acceptodds
Under review as a conference paper at ICLR 2027

From Black-Box Predictors to Scientific Instruments: Biological Rule Faithfulness in Genomic Foundation Models

Abstract

Genomic foundation models are becoming essential tools for AI-driven genomics research. Although their downstream performance has been demonstrated across benchmarks, their black-box nature obscures the biological basis of predictions, limiting assessments of scientific reliability. Here we present a biological rule–based framework for evaluating the scientific reliability of genomic foundation models. Using RNA splicing as a test system, we identify biological component subspaces in frozen genomic foundation models and assess their interactions against experimentally established biological rules. Across six models with distinct architectures and pretraining strategies, we find that similar predictive performance can mask differences in component use and biological rule faithfulness (BRF). To assess whether BRF provides information beyond conventional benchmarks, we examined model pairs with conflicting rankings using measurements reported in independent published experimental studies. Across the evaluated settings, BRF favored the better-performing model in 57.1–88.9% of these comparisons. Grounded in biological rules, our framework reveals differences in underlying representations and task generalization beyond conventional benchmarks. These findings motivate evaluation and training paradigms for developing scientifically reliable foundation models for biological discovery.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Genomic foundation models, Scientific reliability, Model interpretability

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_7qBTOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes biological rule faithfulness (BRF) for evaluating genomic foundation models using RNA splicing. Controlled sequence edits identify component-associated subspaces, whose functional relevance is examined through propagated erasure and matched random interventions. Representations of experimentally assayed sequence states are then substituted into a common reference along joint component subspaces. BRF measures agreement between the resulting model interaction directions and experimental interaction directions. The paper also tests whether fixed, relation-specific BRF rankings help select models on external experimental tasks when conventional benchmark rankings disagree.

Critical strengths
  • The paper asks whether genomic models capture biological relationships that conventional predictive benchmarks may miss. Using splicing as a concrete test system gives this question a clear scientific motivation and makes the framework relevant to biologically informed model assessment.
  • The authors' internal analyses show that similar predictive performance can coexist with different dependence on the selected component subspaces. Their evaluation on external experimental data further suggests that relation-specific scores can provide useful model-selection information. Together, these findings support the framework as a complement to standard benchmarks within the studied tasks.
Critical weaknesses

The paper could offer a clearer interpretation of the BRF differences between models. The authors already acknowledge in Section 7 that pretraining objectives, architectures and data may all contribute. Connecting that discussion more directly to the observed results, including the role of the intervention and readout, would help readers understand what the comparison supports. The numerical reference and random-control results described in Appendix E, along with the training details of the human-PSI and yeast readouts used for BRF, also remain insufficiently reported. These details matter to the interpretation, even though the current set of checkpoints cannot isolate the causal effects of individual training choices.

Questions for authors
  • Could the authors point to or provide the numerical reference and matched-random control results described in Appendix E, and briefly clarify the training data, prediction targets, data splits and fitting procedure for the human-PSI and yeast readouts used in the BRF evaluation? This would help assess the basis for the reported model differences.

  • Could the authors briefly connect Section 7 to the observed BRF differences, considering bidirectional MLM versus causal next-token prediction, parameter scale and pretraining data? In particular, how might contextual access interact with the intervention and pooled readout, and what role might species coverage and splice-relevant sequence content play? I would find it helpful to distinguish patterns supported by the existing results from possible explanations, given the confounding among objectives, architectures and data.

  • The authors might also briefly discuss publicly released models such as NTv3’s MLM-pretrained models, GENERator and Caduceus-PS to place the work in a broader context. I would welcome such a discussion in the final version, without expecting the authors to undertake additional model experiments during the rebuttal period.

Reviewer_cyy9OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces Biological Rule Faithfulness (BRF), a framework for evaluating genomic foundation models by comparing model-derived component interactions with experimentally observed biological relationships. The authors identify component subspaces via controlled sequence perturbations and evaluate the framework across six foundation models using RNA splicing as a case study.

Critical strengths
  1. Introduces a biologically grounded evaluation perspective by examining interactions between biological components rather than focusing solely on prediction accuracy or single-feature attribution.

  2. Connects internal representation analysis with controlled biological perturbations and experimental interaction datasets across six foundation models, providing encouraging evidence of the framework's potential utility.

Critical weaknesses
  1. The framework is validated exclusively on RNA splicing, leaving its applicability to other biological tasks uncertain, particularly when component relationships are less well characterized.

  2. The metric relies primarily on directional agreement between interaction contrasts, which does not necessarily establish mechanistic understanding. Furthermore, BRF rankings show inconsistent alignment with downstream performance across biological relationships, raising concerns about how reliably the metric can guide model evaluation.

Questions for authors
  1. Could the observed directional agreement in identified subspaces arise from correlated or non-specific representations rather than the intended biological mechanisms? Given that BRF primarily measures interaction consistency, how should its relationship to scientific reliability and mechanistic understanding be interpreted?

  2. The paper includes supplementary analyses of continuous interaction contrasts. Would incorporating interaction magnitude or experimental uncertainty directly into BRF substantially alter model rankings compared to the current sign-based formulation?

  3. How can suitable biological components and experimental references be defined for tasks where mechanisms are less characterized than RNA splicing? What are the main practical limitations to extending the framework to other genomic tasks?

  4. The external evaluations show substantial variation across component relationships, including cases where BRF rankings underperform conventional benchmark rankings. What factors determine when BRF is informative for a downstream task, and how should practitioners select relevant component relationships for new evaluations?

Reviewer_4XAHOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper focusses on the problem of selecting reliable genomic foundation model for a give task. The authors propose Biological Rule Faithfulness (BRF) as a method that can evaluate whether a bioFM encodes biologically grounded relationships. Running case study in the paper is RNA splicing. In this context, their approach starts with identifying low-dimensional subspaces corresponding to four biological components ((a) donor sites, (b) acceptor sites, (c) polypyrimidine tracts and (d) branchpoints) through controlled sequence perturbations. Then, use these subspaces to assess whether model-predicted bio interactions align with experimentally observed ones. They show BRF is a better scorer/ranker for selecting performant models than using predictive accuracy.

Critical strengths
  • I like the motivation of evaluating genomic FMs beyond usual approach of predictive accuracy. Evaluating whether GFMs' learned representations are consistent with experimentally established biological relationships can help avoid using brittle models.

  • They carried out a comprehensive evaluation using 6 popular genomic FMs having diverse architectures and pretraining methods. The experiments include validation against experimentally measured biological interactions and showed that BRF can reveal differences between GFMs that are not uncovered by usual predictive accuracy based scoring.

Critical weaknesses

I am not an expert in RNA splicing so my concerns are mostly from a broader ML and scientific discovery perspectives.

It seems to me that the BRF framework assumes experimental data representing a unique biological ground truth. There are scientific applications where multiple mechanisms or context-dependent relationships may explain the same observations. In such applications, if experimental data covers only few possible behaviors, could BRF penalize models that capture other scientifically plausible relationships but misisng from experimental data? This raises questions about whether BRF measures biological understanding more generally or simply agreement with a particular experimental database. Furthermore, in many bio applications, bio data is either noisy or incomplete (confounding factors) if not entirely missing, BRF generalizability in these scenarios is not yet well understood.

Questions for authors
  • It would be useful if authors authors could clarify the aspects of BRF framework that are methodologically novel -- can they point differences wrt existing concept-based interpretability and subspace-intervention approaches?

  • Can BRF support tasks/applications where multiple mechanisms/rules are consistent with the available experimental observations? Would their metric penalize models that use alternative mechanisms/relationships but not represented in the experimental data used in BRF?

  • Can BRF generalize to tasks/applications with unknown or less clearly understood components and interactions?

  • How sensitive is BRF to incomplete, noisy, or potentially conflicting experimental measurements -- a common scenario in bio? Would it be useful to study this sensitivity and faithfulness?

  • Models that deviate from established biological known relationships can still make useful or novel predictions, right? How does BRF balance faithfulness to known/seen rules/settings with extrapolation / predictive performance in such unseen settings?

1 comment
AuthorsOpenReview ↗

Response to Question 3

We thank the reviewer for raising this question. We sincerely apologize that the original manuscript did not sufficiently explain how BRF can be used when component knowledge is incomplete. BRF requires components that can be defined and perturbed, together with corresponding experimental measurements, but does not require prior knowledge of the complete molecular mechanism or interaction direction. We conducted the following additional experiments to examine this issue.

1. Candidate regions provide a starting point when component definitions are unclear.

Using Fan’s regulatory-sequence data [1], we compared two approaches to constructing component subspaces. One used the authors’ component-targeted edits; the other introduced edits only within two predefined 15 bp regions, without using transcription factor names, motif rules, or precise target positions, while matching the number of edits and nucleotide substitution types. Across four models and three cell lines, 83.33% of pairwise model orderings were preserved. The top-ranked model remained unchanged in GM12878 and Jurkat but changed in K562. Region-based BRF differed from component-based BRF by −1.55 to +1.81 percentage points, and the Spearman correlations between their continuous interaction estimates ranged from 0.54 to 0.88.

These results show that, in this experiment, replacing precise component-targeted edits with candidate-region edits preserved most pairwise model orderings. This reduces reliance on precise component annotations, although the candidate regions themselves must still be specified in advance.

2. Combinatorial experiments establish the reference when interactions are unclear.

Under the same background and conditions, four states are measured: the reference state, A-only perturbation, B-only perturbation, and joint A/B perturbation. These measurements test whether the effect of A depends on the state of B. The interaction direction is obtained from the measurements, without prespecifying whether the components should act synergistically or inhibit one another, or first identifying the complete mechanism underlying the relationship. An incomplete understanding of how components interact therefore does not preclude evaluation; the key requirement is that they can be perturbed separately and their combined effects measured.

3. Settings where controlled component perturbations cannot be defined remain to be explored.

We acknowledge the challenge raised by the reviewer. At present, BRF requires, at a minimum, controlled component perturbations defined at a coarse level, such as candidate regions, together with corresponding combinatorial measurements. When components are entirely unknown and such perturbation comparisons cannot yet be established, the current workflow lacks a basis for identifying component subspaces. Without combinatorial measurements, an interaction reference cannot be established either. Future work might use experimental perturbations and their responses to guide subspace searches and subsequently localize candidate components. However, this possibility remains to be validated and lies beyond the scope of the present study.

References

[1] Fan et al. Expanding the DNA Motif Lexicon of the Transcriptional Regulatory Code. bioRxiv, 2025 (preprint).

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.