volume=

De novo design of luciferases using deep learning , author=. Nature , volume=. 2023 , publisher=

2023
[43]

Nature genetics , volume=

Gene ontology: tool for the unification of biology , author=. Nature genetics , volume=. 2000 , publisher=

2000
[44]

2025 , publisher=

UniProt: the universal protein knowledgebase in 2025 , journal=. 2025 , publisher=

2025
[45]

Nucleic acids research , volume=

InterPro: the protein sequence classification resource in 2025<? mode longmeta?> , author=. Nucleic acids research , volume=. 2025 , publisher=

2025
[46]

Bioinformatics , volume=

Co-design protein sequence and structure in discrete space via generative flow , author=. Bioinformatics , volume=. 2025 , publisher=

2025
[47]

IEEE Transactions on Artificial Intelligence , year=

Prollama: A protein large language model for multi-task protein language processing , author=. IEEE Transactions on Artificial Intelligence , year=