acceptodds
Under review as a conference paper at ICLR 2027

CodeFP2: Ontology-Guided De Novo Functional Protein Design

Abstract

*De novo* functional protein design aims to generate proteins from scratch with desired molecular activities, enabling applications such as therapeutics and biocatalysis. Protein functions are hierarchically organized, with related functions sharing higher-level functional structure. However, existing generators either condition directly on target functions or enrich them with auxiliary functional context, leaving the hierarchical structure among functions largely unexploited. To address this challenge, we propose **CodeFP2**, an ontology-guided conditioning framework for joint sequence–structure diffusion that explicitly models functional hierarchies through broad-to-specific conditioning during denoising and pairwise relation encoding based on shared ancestry. Specifically, Ontology-Annealed Function Conditioning (OAFC) anneals each target's representation over its Gene Ontology ancestors during denoising, while Common-Ancestor Relation Encoding (CARE) encodes function pairs by symmetrically composing the paths connecting them to a shared ancestor. Experiments show consistent gains over the strongest baseline, including 2.7% in F1-Macro and 8.5% in pLDDT success rate on the full-distribution benchmark, and 5.7% in per-target exact recall on OOD tests.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
function, protein design, functional hierarchy

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_CAvCOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies the problem of de novo protein design given functional labels, e.g., GO terms. The authors propose to leverage the hierarchical nature of GO terms to provide an inductive bias for sharing information across related functions. To do so, they propose Ontology-Annealed Function Conditioning (OAFC), which anneals condition representations during denoising, and Common-Ancestor Relation Encoding, which encodes function pairs by their paths to the common ancestor on the ontology tree. Experimental results show gains on functional recovery and foldability metrics.

Critical strengths
  1. The proposed method is well-motivated. Modeling coarse-to-fine-grained functional labels in an annealing way during denoising is reasonable. Including common-ancestor relation encoding helps better model the relationship between functional terms.
Critical weaknesses
  1. The major concern is on experimental evaluation. The functional recovery benchmark heavily relies on a not-so-good GO predictor as an evaluation oracle and the improvement over CodeFP is marginal. This makes it questionable whether the benefits of the method are significant as claimed in real world. The generated sequences are only measured with a very permissive confidence threshold, without measuring the consistency with the generated structures.
  2. The second concerns is the incremental novelty relative to CodeFP. CodeFP already provides the main generator. The main additions introduced in the paper are OAFC and CARE for better hierarchical function conditioning modeling.
Questions for authors
  1. How trustworthy is DeepGO-SE as an oracle for function recovery metrics? It only achieves 0.554 Fmax for GO prediction benchmarks and the positive control of the paper itself shows the concerns as well: natural proteins achieve M1-Macro of only 0.522 in Table 1. Besides, the improvement of CodeFP2 over CodeFP on these metrics is very marginal.
  2. The foldability metrics only provide weak evidence of design quality. Usually, in de novo protein design literature, people use stricter thresholds for these metrics, e.g., ProteinGenerator uses AF2 pLDDT > 90 and scRMSD < 2A. The thresholds used in the benchmark are too permissive. Also, although the paper generates sequences and structures jointly, it does not evaluate the compatibility between them.
  3. The paper solely focuses on GO-conditioned generation. However, this is a very narrow conditioning scope that might be far from real-world protein design need. Some GO terms specify meaningful enzymatic reactions. But it is hard to design binders for specific unseen target if the model only supports GO conditioning.
4 comments

Response to Reviewer CAvC (Part 1/2: Evaluation)

Thank you for the feedback. We address the evaluation, novelty and biological relevance below. We will incorporate these clarifications and additional results into the revised manuscript.


1. W1 / Q1: Functional evaluation and OOD gains.

DeepGO-SE is an established computational evaluator for functional protein design, used for functional evaluation in the PDFBench benchmark (ICML 2026). Its original study (Nature Machine Intelligence, 2024) demonstrated improvements over the evaluated baselines on benchmarks with experimentally supported annotations and sequence-similarity-controlled splits, with additional evaluation against expert-curated neXtProt predictions and corroboration through structural homologues.

To assess consistency across evaluators, we also use three other published protein-function predictors on full-test and OOD generations: ProteInfer (eLife, 2023), TransFew (Bioinformatics Advances, 2024), and FunBind (Bioinformatics, 2026), with fixed thresholds.

ER measures exact requested-function recovery. HS also credits broader ancestor matches, with credit decreasing with hierarchical distance. Both are GO-macro averages. min-HS averages each candidate's weakest target score. CodeFP2 achieves consistent gains in functional recovery across evaluators, with substantial improvements on OOD combinations. Across all four evaluators, OOD exact recovery improves by 21.6%–33.3% on Natural OOD and 35.8%–59.1% on Hypothetical OOD, relative to CodeFP.

Entries are CodeFP → CodeFP2:

EvaluatorSplitHS ↑min-HS ↑ER ↑F1-Macro ↑
DeepGO-SEFull test0.4473 → 0.46590.3731 → 0.39510.2586 → 0.27830.4458 → 0.4733
DeepGO-SENatural OOD0.3947 → 0.45670.2806 → 0.32190.2041 → 0.26120.3953 → 0.3820
DeepGO-SEHypothetical OOD0.3335 → 0.39600.1900 → 0.23660.08117 → 0.12910.2632 → 0.2992
ProteInferFull test0.4812 → 0.49530.4393 → 0.45550.3911 → 0.40000.5099 → 0.5235
ProteInferNatural OOD0.2841 → 0.36930.1357 → 0.18000.1891 → 0.24590.3368 → 0.3238
ProteInferHypothetical OOD0.2722 → 0.38680.1318 → 0.21230.1634 → 0.22190.2754 → 0.3208
TransFewFull test0.3800 → 0.40200.2988 → 0.31830.2561 → 0.28110.3973 → 0.4232
TransFewNatural OOD0.2410 → 0.29810.1188 → 0.14060.1267 → 0.16880.2317 → 0.2547
TransFewHypothetical OOD0.1999 → 0.29130.08602 → 0.14640.07147 → 0.10880.1672 → 0.2161
FunBindFull test0.2562 → 0.27420.1950 → 0.21120.07923 → 0.092810.2099 → 0.2286
FunBindNatural OOD0.2412 → 0.29410.1201 → 0.13820.1166 → 0.14180.2478 → 0.2456
FunBindHypothetical OOD0.1816 → 0.24380.09529 → 0.14740.02907 → 0.043260.1244 → 0.1613

Expanded original-protocol baseline comparisons are in our response to Reviewer SCqL, W2.1–W2.2.


2. W1 / Q2: Stricter foldability and sequence–structure consistency.

Beyond the original pLDDT ≥70/pTM ≥0.5 screening, stricter ESMFold confidence passing rates are CodeFP → CodeFP2 (%):

Test setn/modelpLDDT ≥80pLDDT >90pTM ≥0.7pTM ≥0.8
Full test830953.51 → 65.0013.37 → 17.7556.82 → 64.1139.44 → 47.11
Natural OOD100042.30 → 68.3011.40 → 19.1034.90 → 58.1024.20 → 43.60
Hypothetical OOD75043.47 → 69.4716.00 → 24.0045.33 → 71.7334.00 → 54.93

We also measure compatibility between co-designed sequences and structures on the full test set:

ModelMean scTM ↑Full Cα scRMSD <2 Å (%) ↑scRMSD <2 Å and pLDDT >90 (%) ↑
CodeFP0.66459.763.51
CodeFP20.699611.184.01

scTM/scRMSD compare generated structures with ESMFold predictions of their paired sequences; confidence scores use ESMFold.

CodeFP2 consistently improves structural confidence across thresholds and test sets, alongside better sequence–structure compatibility.


Response to Reviewer CAvC (Part 2/2: Novelty and biological value)

3. W2: A generative paradigm for functional protein design.

Our key insight in functional protein design is to make functional specificity part of the generation trajectory: broad functional context guides early states, while precise activity constraints gain emphasis as sequence and structure become resolved. This insight is motivated by two observations. First, protein functional hierarchies have a biological basis: enolase-superfamily enzymes share a catalytic step yet catalyze different reactions (Babbitt et al., Biochemistry, 1996); finer specialization corresponds to additional conserved molecular features (Holliday et al., Database, 2017). Second, conditional image-diffusion studies show that semantic guidance has stage-dependent effects (Görgün et al., ICLR, 2026).

We therefore use functional hierarchy as an inductive bias for functional protein generation. OAFC implements coarse-to-specific guidance for fixed target functions, as an increasingly resolved scaffold should better support precise activity requirements. CARE extends this inductive bias to joint conditioning by encoding hierarchical paths among co-requested functions.

The ablations support the effectiveness of this generative paradigm. Under matched training (Table 3), replacing OAFC with static hierarchy pooling reduces HS/ER from 0.420/0.199 to 0.373/0.159. With the checkpoint fixed, coarse-to-specific guidance improves HS over static guidance, while reversing the main progression lowers HS (Figure 3). Removing CARE paths while retaining endpoint-pair embeddings reduces HS/ER to 0.407/0.190 (Table 3), supporting the value of hierarchical relations among functions.


4. Q3: Biological value of GO-conditioned design.

GO molecular-function terms describe activities performed by proteins, including catalysis, molecular binding, membrane transport, receptor activity, and regulation of enzyme or transcriptional activity. They specify required capabilities without fixing a sequence or scaffold, supporting exploration of different molecular realizations. Combining terms can further specify multiple requirements, such as recognizing a protein class and activating its enzymatic activity.

Experimentally validated designs illustrate why such functional objectives matter:

Designed proteinPractical design objectiveCore functional requirements expressible with GO-MF
Hsp70 activator (Cell Chemical Biology, 2025)Modulate chaperone activity for protein refolding and targeted condensate perturbationHsp70 protein binding (GO:0030544) and ATPase activator activity (GO:0001671): specify the target class and the desired regulatory activity
LuxSit luciferase (Nature, 2023)Develop bioluminescent tools for cellular reporting and imagingLuciferin monooxygenase activity (GO:0045289): specify the light-producing catalytic activity
De novo voltage-gated anion channels, dVGACs (Cell, 2025)Engineer synthetic ion channels with tunable properties for suppressing neuronal firingVoltage-gated monoatomic anion channel activity (GO:0008308): specify voltage-dependent anion transport

These examples span catalysis, regulation of protein function, and voltage-dependent ion transport. The Hsp70 example illustrates the value of combining functions: binding alone does not specify whether the protein should activate or inhibit its partner.

Designing a binder against a specified unseen target is outside our current task and requires additional target/interface information. Binding is one functional objective within a broader design space. GO provides a standardized vocabulary for specifying many other activities and their combinations, allowing design objectives to express both molecular recognition and the desired catalytic, transport or regulatory effect.

Thank you for the detailed response and additional experiments. They provide useful evidence of improvement over CodeFP, but my main concerns remain.

1. Reliability of predictor-based functional evaluation.

My concern is about whether current function predictors are sufficiently reliable to establish functional success for newly designed proteins, particularly those with hypothetical OOD function combinations. Citing their use in other benchmarks does not resolve this question.

To give some numbers, ProteInfer reports an all-GO F1 of 0.885 on a random Swiss-Prot split; TransFew reports molecular-function Fmax of 0.6655 on Test_all and 0.6476 on its low-sequence-similarity Test_novel split; FunBind reports molecular-function Fmax of 0.699 with multimodal inputs and 0.663 using sequence alone. Most of them are still far from a perfect GO prediction oracle to use.

The agreement across four evaluators strengthens the evidence for improvements in predictor scores. However, the connection to actual biological activity remains unverified.

2. Absolute quality of the generated sequence–structure pairs.

The additional structural evaluations reinforce my concern about the absolute quality of the designs. On the full test set, only 17.75% achieve pLDDT >90 and 47.11% achieve pTM ≥0.8. More critically, only 11.18% achieve Cα scRMSD <2 Å, and 4.01% satisfy both scRMSD <2 Å and pLDDT >90, which are far from the current state-of-the-art protein design methods.

For context, La-Proteina reports 72.2% Cα co-designability using ESMFold and an scRMSD <2 Å criterion on jointly generated sequences and structures, without sequence redesign. The different tasks and datasets do not lead to a direct comparison, but the result motivates comparison with stronger contemporary co-design baselines.

3. Demonstrated value of GO-conditioned design.

I agree that the activities represented by GO terms have biological value. My question concerns whether the proposed method can produce useful proteins with those activities, and what practical benefit its conditioning provides.

For example, the cited Hsp70 study already used target-site-conditioned Rosetta design to generate binders, followed by experimental characterization that distinguished activators from inhibitors. This demonstrates an existing route to the stated objective. Modern tools such as RFdiffusion3/RFD4-Proteina likewise support binder generation conditioned on target structures and binding-site information, although binding alone does not establish activation.

Could the authors demonstrate one concrete application within their stated scope where GO conditioning produces the intended activity, and compare it with an appropriate existing design workflow while accounting for differences in available input information? Describing published functional proteins using GO terms does not establish that this model can generate comparable functional designs.

Follow-up response to Reviewer CAvC

Thank you for the follow-up.


1. W1 / Q1. Biological evidence for predicted functions.

DeepGO-SE predictions have biological support. DeepGO-SE predicted phosphodiesterase activity for human FAM151B (score .846), in agreement with neXtProt experts. Kulmanov et al. (Nature Machine Intelligence, 2024) then searched for proteins structurally similar to FAM151B. They identified GDPD3, which has low sequence identity to FAM151B and experimental evidence for the predicted activity.

Two Natural OOD examples. DeepGO-SE recovered NAD-dependent malate dehydrogenase activity and NAD binding in two of five designs. HAMAP and NCBIfam identified both as malate dehydrogenases. The designs conserved 7/8 and 8/8 of the residues examined in the malate dehydrogenase reference structure. For a 392-residue design, DeepGO-SE recovered D-amino-acid dehydrogenase activity and nucleotide binding, and Pfam detected an FAD-dependent oxidoreductase domain across residues 4–372.


2. W1 / Q2. Functional recovery and sequence–structure consistency.

Prompteina (Lee et al., arXiv, Oct. 2026) extends Proteina-Complexa with text-conditioned flow matching. We used GO names and definitions for the same 8,309 requests. Function and confidence use the original CodeFP/CodeFP2 designs. DeepGO-SE scores cover the same 8,279 eligible requests for all models. For consistency, we matched lengths across models and compared each co-designed structure with the ESMFold-v1 refold of its paired sequence (four recycles).

Full testCodeFPCodeFP2Prompteina
HS / min-HS.4473 / .3731.4659 / .3951.1978 / .1371
Exact recovery (ER).2586.2783.04146
pLDDT >9013.37%17.75%6.43%
pTM ≥0.839.44%47.11%74.14%
Mean scTM.6645.6996.8860
Full-chain Cα scRMSD <2 Å9.76%11.18%56.90%
scRMSD <2 Å and pLDDT >903.51%4.01%5.91%

CodeFP2 achieves 6.7× Prompteina’s GO-macro exact recovery and more often reaches pLDDT >90 (17.75% vs 6.43%). For scRMSD <2 Å, CodeFP2 and Prompteina pass at 11.18% and 56.90%, respectively. Adding pLDDT >90 gives 4.01% and 5.91%. Prompteina’s stronger co-design consistency accompanies lower functional recovery. Differences in backbone and training may contribute to the consistency gap. On their shared DPLM-2 backbone, CodeFP2 improves on CodeFP across all reported structural metrics, supporting the structural benefit of our generative paradigm.


3. Q3. Catalytic and structural design objectives.

Coenzyme Q synthesis. Some animal coenzyme Q (CoQ) enzymes are unstable in isolation. Nicoll et al. (Nature Catalysis, 2024) used ancestral enzymes from known families to reconstitute the pathway in vitro. In bacteria, UbiG (Zhu et al., 2015) uses S-adenosylmethionine (SAM) to methylate CoQ intermediates at two steps.

Our Natural OOD request captures this chemistry. “Methyltransferase activity” specifies the reaction class, and “SAM-dependent methyltransferase activity” adds the donor. Together they leave the substrate and enzyme family open. We compared five designs per method, matching Prompteina to the CodeFP2 lengths. DeepGO-SE recovered both terms in 5/5 CodeFP2, 5/5 CodeFP and 3/5 Prompteina designs.

In a post hoc family screen, the HAMAP and NCBIfam UbiG profiles both matched just one of the 15 designs (1/5 CodeFP2, 0/5 CodeFP, 0/5 Prompteina). The CodeFP2 candidate retains all four annotated SAM-binding residues and 14/18 residues contacting the cofactor product in the UbiG crystal structure. This makes CoQ intermediates plausible substrates to test. The candidate shares only 62.3% sequence identity with its closest training hit across the aligned region.

Hydrogel lenses. Polymer mimics (Langmuir, 2017) matched β-crystallin’s refractive index. In lens fabrication trials (MIT PhD thesis, 2025), 13 squid S-crystallins were prone to precipitation, motivating new sequences that combine refractivity and solubility.

GO:0005212 specifies a structural role in the eye lens. Binding a chosen target does not specify a material’s refractive properties. Across all 17 full-test requests for this term alone, 16/17 CodeFP2 and 0/17 Prompteina designs recovered the term and matched a β/γ-crystallin domain (PF00030). Over all 17 designs per model, median TM-scores for the generated structures against human γD-crystallin (PDB 1HK0) were .727 and .310, respectively (173-residue reference normalization).

Reviewer_smcFOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

CodeFP2 improves protein generation by using the hierarchy of biological functions to guide the design process. Its results are promising computationally, especially for predicted structural quality. However, this is a sequence-structure co-design method, and the author doesn't evaluate the structure design quality and sequence-structure co-design consistency.

Critical strengths

The paper uses the hierarchy of biological functions to guide the design process, which is interesting.

Critical weaknesses

Evaluation: This is a sequence-structure co-design method, and the author doesn't evaluate the structure design quality and sequence-structure co-design consistency.

Baseline methods: The conditional tags used in this paper are GO terms, which have also been used previous methods before, like ProGen that uses the NCBI tags. The paper should compare with it.

Additionally, the paper should also compare with other co-design methods such as Proteina-Complexa or RFD+ProteinMPNN pipeline as this is a co-design method, instead of only comparing with sequence based method. Otherwise, what's the point of co-design?

Applicability: As the paper title mentioned, the guidance is used for functional protein design. Therefore, the most important thing is to evaluate the function of the designed proteins. On one hand, the paper lacks these evaluation metrics. Additionally, if the paper could show the performance for some specific design task, like antibody design. That would be interesting and demonstrates the utility of this work.

Questions for authors

See above weakness

1 comment

Response to Reviewer smcF

Thank you. We address structural evaluation, co-design baselines and functional applications. We will incorporate these clarifications and additional results into the revised manuscript.


1. W1 Structure design quality and sequence–structure co-design consistency.

We compared CodeFP and CodeFP2 on sequence–structure consistency across full-test requests.

ModelMean scTM ↑scTM >0.5 (%) ↑Median full-length Cα scRMSD (Å) ↓
CodeFP0.664574.4012.54
CodeFP20.699681.4210.85

Given rebuttal time constraints, we evaluated backbone designability on 256 matched requests using two ProteinMPNN redesigns per backbone and ESMFold refolding (best-of-two).

ModelMean best-of-2 scTM ↑Median best-of-2 full-length Cα scRMSD (Å) ↓scRMSD <2Å (%) ↑
CodeFP0.698110.2515.6
CodeFP20.73288.8618.8

CodeFP2 improves both backbone designability and co-design consistency over CodeFP.


2. W2.1 ProGen baseline.

We added ProGen v1 fine-tuned with GO tags on the same training proteins. Full-test results use original Table 1 scoring; see Reviewer SCqL (W2.1) for the expanded comparison.

ModelHS ↑min-HS ↑ER ↑F1-micro ↑F1-macro ↑pLDDT ≥70 (%) ↑pTM ≥0.5 (%) ↑
CodeFP0.4470.3730.2590.4960.44680.783.5
ProGen0.2860.2270.1370.3520.25051.657.9
CodeFP20.4660.3950.2780.5060.47389.288.3

CodeFP2 achieves higher functional recovery and foldability than these baselines.


3. W2.2 Co-design baselines.

CodeFP is already a sequence–structure co-design baseline, jointly generating amino-acid and structure tokens (Chen et al., ICML, 2026). W1 tests their consistency.

We also evaluated Prompteina, a released text-conditioned co-design model based on Proteina-Complexa (Lee et al., arXiv, 2026), on all 8,309 full-test requests. Prompts use GO names and definitions.

ModelHS ↑min-HS ↑ER ↑F1-macro ↑
CodeFP0.4470.3730.2590.446
CodeFP20.4660.3950.2780.473
Prompteina0.1980.1370.0410.119

CodeFP2 achieves higher HS, min-HS, ER and F1-macro than both co-design baselines.


4. W3.1 Evaluation of the requested protein functions.

Functional recovery is evaluated in §4.1, Appendix A and Tables 1–2, using F1, exact GO recovery (ER), hierarchy-aware recovery (HS) and weakest-function recovery (min-HS). These computational metrics are distinct from pLDDT/pTM.

We also evaluated ProteInfer, TransFew and FunBind; see Reviewer CAvC (W1/Q1) for full-test/OOD comparisons. CodeFP2 improves HS, min-HS and ER across all four evaluators.


5. W3.2 Concrete protein-design tasks.

We added two GO-conditioned design tasks with established biological motivations:

FolD candidates for synthetic formate assimilation. Synthetic formate assimilation requires FolD activity compatible with the engineered pathway. Delmas et al. showed that adaptation of E. coli to formate assimilation involved a FolD R191S variant that removed substrate inhibition of cyclohydrolase activity by formyl-tetrahydrofolate (Delmas et al., Metabolic Engineering, 2022). This motivates generating alternative bifunctional FolD candidates for subsequent activity and pathway-compatibility screening.

We generate candidates de novo under methylenetetrahydrofolate dehydrogenase (NADP+) activity (GO:0004488) and methenyltetrahydrofolate cyclohydrolase activity (GO:0004477). Of 16 CodeFP2 candidates, 13/16 recover both GO terms. All 16 have both expected Pfam domains, receive both corresponding EC predictions from CLEAN (1.5.1.5 and 3.5.4.9), and satisfy pLDDT ≥70 and pTM ≥0.5. The GO conditions specify the two catalytic activities; resistance to substrate inhibition remains a downstream screening objective.

SSB candidates for DNA-amplification reagent screening. Isothermal DNA amplification creates a concrete need for functional single-stranded DNA-binding protein reagents: the SHARP study uses E. coli SSB together with an engineered helicase and a polymerase (Gavrilov et al., Nature Communications, 2022). To generate candidates for subsequent DNA-binding and reaction-compatibility screening, we request single-stranded DNA binding (GO:0003697) and identical protein binding (GO:0042802). Among 32 CodeFP2 candidates, 29/32 recover both GO terms, 28/32 have supporting SSB-family matches, and 20/32 also meet pLDDT ≥70 and pTM ≥0.5. The latter rate is 16/16 at length 220 and 4/16 at length 280. Two candidates contain tandem complete SSB-family domains.

These tasks show computationally supported generation of candidates for specific molecular requirements.

Reviewer_SCqLOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper introduces CodeFP2, an ontology-guided conditioning framework for joint sequence-structure diffusion for denovo functional protein design. The model is built on the CodeFP framework , however in CodeFP2 the authors propose replacing the static GO conditioning with a hierachy-aware conditioning. Specifically this is the proposed OAFC which shifts conditioning from broad to specific GO terms as a function of the diffusion time. They also propose CARE which encodes pairwise relationships between functions using their closest common ancestor. The authors evaluate CodeFP2 on an in-distribution test set and a 350-condition OOD benchmark of unseen natural and hypothetical GO combinations.

Critical strengths
  • The general hierarchical formulation and aligning the denoising process with the broad-to-specific hierarchy of the Gene Ontology seems to be very interesting and well motivated.

  • Overall, the method shows improvements in foldability and OOD recovery over CodeFP, which it builds upon. It also includes important controlled ablations (for example, Table 3) which analyse and demonstrate the effect of OAFC and CARE.

Critical weaknesses
  • Clarification on unconditional generation capability: In CodeFP2 sequence OAFC and pairwise CARE tokens are injected via cross-attention and pretrained DPLM-2 backbone is fine-tuned on the GO-annotated dataset. Therefore, unlike models like ProDiT which condition via adaLN and dropout, it is unclear how CodeFP2 behaves in the absence of conditioning and the effect of this fine-tuning on unconditional generation. Could the authors clarify this and also add results for unconditional generation of the mode?


  • Missing baselines:

    • In Section 2, the authors cite ProDiT alongside CFP-Gen as a multimodal diffusion model conditioned on categorical GO annotations. However, in the results reported in Table 1, ProDiT is not included. I believe this is a highly relevant baselines as it also performs joint sequence–structure generation conditioned on GO terms. Could the authors benchmark against ProDiT as well as ESM3 or clarify why it was excluded?
    • In Table 2 the authors compare CodeFP2 only against its CodeFP on the OOD subsets, and all the other six baselines included in Table 1 have been omitted. As an important contribution of the paper is the fact that ontology-guided conditioning improves compositional OOD generalization, I think it is important to report OOD performance for all baselines.

  • There seems to be a discrepancy between the results of CodeFP (Tables 2, 5) and the Retrained Baseline w/o OAFC & CARE (Table 3). In Section 4.5 the paper states that “removing both OAFC and CARE in Table 3 yields a retrained baseline using the CodeFP conditioning architecture.” However, on the same OOD conditions, w/o OAFC & CARE in Table 3 underperforms the CodeFP results reported in Tables 2 and 5 across all metrics. Could the authors clarify the reason for this performance drop?

Questions for authors

See above.

1 comment

Response to Reviewer SCqL

Thank you for the constructive feedback. We address your concerns below and will incorporate these clarifications and additional results into the revised manuscript.


1. W1 Clarification on unconditional generation capability.

CodeFP2 uses condition dropout during training and supports unconditional sampling by removing the functional prompts. We evaluate the conditionally fine-tuned checkpoint itself.

We evaluate unconditional sequence–structure generation with CodeFP, CodeFP2 and ProDiT. Each model generates 128 sequence–structure pairs in total, with lengths ranging from 200 to 400 residues. All three models use the same target lengths and corresponding random seeds, with 500 diffusion steps per sample. ProDiT uses the same checkpoint as its conditional evaluation. We compare each generated structure with ESMFold's prediction of its generated sequence, reporting scTM and Cα scRMSD.

ModelMean scTM ↑scTM >0.9 (%) ↑Cα scRMSD <2 Å (%) ↑
CodeFP0.894964.8443.75
ProDiT0.657633.5923.44
CodeFP20.905771.8852.34

CodeFP2 retains strong unconditional sequence–structure consistency after functional fine-tuning.


2. W2.1 ProDiT and ESM3 baselines.

We include ProDiT, ESM3 and ProGen in the test comparison. ProGen is trained on the same training set as CodeFP2. ESM3 uses the corresponding reference protein's InterPro annotations.

ProDiT's released interface conditions on one GO term at a time. For a multi-GO request, we compute each term's guidance relative to the unconditional prediction and average these directions at every denoising step.

HS rewards exact functional recovery and gives partial credit for recovering a corresponding broader function; min-HS averages the weakest function score within each request; ER (Exact Recall) measures exact recovery only (§4.1; Appendix A.1).

ModelHS ↑min-HS ↑ER ↑F1-micro ↑F1-macro ↑pLDDT ≥70 (%) ↑pTM ≥0.5 (%) ↑
CodeFP0.4470.3730.2590.4960.44680.783.5
CFP-Gen0.4150.3470.2410.4290.37075.572.3
ProDiT0.2530.1990.0780.4170.26586.289.7
ESM30.1070.0740.0080.2210.03030.027.2
ProGen0.2860.2270.1370.3520.25051.657.9
CodeFP20.4660.3950.2780.5060.47389.288.3

CodeFP2 achieves the best functional recovery across HS, min-HS, ER and F1 in this comparison, together with strong foldability.


3. W2.2 OOD performance beyond CodeFP.

Our original comparison focuses on CodeFP because it is the strongest prior baseline for functional recovery and foldability in Table 1 and shares the DPLM-2 backbone. We now include CFP-Gen, ProGen and ProteoGAN on the OOD test set, with 350 conditions and five candidates per condition. CFP-Gen, ProGen and ProteoGAN are trained on the same 95,627-protein training set as CodeFP2. ProteoGAN retains its native variable output lengths. We retain the two-subset presentation of Table 2.

SubsetModelHS ↑min-HS ↑ER ↑F1-micro ↑F1-macro ↑pLDDT ≥70 (%) ↑pTM ≥0.5 (%) ↑
NaturalCodeFP0.3950.2810.2040.5260.39571.062.3
NaturalCFP-Gen0.2730.1880.1200.4050.23648.837.6
NaturalProGen0.2380.1630.0910.3520.15634.840.6
NaturalProteoGAN0.1410.0900.0180.2810.0640.00.0
NaturalCodeFP20.4570.3220.2610.5340.38290.786.2
HypotheticalCodeFP0.3330.1900.0810.4280.26362.567.3
HypotheticalCFP-Gen0.2270.1430.0440.3440.17128.922.0
HypotheticalProGen0.2310.1290.0370.3160.14243.252.1
HypotheticalProteoGAN0.1770.0940.0110.3510.0850.00.0
HypotheticalCodeFP20.3960.2370.1290.4520.29987.188.4

CodeFP2 achieves the highest HS, min-HS and ER on both OOD test subsets, with strong foldability.


4. W3 CodeFP versus the retrained baseline without OAFC and CARE.

The training recipes differ. As stated in §4.1 (Implementation details) and Appendix C.3, CodeFP2 and all its ablations use full-parameter fine-tuning. CodeFP keeps the pretrained DPLM-2 backbone frozen and trains the added components. Thus, removing OAFC and CARE gives CodeFP-style conditioning retrained under the CodeFP2 recipe; it does not reproduce CodeFP's training configuration. Table 3 holds the training recipe fixed across ablations. We will make this distinction explicit in its caption.

Comments on the paper

1 comment
Area_Chair_7JACOpenReview ↗

Please respond to the rebuttals

Hi reviewers,

Thank you for your reviewing service. The authors have provided fairly extensive rebuttals. Please confirm that you have read them, state which of your concerns are or are not addressed, and whether that changes your recommendations.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.