acceptodds
Under review as a conference paper at ICLR 2027

Leveraging Biological Priors for Protein Fitness Prediction

Abstract

Accurately predicting the properties of proteins is a foundational problem in protein science and engineering. Models trained across experimental assays can learn transferable patterns, but their success depends on the quantity and quality of available data. On the other hand, biological priors informed by signals such as sequence and structure are valuable for predictions in the low-data regime, but they may encode restrictive assumptions that do not accurately reflect empirical evidence. We introduce \method, a data-efficient protein fitness prediction framework that bridges these approaches. First, we incorporate sequence and structural information through protein language model embeddings and inverse folding features. Second, rather than relying solely on experimental data, we turn biological models into synthetic data generators, pretraining on the resulting fitness prediction tasks before finetuning on real assays. We find that synthetic pretraining improves downstream performance over training on experimental data alone, and our approach outperforms strong baselines across held-out ProteinGym assays.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
protein fitness, synthetic pretraining, prior-data fitted networks

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

This paper trains a transformer that fuses the sequence and structure embedding of variants (context) to predict the fitness. It is tested on two settings: real experimental labels and synthetic labels predicted from Kermut. The experiments are done in ProteinGym training/test/val split.

Critical strengths

Experiments and methodology are sound.

Critical weaknesses
  1. I think this is sound paper with good empirical gains. However, the novelty and impact is limited to recommend accept. The idea of encoding multiple variants and training a transformer to predict fitness already existed. (ProteinNPT and PRIMO).
  2. Synthetic-only setting still depends on experimental supervision. L344-345 "This allows us to generate tasks over a distribution of GP settings which are informed by experimental data." I think it is misleading to call it synthetic-only.
  3. I also thought about if this method can be useful for practitioners. But Figure3 suggest at context size of 32-64 Proficient performs similar to Kermut. So the case where Proficient can be useful is very narrow, low-N setting where N=4-32.
Questions for authors
  1. How does choice of context affects the performance? What if you split the evaluation assays to make sure, instead of random, the context is chosen from held-out positions? I think ProteinGym has supervised split provided within assay (modulo, random, ...), what's the performance if you select context only from one bucket of the split and evaluate on the rest?
  2. Synthetic pretraining lets you train outside the ProteinGym training assays. Is there particular reason for limiting to ProteinGym training assays?
  3. How's the performance if synthetic training uses zero-shot scores instead of experimentally fitted scores?

However I think these will not materially affect my rating. I think the novelty and impact is fundamentally limited. For revision and future submission I would recommend exploring methods for other synthetic pretraining (learning from stability labels, learning from AlphaFold confidence, etc) or studying deeper in "why" rather than just showing the performance. Or submitting to soundness-prioritized journals like TMLR might make sense.

5 comments
AuthorsOpenReview ↗

Response [1/3]

Thank you for recognizing the soundness of our experiments and the empirical improvements. We first clarify an important distinction that affects the assessment of our method’s practical impact, then explain our methodological and scientific contributions and address your questions.

Figure 3 suggests that at context sizes of 32-64, PROFICIENT performs similar to Kermut.

Figure 3 does not show our final method but instead evaluates our real-only ablation. The purpose of Figure 3 is explicitly to demonstrate the limitations of training a model only on real data, which motivates the synthetic pretraining that we use in Sections 4 and 5. Our full PROFICIENT method, seen in Section 5 and Figure 1, combines synthetic pretraining with real-assay finetuning and improves over Kermut across all evaluated context sizes, including N=64N=64 and N=128N=128.

Furthermore, low-data prediction is practically relevant to many protein engineering settings. Previous studies have engineered fluorescent proteins and enzymes using as few as 24 experimentally measured variants [1], identified high-performing protein variants by testing only 12 candidates per round [2], and designed functional fluorescent proteins using just 64 training examples [3]. These studies highlight the practical importance of learning from small experimental datasets, particularly when measurements are expensive or time-consuming.

The idea of encoding multiple variants and training a transformer to predict fitness already existed

ProteinNPT and PRIMO do establish the use of transformers to predict fitness, but train their fitness predictors on empirical data. They do not use biologically informed synthetic fitness tasks to complement learning from real assays. Our methodological contribution is a biological task generator and a training framework that encodes domain knowledge through synthetic tasks and combines it with learning from experimental assays.

Our contributions are both methodological and empirical: we develop a data generator for protein fitness landscapes and a new framework that combines synthetic pretraining with experimental data. We also establish when and how these different training sources improve protein fitness prediction.

Methodological Novelty

We combine biological priors with experimental data for in-context fitness prediction.

  • PFNs primarily provide a framework for learning from purely synthetic tasks. Our framework combines synthetic pretraining with real assays for in-context protein fitness prediction.
  • Kermut incorporates biological priors and fits measurements from each target assay, but it does not learn across experimental assays. Our framework combines both sources of knowledge.
  • Metalic and PRIMO meta-train on experimental data alone. Our paper demonstrates that adding biologically informed synthetic pretraining improves over training on real assays alone. Furthermore, their best-performing variants require assay-specific parameter updates, while our model predicts new assays entirely in-context without parameter updates.
  • METL combines both synthetic and real data, but it requires assay-specific finetuning.
  • ProteinNPT trains a separate model for each assay on experimental data, without learning across assays.

We construct a novel prior distribution over protein fitness landscapes.

  • Designing domain-specific priors is a significant methodological contribution in recent ML work, including KinPFN for RNA folding times (ICLR 2025) [4], MapPFN for single-cell perturbation effects (NeurIPS 2026) [5], and Prior-Fitted Functional Flows for pharmacokinetics (UAI 2026) [6]. Our work develops a synthetic training distribution over protein fitness landscapes in this same spirit.
  • METL also uses synthetic data, but each example maps a sequence to a fixed target. In our setting, synthetic data plays a fundamentally different role: rather than specifying a fixed set of target functions, it specifies a distribution over functions.
  • Kermut is a GP fitted to individual assays. We instead turn its biological kernel into a training task generator, producing a synthetic predictor that outperforms Kermut for N≤32N \le 32 and matches it at larger context sizes (Fig. 5). This requires new techniques described in Section 4:
    1. Because Kermut refits its hyperparameters on each target assay while our model cannot, we choose to build this uncertainty into the training distribution. We developed a prior over GP hyperparameters that generates biologically informative training tasks.
    2. We found that Kermut often performed worse than its zero-shot score input. We therefore blend sampled fitness values with zero-shot scores to preserve useful evolutionary signal in the generated training tasks.
AuthorsOpenReview ↗

Response [2/3]

Scientific Contributions

We substantially improve accuracy and label efficiency for protein fitness prediction. Across 33 held-out assays, PROFICIENT achieves a mean Spearman correlation of 0.671 with only four labeled variants, outperforming other baselines even when those baselines are given 32 labels. It also achieves higher mean Spearman correlation than other baselines at every evaluated context size, without requiring assay-specific parameter updates.

We find that biological priors and experimental data have complementary strengths. We might expect synthetic landscapes to be an imperfect substitute for real data. Instead, we find that the two are strong in different regimes. Training on real assays yields strong predictions from very few measurements, but the performance saturates as the context grows. Biological priors, in turn, perform well for larger contexts.

We demonstrate that our framework provides an effective way to combine biological priors with empirical data. Since the two are strong for different context sizes, we might expect that combining them would improve one regime at the expense of the other. Instead, our framework outperforms both standalone models at every context size, indicating that it captures useful predictive information from both sources.

We show that even biological priors which do not lead to strong standalone predictions can improve performance over models which are only trained on experimental data. Our new ablations (Fig. 8) demonstrate the performance of models using various biological priors. Adding synthetic pretraining consistently leads to improvements over real-only training, demonstrating the benefits of our framework.

Our contributions extend beyond combining existing components: we introduce methodological innovations and establish new findings about how biological priors and experimental data contribute to protein fitness prediction.

References

  1. Biswas et al., Low-N Protein Engineering with Data-Efficient Deep Learning, Nature Methods 2021
  2. Chu et al., Accurate Top Protein Variant Discovery via Low-N Pick-and-Validate Machine Learning, Cell Systems 2024
  3. Gelman et al., Biophysics-Based Protein Language Models for Protein Engineering, Nature Methods 2025
  4. Scheuer et al., KinPFN: Bayesian Approximation of RNA Folding Kinetics using Prior-Data Fitted Networks, ICLR 2025
  5. Sextro et al., MapPFN: Learning Causal Perturbation Maps in Context, NeurIPS 2026
  6. Ojeda et al., Prior-Fitted Functional Flows: In-Context Generative Models for Pharmacokinetics, UAI 2026
AuthorsOpenReview ↗

Response [3/3]

I would recommend exploring methods for other synthetic pretraining

Thanks for the suggestion. Your suggestion motivated us to add new ablations to better understand the impact of the biological prior. Specifically, we have added Section 5.3, where we compare four alternative biological priors:

  1. RBF Kernel: Samples from a GP with an RBF kernel between ESM-2 sequence embeddings and a constant mean function
  2. RBF + ESM2 mean: Uses the same RBF kernel, with ESM-2 zero-shot scores as the mean function
  3. Kermut Kernel: Samples from a GP with the Kermut kernel and a constant mean function
  4. Kermut + ESM2 mean: Uses the Kermut kernel with ESM-2 zero-shot scores as the mean function, without PROFICIENT's additional zero-shot label blending described in Section 4

We see in Figure 8 that:

  • Every tested prior improves downstream performance over real-only training, including priors with relatively weak standalone performance.
  • Real-assay finetuning narrows the performance gap between different priors, showing that experimental learning can compensate for some of the limitations of the synthetic data distribution.
  • Prior design remains important. Even though finetuning narrows the performance gap between priors, the relative ordering is broadly preserved, and stronger priors tend to remain stronger after finetuning.

These results also motivate the exploration of alternative biological models in future work, and we have highlighted these directions in our discussion.

The synthetic-only settings still depends on experimental supervision.

We use "synthetic" to distinguish training the predictor on generated labels from training it directly on experimental labels. As we describe in Section 4, experimental measurements are used to inform a small set of GP hyperparameters, which capture broad properties of fitness landscapes such as their smoothness and the relative importance of sequence and structural similarity. Our goal is to use this compact information to generate useful training tasks, not to eliminate experimental information from the pipeline. The relevant question is therefore whether these generated tasks can add value compared to directly training on experimental assays, and we see in Figure 1 that the biological priors do provide complementary sources of information. We have updated Section 4 to make this distinction explicit.

How does the choice of context affect the performance?

We evaluate sensitivity to context selection using five independently sampled context sets per assay and context size, reporting the mean and standard deviation across draws. This measures the robustness of our results across the sampled contexts.

For the rebuttal, we have also added evaluations using the different ProteinGym supervised splits, reported in Appendix A.2. We sample the context exclusively from one fold and evaluate on single-mutant variants in the remaining four folds. We find that PROFICIENT retains its advantage over Kermut and the synthetic and real-only ablations across the three splits. These results demonstrate that combining synthetic pretraining with real-assay finetuning is beneficial across many contexts.

Why restrict synthetic generation to ProteinGym?

ProteinGym provides a standardized benchmark and lets us use the same setup as existing methods such as PRIMO. This makes the comparison more informative, because we can test the benefit of including biological priors while using the same available training assays. ProteinGym also spans diverse proteins and measured properties, including stability, activity, and fluorescence, which provides a good foundation for developing our method. We agree that our framework is not restricted to this dataset, and extending generation to additional settings would be an interesting direction for future work.

How's the performance if synthetic training uses zero-shot scores instead of experimentally fitted scores?

Could you clarify whether you mean using zero-shot scores directly as training labels? We sample synthetic labels from the GP prior, not its fitted posterior, and already incorporate zero-shot scores through the GP mean function and label blending (Section 4).

We hope these clarifications and new experiments support reconsideration of the paper’s novelty and impact. Together, our biological task generator, training framework, and new findings establish how biological priors can improve learning from experimental assays.

Reviewer_Nt2zOpenReview ↗
Rating
4
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces PROFICIENT, a data-efficient framework for protein fitness prediction that combines biological priors with experimental data. The central innovation is using Gaussian processes to generate synthetic fitness landscapes for pretraining, followed by fine-tuning on real experimental assays. The authors demonstrate strong performance across 33 held-out ProteinGym assays, particularly in low-data settings.

Critical strengths

I think this is an excellent paper. The central idea of using biological priors to generate synthetic fitness landscapes for neural network pretraining is clever and potentially broadly applicable. Rather than imposing restrictive assumptions directly on a model, this approach allows the model to learn from those assumptions and subsequently refine them using experimental data.

Critical weaknesses

I have no major concerns. One limitation is that the synthetic landscapes are generated using a particular Gaussian process kernel, and it remains unclear how well the approach generalizes to fitness landscapes that differ substantially from these assumptions. However, I view this primarily as an opportunity for future work.

Questions for authors

How sensitive is PROFICIENT to the choice of biological prior used to generate synthetic training tasks? Would alternative kernels or other biological models provide similar benefits?

1 comment

Thank you for your thoughtful and encouraging review. Your suggestion to explore alternative priors motivated new experiments that strengthened both the paper and our understanding of the approach.

How sensitive is PROFICIENT to the choice of biological prior?

We added Section 5.3 to test sensitivity to prior choice, comparing four alternative biological priors:

  1. RBF Kernel: Samples from a GP with an RBF kernel between ESM-2 sequence embeddings and a constant mean function
  2. RBF + ESM2 mean: Uses the same RBF kernel, with ESM-2 zero-shot scores as the mean function
  3. Kermut Kernel: Samples from a GP with the Kermut kernel and a constant mean function
  4. Kermut + ESM2 mean: Uses the Kermut kernel with ESM-2 zero-shot scores as the mean function, without PROFICIENT's additional zero-shot label blending described in Section 4

We see in Figure 8 that:

  • Every tested prior improves downstream performance over real-only training, including priors with relatively weak standalone performance.
  • Real-assay finetuning narrows the performance gap between different priors, showing that experimental learning can compensate for some of the limitations of the synthetic data distribution.
  • Prior design remains important. Even though finetuning narrows the performance gap between priors, the relative ordering is broadly preserved, and stronger priors tend to remain stronger after finetuning.

These results also motivate the exploration of alternative biological models in future work, and we have highlighted these directions in our discussion.

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

This paper introduces a data-efficient protein fitness prediction framework (PROFICIENT) that incorporates biological knowledge through its inputs and its training procedure. The method combines protein language model embeddings and inverse folding features in a flexible protein encoder. Authors turn biological priors into synthetic data generators and pretrain the model on synthetic tasks generated from these priors. After pretraining, they finetune the model on real assay data, allowing the model refine the biological relationships learned during pretraining to the specific structure of real fitness landscapes. The authors evaluate PROFICIENT on 33 assays with different context sizes. The experimental results show that synthetic pretraining followed by real-assay finetuning achieves a higher performance, compared with training only using synthetic data or real data, respectively.

Critical strengths
  1. This paper combines protein language model embeddings and inverse folding features in a flexible protein encoder to capture complementary sequence and structural information. Then, pretrain the model on synthetic tasks generated from biological priors, and finetune the model on real assay data. The results show that such precedure achieves better prediction performance than training using real data only or synthetic data only.

  2. The authors evaluate the proposed framework PROFICIENT on 33 assays with different context sizes. The results show that PROFICIENT outperforms other baselines.

Critical weaknesses
  1. Figures 1 and 2 present performance comparisons averaged over 33 assays. However, per-assay results are needed to determine whether the improvements are consistent across most tasks or driven by a few specific assays. I would also encourage the authors to provide a quantitative comparison table with existing methods and baselines, rather than relying solely on line charts and bar charts.

  2. I have some concerns regarding the methodological novelty. The core idea is similar to Mutational Effect Transfer Learning (METL) [1], which pretrains on biophysical simulation data and fine-tunes on experimental data. The synthetic task formulation also closely follows PFNs [2], while Kermut [3] is another relevant approach. Overall, PROFICIENT appears to primarily combine existing ideas. The authors are encouraged to better clarify its technical novelty and differences from these methods. The paper would benefit from an overall framework diagram illustrating the proposed method and model architecture. Although textual descriptions are presented, it remains unclear which components represent the key innovations and which are adopted from existing methods.

[1] Gelman S, Johnson B, Freschlin C R, et al. Biophysics-based protein language models for protein engineering[J]. Nature Methods, 2025, 22(9): 1868-1879.

[2] Müller S, Hollmann N, Arango S P, et al. Transformers can do bayesian inference[J]. arXiv preprint arXiv:2112.10510, 2021.

[3] Groth P M, Kerrn M H, Olsen L, et al. Kermut: Composite kernel regression for protein variant effects[J]. Advances in Neural Information Processing Systems, 2024, 37: 29514-29565.

Questions for authors
  1. Could the authors clarify the key technical innovations of PROFICIENT compared with METL [1], PFNs [2], and Kermut [3]? What are the fundamental differences beyond combining existing approaches?

  2. Could the authors provide per-assay comparison results for the 33 assays? Are the reported improvements consistent across most assays, or mainly driven by a few specific tasks?

5 comments
AuthorsOpenReview ↗

Response [1/2]

Thank you for your review and constructive suggestions. We have strengthened the paper with new experiments and expanded comparisons, and we highlight the resulting findings below.

I have some concerns regarding novelty

Our contributions are both methodological and empirical: we develop a data generator for protein fitness landscapes and a new framework that combines synthetic pretraining with experimental data. We also establish when and how these different training sources improve protein fitness prediction.

Methodological Novelty

We combine biological priors with experimental data for in-context fitness prediction.

  • PFNs primarily provide a framework for learning from purely synthetic tasks. Our framework combines synthetic pretraining with real assays for in-context protein fitness prediction.
  • Kermut incorporates biological priors and fits measurements from each target assay, but it does not learn across experimental assays. Our framework combines both sources of knowledge.
  • Metalic and PRIMO meta-train on experimental data alone. Our paper demonstrates that adding biologically informed synthetic pretraining improves over training on real assays alone. Furthermore, their best-performing variants require assay-specific parameter updates, while our model predicts new assays entirely in-context without parameter updates.
  • METL combines both synthetic and real data, but it requires assay-specific finetuning.
  • ProteinNPT trains a separate model for each assay on experimental data, without learning across assays.

We construct a novel prior distribution over protein fitness landscapes.

  • Designing domain-specific priors is a significant methodological contribution in recent ML work, including KinPFN for RNA folding times (ICLR 2025) [1], MapPFN for single-cell perturbation effects (NeurIPS 2026) [2], and Prior-Fitted Functional Flows for pharmacokinetics (UAI 2026) [3]. Our work develops a synthetic training distribution over protein fitness landscapes in this same spirit.
  • METL also uses synthetic data, but each example maps a sequence to a fixed target. In our setting, synthetic data plays a fundamentally different role: rather than specifying a fixed set of target functions, it specifies a distribution over functions.
  • Kermut is a GP fitted to individual assays. We instead turn its biological kernel into a training task generator, producing a synthetic predictor that outperforms Kermut for N≤32N \le 32 and matches it at larger context sizes (Fig. 5). This requires new techniques described in Section 4:
    1. Because Kermut refits its hyperparameters on each target assay while our model cannot, we choose to build this uncertainty into the training distribution. We developed a prior over GP hyperparameters that generates biologically informative training tasks.
    2. We found that Kermut often performed worse than its zero-shot score input. We therefore blend sampled fitness values with zero-shot scores to preserve useful evolutionary signal in the generated training tasks.
AuthorsOpenReview ↗

Response [2/2]

Scientific Contributions

We substantially improve accuracy and label efficiency for protein fitness prediction. Across 33 held-out assays, PROFICIENT achieves a mean Spearman correlation of 0.671 with only four labeled variants, outperforming other baselines even when those baselines are given 32 labels. It also achieves higher mean Spearman correlation than other baselines at every evaluated context size, without requiring assay-specific parameter updates.

We find that biological priors and experimental data have complementary strengths. We might expect synthetic landscapes to be an imperfect substitute for real data. Instead, we find that the two are strong in different regimes. Training on real assays yields strong predictions from very few measurements, but the performance saturates as the context grows. Biological priors, in turn, perform well for larger contexts.

We demonstrate that our framework provides an effective way to combine biological priors with empirical data. Since the two are strong for different context sizes, we might expect that combining them would improve one regime at the expense of the other. Instead, our framework outperforms both standalone models at every context size, indicating that it captures useful predictive information from both sources.

We show that even biological priors which do not lead to strong standalone predictions can improve performance over models which are only trained on experimental data. Our new ablations (Fig. 8) demonstrate the performance of models using various biological priors. Adding synthetic pretraining consistently leads to improvements over real-only training, demonstrating the benefits of our framework.

Our contributions extend beyond combining existing components: we introduce methodological innovations and establish new findings about how biological priors and experimental data contribute to protein fitness prediction.

In response to your feedback, we have also updated our paper: we added a contribution section at the end of our Introduction, and we have further revised our Related Work to better differentiate our method from prior work. Thank you for suggesting a diagram to clarify our key innovations! We have added a comparison table that compares the training data, cross-assay learning, and assay-specific adaptation of our method and related approaches, and we hope this further clarifies our methodological contributions.

Per-assay results are needed

Thank you for the suggestion! We have provided new per-assay results in Appendix A, Tables 1 and 2. We find that PROFICIENT consistently outperforms other evaluated baselines across most assays, performing the best on 32/33 assays at N=4N=4 and 31/33 assays at N=32N=32. These results demonstrate that our improvements are consistent across assays rather than driven by only a few specific tasks.

Please let us know if you have any other questions.

References

  1. Scheuer et al., KinPFN: Bayesian Approximation of RNA Folding Kinetics using Prior-Data Fitted Networks, ICLR 2025
  2. Sextro et al., MapPFN: Learning Causal Perturbation Maps in Context, NeurIPS 2026
  3. Ojeda et al., Prior-Fitted Functional Flows: In-Context Generative Models for Pharmacokinetics, UAI 2026
AuthorsOpenReview ↗

Response Followup

Thank you again for your constructive feedback! We wanted to follow up on your questions:

  • We have provided new per-assay results: Our method outperforms the other approaches on 32/33 assays at N=4N=4 and 31/33 assays at N=32N=32, indicating that our method consistently provides better predictions.
  • We have clarified our methodological and scientific contributions.: We have provided a list of our contributions at the end of the introduction, expanded upon the novelty of our method compared with existing methods, and provided a new table highlighting the differences in approaches.

We hope these additions address your questions, and we would appreciate it if you could reconsider your evaluation of the paper. Please let us know if there is anything else we can clarify.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response

We thank the reviewers and the AC for their careful reading and constructive feedback. We include this general response to clarify our paper's methodological and scientific contributions and summarize the new evidence added during rebuttal.

Contributions

  • We develop PROFICIENT, a data-efficient framework for in-context protein fitness prediction. It combines biological synthetic pretraining with real-assay finetuning and predicts new assays without parameter updates. Kermut does not learn across experimental assays; Metalic and PRIMO meta-train on experimental data alone; METL requires assay-specific finetuning.

  • We show that biological priors and experimental data have complementary strengths. Real-assay training yields strong predictions from very few measurements but saturates as the context grows, while biological priors perform well for larger contexts.

  • We construct a prior distribution over protein fitness landscapes. We turn Kermut's biological kernel into a training task generator, building uncertainty over GP hyperparameters into the training distribution and blending sampled fitness values with zero-shot scores (Section 4). Designing domain-specific priors is itself a methodological contribution, as illustrated by KinPFN for RNA folding times [1], MapPFN for single-cell perturbation effects [2], and Prior-Fitted Functional Flows for pharmacokinetics [3].

  • We demonstrate that encoding biological knowledge as training data improves protein fitness prediction. Synthetic pretraining followed by real-assay finetuning outperforms training on either source alone at every context size (Fig. 1). Even priors with weak standalone predictions improve over real-only training (Fig. 8).

  • We improve accuracy and label efficiency for protein fitness prediction. Across 33 held-out assays, PROFICIENT achieves a mean Spearman correlation of 0.671 with only four labeled variants, outperforming other baselines even when those baselines are given 32 labels.

Rebuttal Updates

  • Prior ablations. We have added Section 5.3, where we compare four alternative biological priors. Every tested prior improves downstream performance over real-only training, including priors with relatively weak standalone performance. Real-assay finetuning narrows the performance gap between different priors, while stronger priors tend to remain stronger after finetuning.
  • Per-assay results. We have provided new per-assay results in Appendix A, Tables 1 and 2. PROFICIENT performs best among the compared baselines on 32/33 assays at N=4N=4 and 31/33 assays at N=32N=32. These results demonstrate that our improvements are consistent across assays rather than driven by only a few specific tasks.
  • Supervised split settings. We have added evaluations using ProteinGym's random, modulo, and contiguous supervised splits. We sample the context exclusively from one fold and evaluate on single-mutant variants in the remaining four folds. PROFICIENT retains its advantage over Kermut and both the synthetic and real-only ablations across all three split settings (Appendix A.2).
  • Clarified contributions and related work. We have added a contribution section at the end of our Introduction and revised our Related Work to better differentiate our method from prior work. We have also added a comparison table in Appendix B that compares the prediction tasks, training data, and adaptation procedures of PROFICIENT and related methods.

References

  1. Scheuer et al., KinPFN: Bayesian Approximation of RNA Folding Kinetics using Prior-Data Fitted Networks, ICLR 2025
  2. Sextro et al., MapPFN: Learning Causal Perturbation Maps in Context, NeurIPS 2026
  3. Ojeda et al., Prior-Fitted Functional Flows: In-Context Generative Models for Pharmacokinetics, UAI 2026

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.