acceptodds
Under review as a conference paper at ICLR 2027

Extending Controllability from Composition and Topology to Domain Identity in Language-Model Protein Generation

Abstract

Protein language models can generate plausible sequences, but reliable control over multiple user-specified properties remains limited. We present Aria, which tunes Llama-3.1-8B-Instruct on  350k Swiss-Prot proteins described by a random subset of thirteen physicochemical, developability, structural, and functional descriptors written into the prompt, and which measures adherence by recomputing every descriptor from the generated sequence using the same code that wrote the prompt. On cluster-separated evaluation proteins, Aria improves adherence across directly measurable properties, including length, hydrophobicity, cysteine count, transmembrane topology, and signal peptides. Prompt-mismatch controls show that these gains depend on the requested specification. Exact Pfam-domain recovery remains difficult and increases with family exposure. We next test whether natural-sequence-anchored direct preference optimization can improve weaker structural and functional axes. One optimization stage raises mean ESMFold pLDDT by 8.2 points in PF00416, ribosomal protein uS13, and increases exact requested-domain recovery by 23.6-46.0 percentage points across three additional optimized families, with corresponding increases in predicted structural similarity. However, applying the resulting family-specific adapters to six families not used for preference optimization does not improve target-domain recovery over the supervised model. These results show that a general instruction-tuned model can acquire broad protein-property control without protein-specific pretraining, while natural-sequence-anchored preference optimization provides substantial but family-local gains.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Protein design, Large language models, Protein language models, Controllable protein generation, Property-conditioned sequence design, Instruction tuning, Direct preference optimization, Structure-guided optimization

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_K7ysOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents a multi-attribute controllable protein sequence generation framework based on a general language model - Aria. The authors fine-tuned Llama-3.1-8B-Instruct using QLoRA and the training data came from approximately 349,000 Swiss-Prot protein sequences. This model demonstrated significant improvement in the adherence to sequence-level attributes (such as length, hydrophobicity, cysteine count, transmembrane topology, and signal peptide). However, in extensive evaluations, the ability to recover the required Pfam domains was still limited.The authors also introduced a DPO method based on natural sequences, that is, using natural protein sequences as the preferred response and rejecting the generated sequences with low confidence. In experiments on multiple protein families, the results showed that the confidence of ESMfold structure prediction, the recovery rate of Pfam domains, and the predicted structural similarity all improved. However, these improvements mainly relied on specific protein families and there is no sufficient evidence to show that they have effective transferability across families.

Critical strengths

1.Systematic assessment of the controllability of multi-attribute proteins.This paper conducts a comprehensive experimental analysis of controllability at various levels, including sequence composition, topological structure, and domain identity. By using grouped separated evaluation data, rapid misalignment control, and multiple specific attribute indicators, the empirical results are further strengthened. The attributes that can be directly calculated from the sequence and those that require external predictors are distinguished, indicating that after the specific instructions of proteins are optimized, general-purpose language models can or cannot reliably control certain aspects.

  1. Significant empirical improvement and detailed analysis of preference optimization.The preference optimization experiment based on natural sequences shows that in multiple protein families, the prediction of structural confidence and the recovery rate of the requested domains have both improved.
Critical weaknesses
  1. Limited methodological novelty and insufficient baselines.The framework primarily combines existing instruction tuning and DPO techniques. Comparisons with the untuned Llama model and ProLLaMA are insufficient, as ProLLaMA cannot receive the same property specifications. Stronger baselines under matched conditioning settings are needed.

  2. Insufficient DPO ablations and limited generalization.The improvements from anchored DPO may partly result from additional supervised learning on natural sequences. Comparisons with continued SFT or NLL-only training are missing. Moreover, overall Pfam recovery remains low (2.9%), and family-specific gains transfer poorly, raising concerns about generalizability.

Questions for authors

1.The comparison with ProLLaMA is somewhat difficult to interpret, since ProLLaMA does not receive the same property specifications as Aria. Could the authors include a protein-specific baseline adapted to the same conditioning inputs? Alternatively, a retrieval-based or constraint-filtering baseline would help establish how much improvement comes from the proposed approach. 2. I am not fully convinced that the improvements after anchored DPO are primarily due to preference optimization. Since natural protein sequences are used as preferred responses and an additional NLL loss is included, some gains might simply come from further supervised training. A comparison with continued SFT under a similar training budget would help clarify this point. 3. While the improvements in individual properties are encouraging, it remains unclear how often the generated proteins satisfy multiple constraints simultaneously. In particular, the strict composite metric does not include Pfam-domain recovery. It would be useful to report joint success rates incorporating domain identity and structural properties, especially for protein families with limited representation in the training data.

Reviewer_frwYOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents Aria, a controllable protein sequence generator built by instruction-tuning Llama-3.1-8B-Instruct on approximately 350K Swiss-Prot sequences conditioned on subsets of 13 physicochemical, structural, and functional properties. The authors further apply natural-sequence-anchored Direct Preference Optimization (DPO), using natural proteins as preferred responses and model-generated sequences as rejected responses, with ESMFold pLDDT guiding preference-pair selection. The experiments show that supervised fine-tuning substantially improves adherence to compositional properties (e.g., length, hydrophobicity, and cysteine count) and topological features (e.g., transmembrane helices and signal peptides). However, exact Pfam domain recovery remains limited at 2.9%. Subsequent family-specific DPO improves predicted structural confidence, domain recovery, and structural similarity across several evaluated protein families, although improvements in property adherence are inconsistent. The authors also compare Aria with ProLLaMA and investigate cross-family transfer. They find that Aria achieves better multi-property adherence, while domain recovery varies across models and families. Importantly, DPO improvements do not consistently transfer to unseen protein families. Overall, the study demonstrates the feasibility of adapting general-purpose LLMs for multi-property protein generation while highlighting the difficulty of controlling domain identity and generalizing family-specific optimization.

Critical strengths
  • The paper demonstrates that a general-purpose LLM can be adapted to generate protein sequences conditioned on multiple user-specified properties without protein-specific pretraining. Supervised fine-tuning substantially improves adherence to physicochemical and topological constraints. The prompt-mismatch controls further support that these improvements arise from conditioning on the requested properties rather than simply learning common sequence distributions. This provides useful empirical evidence for the feasibility of instruction-based, multi-property protein generation.
  • The paper evaluates controllability across different levels of biological complexity, revealing a clear gap between composition, topology, and domain identity. The additional DPO experiments show substantial improvements in domain recovery and predicted structural similarity across several protein families, while cross-family transfer experiments reveal that these improvements are largely family-specific. Although the underlying methods are relatively straightforward, these experiments provide useful insights into both the capabilities and limitations of current approaches to controllable protein generation.
Critical weaknesses
  • The DPO stage consistently uses natural sequences from the target protein family as preferred responses, together with an auxiliary negative-log-likelihood loss. Consequently, improvements in domain recovery may primarily result from further fitting the model to the target family's natural sequence distribution, rather than from optimizing structural confidence. The paper lacks critical controlled comparisons, such as family-specific SFT or DPO with natural positive sequences but without pLDDT-based preference selection. Although a self-play DPO comparison is provided, multiple training factors differ simultaneously, making causal attribution difficult. This limitation is consequential because the DPO stage is the paper's main approach to extending controllability toward domain identity, yet the specific contribution of its optimization strategy remains unclear.
  • While supervised fine-tuning substantially improves compositional and topological controllability, exact Pfam domain recovery remains extremely low (2.9%) in the broad evaluation. The subsequent DPO improvements are largely restricted to families explicitly used for preference optimization. In the cross-family transfer experiments, domain recovery decreases in four of six predefined nearby-family pairs and remains unchanged in the other two. Furthermore, strict-composite performance decreases in 25 of 36 source-target-seed comparisons. These findings suggest that the approach relies heavily on target-family-specific training rather than learning broadly transferable principles of domain-level protein generation. This substantially limits the generality of the proposed framework and the significance of its claimed extension to domain identity.
Questions for authors
  • Given the low overall domain recovery rate after SFT (2.9%) and the largely negative cross-family transfer results, to what extent does the proposed framework actually learn generalizable domain-level controllability?
Reviewer_CpATOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper adapts Llama-3.1-8B-Instruct to generate protein sequences from partial textual specifications covering thirteen descriptors. Supervised tuning on approximately 349,000 Swiss-Prot proteins improves adherence to directly computable properties and predicted topology, while broad Pfam-domain recovery remains low. A second stage uses natural sequences as preferred responses in a DPO objective with an auxiliary likelihood term. Family-specific optimization improves predicted confidence, domain recovery, and structural similarity, but does not consistently transfer to other families. The paper also studies prompt-mismatch controls, reward hacking, and the difference between family membership and compliance with additional property requests.

Critical strengths
  1. The paper exposes meaningful differences between kinds of control.

  2. The paper's evaluation asks whether the model follows the actual specification. Recomputing the six directly measurable descriptors avoids relying only on sequence plausibility.

Critical weaknesses
  1. Appendix A states that clustering uses 30% identity and 80% coverage and that no held-out protein exceeds 30% identity to any training protein (p. 12). However, Appendix C reports that natural held-out PF00416 proteins have mean nearest-training identity 0.707 and that 29 of 260 exceed 90% identity.

  2. Every preferred response is a natural sequence already used in supervised training, and the objective includes a preferred-response negative-log-likelihood term with weight 1.0. There is no matched continued-SFT control using those same natural sequences.

Questions for authors
  1. How does the combined DPO-plus-likelihood objective compare with continued SFT on the same preferred natural sequences, matched for update budget? Does preference optimization improve held-out domain recovery or joint success beyond that baseline?

  2. For a subsample of PF00467/PF03652 generations, do the TM-score and confidence gains reproduce with a second structure predictor?

2 comments
AuthorsOpenReview ↗

Response to question 2

We thank the reviewer for suggesting this control. We agree that variability between structure predictors can affect the results. As per reviewer’s suggestion, for both families, uL24 (PF00467) and YqgF (PF03652), we used OmegaFold to predict structures for the same generated sequences and their accession-matched natural references. We then repeated the TM-score comparison using the same alignment procedure as for ESMFold. At the round-three DPO checkpoint, both mean predicted confidence and mean TM-score increased relative to supervised Aria in both families at both seeds.

For uL24, the OmegaFold pLDDT gains were 9.2 and 5.3 points at the two seeds, and the corresponding TM-score gains were 0.294 and 0.250. For YqgF, pLDDT increased by 14.6 and 13.7 points, and TM-score increased by 0.156 and 0.167. The TM-score gains were close to those obtained with ESMFold on the same generations: 0.290 and 0.242 for uL24, and 0.153 and 0.185 for YqgF. The matched continued-SFT checkpoints also improved both OmegaFold measures in both families at both seeds. We added a section to the appendix (Section M) to report these findings.

Thank you for the detailed responses and additional experiments. The clarification addresses most of my concerns. I will maintain my positive score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.