acceptodds
Under review as a conference paper at ICLR 2027

LAVEP:LANGUAGE GUIDED PROTEIN LANGUAGE MODELS FOR VARIANT EFFECT PREDICTION

Abstract

Predicting how sequence variants alter protein function is central to protein engineering and variant interpretation. Yet protein fitness is assay-dependent, while most zero-shot protein predictors are assay-agnostic. Protein language models (PLMs) provide transferable sequence priors, but typically assign a fixed score to a mutation regardless of whether an experiment measures binding, activity, stability, expression, or another phenotype. Supervised methods can capture assay specificity, but rely on target-assay labels that are experimentally costly and time-consuming to obtain and often scarce in practice, limiting their use for new assays. We introduce a natural-language-guided framework that uses assay descriptions to tell a frozen PLM what “fitness” means in the experiment at hand. We treat language as an open-vocabulary task specification: assay context is grounded in the protein representation and used to steer the PLM's mutation prior toward the measured phenotype, without fine-tuning the PLM or requiring target-assay labels. We further establish a fundamental limit of assay-agnostic prediction. When the same variants are ranked differently across assays, any shared predictor must compromise between distinct fitness landscapes, imposing an intrinsic ceiling on attainable performance. Conditioning on assay language allows the model to represent these experimental objectives separately and escape this shared-ranking constraint. Across homology-separated ProteinGym assays, our approach consistently improves zero-shot mutation-effect prediction on unseen assays across ranking and classification objectives, showing the value of assay language precisely where new experimental labels are limited.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
protein fitness prediction, protein language model, assay-conditioned learning, zero-shot and few-shot learning

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_pc3KOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

LAVEP uses natural-language assay descriptions to condition a frozen protein language model on the experimental phenotype. It learns the conditioning module across assays and supports few-shot adaptation through kernel-based residual correction.

Critical strengths
  1. The paper addresses the limitation of assigning identical mutation scores to different experimental objectives.
  2. Homology-separated evaluation and text-intervention experiments provide useful evidence for assay-aware prediction.
  3. The framework supports both prediction without target-assay labels and adaptation from limited labeled variants.
Critical weaknesses
  1. Existing controls do not fully isolate detailed language semantics from supervised transfer.
  2. Cross-assay support pooling complicates comparisons under nominal few-shot budgets.
  3. The central multi-assay evidence covers few proteins, and some claims overstate the reported results.
Questions for authors
  1. The no-text intervention disables conditioning only at inference, so it does not separate language benefits from cross-assay supervised training. Please retrain matched controls with no text and with categorical phenotype labels to establish the value of detailed assay descriptions.
  2. In Eq. (7), each protein-to-text attention operation has only one text key/value, so its attention weight is always one. Please clarify this operation’s role and how it differs from simply broadcasting a projected text vector.
  3. Pooling support variants across assays means that a nominal K-shot prediction can use more than K labels per protein. Please report total label budgets and clarify whether competing methods receive the same cross-assay support information.
  4. Only two proteins in the multi-assay analysis come from the held-out test split. Please report these results separately and verify that the added NUDT15 and LDLR pairs have no protein or homology overlap with training.
  5. Section 5.3 claims the strongest multi-assay performance across all budgets, but Table 5 favors PRIMO at K = 16: 0.170 versus 0.149. Please correct this statement and add uncertainty estimates for the multi-assay comparisons.
1 comment
Reviewer_pc3KOpenReview ↗

response

Based on the author's response, I believe most of my concerns have been addressed; therefore, I will maintain my rating.

Reviewer_Kc2QOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The effect of a variant is assay specific so a single PLM score can't take that into account. LAVEP takes the assay description and splits it into 5 fields, uses a frozen text encoder to encode those fields and then uses bidirectional cross-attention to link those fields to the mutation related features of ESM-2. Gated corrections then pass through the frozen LM head to produce an assay conditioned log-odds which is combined with the original PLM score (a kernel ridge residual is added if labels are available). The authors have derived an upper bound on the average Spearman rank correlation that any assay agnostic predictor can achieve and have also stated a sufficient condition under which conditioning will do better than that upper bound. The evaluation is performed on a homology separated ProteinGym split in both zero shot and few shot regimes, an assay text intervention study and cross assay rank agreement for 4 proteins.

Critical strengths
  1. The framing is correct in that if you view the assay description as being analogous to a task specification with an open vocabulary then it's better suited to that than as a closed category label. In the zero-shot case there is no requirement for labels from the target assay and that's precisely the situation we have with a new assay.

  2. That's why I think the primary claim is correct. The method used in the text intervention study provides a far more rigorous control than a simple random mismatch. You're only changing the bits that are specific to the assay and the protein context is held constant. The fact that a rewrite which preserves the meaning has such a minimal effect goes directly against the idea that surface phrasing is important. Such high quality controls are very rare in the protein text world.

Critical weaknesses
  1. The whole argument is based on one seed really. The module ablation shows a case with no intervals and the gaps between the various variants are small and can be attributed to seed noise. The authors do point out that retrained ablations can in fact make up for this via the other pathways. My own approach would be to do the ablation with a number of seeds with intervals and include a control where the text is left as is but cross-attention is replaced by something very simple (e.g. FiLM style modulation from the pooled field embeddings). The architectural claim is only going to be weakened if the full model fails to separate in that scenario.

  2. The cross-assay study data is for four proteins, two of which are not from the benchmark. Only two of the multi-assay proteins are from the held out test set, the other two having come from another source. In one case the agreement is actually quite poor (although this is not mentioned in the main text). It would seem that either the number of multi-assay proteins should be increased (with appropriate caveats regarding those from other sources) or the individual protein results should be included in the main text and the case of poor agreement should be dealt with explicitly.

  3. The theory is right, it just has nothing to do with the method. Theorem 1 is a valid Cauchy-Schwarz bound (which is true), it's an upper bound and there's no indication that there are assay agnostic predictors which actually achieve anything near to it. The sufficient condition in Corollary 1 has never been tested on any data.

  4. The two systems are not capacity comparable as LAVEP has an 8B text tower plus some additional trainable parameters and is trained on a number of assays jointly. Neither METALIC nor PRIMO have the text tower and the label free baselines have it neither. It's therefore impossible for the reader to tell whether the effect is due to language conditioning or to the extra capacity and supervision.

  5. The benefit obtained with a small number of labels goes down as the number of labels increases (even though the text doesn't state this). With the highest budget the relative gain over ProteinNPT is very small and Fig 2C shows that the advantage mentioned in the text actually becomes negative after a certain moderate number of labels have been reached. The fair conclusion is that the method does make a difference when the number of labels is low.

Questions for authors
  1. What are the zero-shot figures if we don't include the protein information at test time?

  2. The interpolation weight will tend to favour the conditioned score. How sensitive is performance to that weight and given that the weight was derived from validation assays, how should it then be set for a completely new assay?

  3. The cross-assay agreement for LDLR is definitely lower than that for the other three. What's the reason for this? Are the two readouts in fact correlated, or have the descriptions failed to properly differentiate between them? If the former is the case, would a rho_ab be available for that pair?

  4. The reduction in the number of related mismatches is tiny. Doesn't that point to the model being good at distinguishing between quite different assays and having very poor resolution with respect to the related assays on the same protein? Doesn't that contradict the assertion in Section 5?

Reviewer_pCacOpenReview ↗
Rating
2
Confidence
3(+1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes LAVEP, a language-guided framework for protein variant effect prediction. The central idea is that mutation effects depend on the phenotype and experimental conditions being measured, whereas conventional protein language models generally assign the same mutation score regardless of the assay. LAVEP addresses this problem by encoding assay descriptions into five semantic fields: experimental context, measured readout, score semantics, experimental conditions, and protein information. These text representations interact with mutation-specific protein representations through bidirectional cross-attention, allowing the model to adjust its predictions according to the assay. The framework supports both zero-shot prediction on unseen assays and few-shot adaptation through kernel ridge regression. Experiments on homology-separated ProteinGym assays show improvements over several existing assay-agnostic and cross-assay prediction methods.

Critical strengths
  1. An interesting and relevant problem formulation The paper highlights an important limitation of conventional protein variant prediction: the meaning of fitness is not universal across experiments. Using natural language as an open-vocabulary task specification is a reasonable way to provide experimental context without requiring target-assay labels. The five-field representation also gives the framework a relatively interpretable structure.
  2. Broad evaluation of assay-aware prediction The authors evaluate zero-shot and few-shot settings and include text-intervention experiments to examine whether the model actually uses assay descriptions. The homology-separated test split is also appropriate for assessing transfer to unseen assays. LAVEP achieves a zero-shot Spearman correlation of 0.485, compared with 0.455 for the strongest assay-agnostic baseline, and continues to perform well with limited support labels.
Critical weaknesses
  1. Limited evidence for cross-assay ranking changes One of the most interesting claims is that the same mutation can receive different rankings depending on the assay being evaluated. However, only two proteins in the held-out ProteinGym test set have multiple eligible assays. The authors add two further protein pairs from MaveDB, resulting in an analysis involving only four proteins. This is a relatively small basis for a general claim about assay-dependent mutation effects. More examples would help establish whether the method captures a broadly applicable biological relationship rather than patterns specific to a few proteins.
  2. Limited statistical validation of the architectural contributions The main implementation uses a single training seed, and the module ablations are also reported as single-seed results. Some of the architectural differences are relatively small. For example, the full model obtains a Spearman correlation of 0.4855, while removing text grounding gives 0.4820. Without repeated runs and uncertainty estimates, it is difficult to determine how robust these differences are. The theoretical analysis is mathematically reasonable, but its main ranking bound follows directly from Cauchy–Schwarz. Its contribution is therefore more useful as a problem formulation than as a substantial independent theoretical advance.
Questions for authors
  1. Can the authors provide multiple-seed results and confidence intervals for the main benchmark and module ablations? This would help establish whether the reported improvements are robust.
  2. Can the authors expand the same-variant cross-assay evaluation beyond four proteins? It would also be useful to clarify how assay descriptions are constructed and how their completeness is assessed when experimental conditions are missing.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.