acceptodds
Under review as a conference paper at ICLR 2027

CHIME: Multimodal Prediction of Histone Modification Profiles across Adult Mouse Brain Cell Types

Abstract

Comprehensive profiling of histone modifications across the adult mouse brain remains experimentally costly. A recent integrative single-cell epigenomic atlas profiles over 2.5 million nuclei yet covers only nine brain regions, leaving the cerebellum, olfactory bulb, and additional midbrain areas unprofiled. Completing histone modification maps therefore requires predicting active and repressive chromatin features for cell types lacking histone measurements. We introduce CHIME, which integrates chromatin accessibility, DNA methylation, RNA expression, and DNA sequence features to predict four continuous histone modification signals at 200-bp resolution. CHIME leverages a distance-aware representation to integrate local and broader genomic context, paired with a position-preserving decoder that predicts cell-type-specific deviations from a shared histone profile. Evaluated via five-fold cross-validation across 118 cell types with test targets completely withheld, CHIME achieves a mean positional correlation of 0.737 and surpasses all evaluated baselines on every metric across all folds. Matched removals show complementary contributions from chromatin accessibility, DNA methylation, and gene expression. With reference profiles from training cell types available on each chromosome, models trained on chromosome 19 transfer without parameter updates to 18 additional autosomes, achieving a mean correlation across genomic positions of 0.734. These results support computational extension of brain epigenetic maps to cell populations lacking histone measurements.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
multimodal learning, epigenomic imputation, histone modification prediction, cell-type generalization, sequence prediction, regulatory genomics

Reviews

Among:
Shape:
1234this paper, Top 17%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.75 over 4 reviews, above 83% of the others.
Reviewer_8LyQOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This authors propose CHIME, a multimodal model for predicting four histone modification profiles at 200-bp resolution in adult mouse-brain cell types lacking histone measurements. This is an important area of research and can have broad impact on the community. Overall, inputs include approximately 1 Mb of cell-type-specific chromatin accessibility and DNA methylation, a 64-dimensional cell-type RNA representation, and a frozen Evo 2 embedding of the central 1-kb DNA sequence. The architecture combines position-preserving local convolutions, global attention, and an RNA/DNA-gated distance-bucket representation. It predicts cell-type-specific deviations from a locus-wise histone reference profile computed using training cell types.

The paper shows detailed and thorough work on distinguishing shared genomic signal from true cell-type-specific prediction, reporting correlations across cell types and after subtracting the training-cell reference is a strength. Whole-chromosome evaluation and discussion of amplitude compression and reference-assisted transfer are appreciated. But the current experiments don't establish the paper’s broader claim of extending maps to genuinely novel brain populations or regions. The split distributes cell types within 15 major groups across folds, so it primarily evaluates interpolation within the atlas taxonomy rather than withholding an entire lineage or region. The paper is promising but needs work on harder distribution-shift evaluation, better matched controls, and amplitude-sensitive biological validation.

Critical strengths
  1. Computational completion of cell-type-specific histone maps could help prioritize expensive experimental profiling.

  2. The paper explicitly addresses the known pitfall that shared locus activity can make cross-cell-type prediction appear stronger than it is.

  3. Test-cell histones are excluded from model fitting, target scaling, and reference construction within each fold.

  4. Evaluating all eligible stitched positions is more persuasive than reporting only hand-selected regulatory regions.

  5. The authors acknowledge amplitude compression, the need for training-cell reference profiles during chromosome transfer, dependence among cross-validation folds, and the descriptive nature of the fold-level tests.

  6. CHIME’s cell-type-specific metric exceeds the implemented baselines in all five reported folds, and Table 5 in appendix shows the substantially weaker results for the two repressive marks rather than hiding them in the four-mark average.

Critical weaknesses
  1. The split distributes cell types within 15 major groups across folds. Most or all test groups are therefore likely represented in training, although the manuscript does not report the exact group-by-fold counts. A model may recognize broad lineage or subclass identity and interpolate a group-associated histone pattern. This can make the task easier.

  2. CHIME’s final prediction directly adds a locus-specific mean histone profile μp,m\mu_{p,m} computed from training cell types; the neural network itself predicts a residual around this profile. This is a sensible design, but Table 1 does not provide the exact performance of μ\mu alone. Consequently, the headline rpos=0.737r_{\mathrm{pos}}=0.737 cannot be interpreted as the amount of signal learned from the target cell’s modalities. Figure 3 visually compares an earlier MSE-only model with the reference, but it does not supply the required numeric reference baseline for the final model.

  3. Pearson correlation is insensitive to scale. Section 5 and Figure 5 acknowledge that predicted cell-type deviations are compressed in amplitude, and Figure 2b warns that its measured and predicted heatmaps use separate scales. The missing issue is quantitative calibration: a model can have a positive residual correlation while badly underestimating biologically relevant effect sizes.

  4. According to section B, the MLP and CNN predict raw histone targets, do not add the training-cell reference, do not use the same deviation representation or soft-clipped targets, optimize MSE alone, and use different checkpoint rules. CHIME additionally has a correlation loss explicitly aligned with rcellr_{\mathrm{cell}}. Thus, the comparison changes output parameterization, objective, preprocessing, and architecture simultaneously.

  5. The authors should separately ablate the distance branch, global attention, and correlation loss while matching all other training and selection choices. Multiple seeds should be run across more than one fold. The main text should identify the Table 1 objective confound immediately and avoid presenting its 0.031 gap as a branch-only effect.

Questions for authors
  1. How exactly were subclass-level populations mapped among the Paired-Tag, ATAC, and methylation atlases? How many mappings were ambiguous, and how sensitive are the results to cell counts and the mCG coverage waivers?

  2. Are all modality-removal rows in Table 1 trained with exactly the same correlation loss, target transform, and checkpoint rule as the full model? The relationship among Tables 1, 4, 6, and 8 is difficult to follow.

  3. Why do the 24 regional diagnostics in Table 13 use the earlier MSE-only model rather than the final CHIME model?

  4. What are the exact cell types in each fold, and what is each test population’s nearest training population by taxonomy and input similarity?

Reviewer_WL1cOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents a model for imputing histone mods in new cell types, using as input DNA sequence, RNA-seq, ATAC-seq and histone mods in other cell types. The motivation is to characterize cell types in mouse brain for which such profiles are missing. The model is compared to state-of-the-art baseline methods and appears to deliver better performance.

Critical strengths

This is a real problem, and the model appears to make a significant improvement compared to state-of-the-art methods.

I really enjoyed the framing of the scientific hypotheses (lines 300-303) and the systematic way in which they were addressed.

Critical weaknesses

The paper was largely clear, except for the description of the model itself. In particular, I found Figure 1b to be very difficult to understand, and the accompanying text did not help very much. It is not clear to me what "delta window" and "raw + delta" refer to, nor what "per-bin stems" are. The variable "c" is not defined in the figure, and is only defined in Section 3.4. In general, making the figure more abstract and then writing the methods section more clearly would have been preferable.

Questions for authors

Why did you limit the training set to chromosome 19? It seems like you could have a more parameter-rich model if you had >10x more training data.

Minor additional questions:

Is each panel in Figure 2b ordered independently? Wouldn't it make more sense to order them the same, so for i on the left corresponds to row i on the right?

How did you select those particular four modifications to profile?

What is the validity mask? It is introduced on line 156 but never defined.

Incidentally, I think the chromosome transfer experiment is worth reporting, but you should probably acknowledge that, in practice, this setting will not arise.

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

CHIME predicts four histone modification signals from chromatin accessibility, DNA methylation, RNA expression, and frozen DNA embeddings. It combines local convolutions, global attention, and a gated distance-dependent representation, then predicts deviations from the mean histone profile of training cell types. Evaluation holds out histone measurements for cell types across five folds, with training on chromosome 19 and reference-assisted transfer to other autosomes.

Critical strengths

1- The evaluation directly tests cell-type-specific prediction rather than relying on overall signal reconstruction. With all histone targets withheld for test cell types, the paper evaluates both differences across cell types at each locus and deviations from the training-cell reference. CHIME improves the primary cell-type correlation to 0.192, compared with 0.134 for the strongest evaluated baseline on this metric. This is consequential because it provides evidence that the model captures target-specific information beyond shared genomic structure.

2- The model demonstrates reference-assisted transfer across chromosomes. Models trained on chromosome 19 are applied without parameter updates to 18 additional autosomes, achieving mean positional and residual correlations of 0.734 and 0.247, respectively. This is important because it supports the practical value of learning a mapping that extends beyond the genomic loci used for model fitting.

Critical weaknesses
  1. Some ablations change both the model and the training loss, while the CNN baseline uses a different training setup. It is therefore unclear how much of the improvement comes from the proposed architecture.
  2. Related cell populations remain in training, and chromosome transfer still uses measured histone references. It is not clear how the results establish how well CHIME would work on the missing populations that motivate the paper
Questions for authors
  1. Some ablations use MSE alone, whereas the full model also uses a correlation loss. What evidence separates the contribution of the architectural components from that of the loss?
  2. How does the CNN perform when trained with the same reference decomposition, target transformation, loss, and checkpoint-selection procedure as CHIME?
Reviewer_CJFWOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors propose CHIME, a multimodal predictive framework based on multi-omics data that provides a computational extension of brain epigenetic maps to cell populations lacking histone measurements. The authors have benchmarked on several high impact downstream tasks, as well as tried to demonstrate fidelity to known biological pathways for trustability, but the key takeaway is that this pipeline should work even with partially missing histone data, thus having higher usability potential in real world settings.

Critical strengths

Same as paper summary, repeated here:

The authors propose CHIME, a multimodal predictive framework based on multi-omics data that provides a computational extension of brain epigenetic maps to cell populations lacking histone measurements. The authors have benchmarked on several high impact downstream tasks, as well as tried to demonstrate fidelity to known biological pathways for trustability, but the key takeaway is that this pipeline should work even with partially missing histone data, thus having higher usability potential in real world settings.

Critical weaknesses

I have two major comments here from the methodology point of view that I feel strongly about, that is regarding robustness of information flow and reliability of downstream predictions -

  1. regarding robustness of information flow, though you have showed that information flows through biologically relevant concepts and pathways, does the work measure any information leakage between these, which would confound the apparent interpretability (think of concept leakage as an indirect analogy)... is transparency overestimation due to information leakage a possibility here?

  2. regarding reliability of downstream predictions, you have mentioned yourself "Reported fold SD cannot be used as the uncertainty of a difference between two models evaluated on a single fold. Such a comparison requires paired measurements for the competing models across the same folds and seeds." - so what is the solution then, as a semi-rigorous UQ is extremely crucial when using AI in this kind of scenario, perhaps a calibration based frequentist approach (like conformal prediction as an example, but could be something else), can be used to provide a marginal coverage, is that possible, and if not, explain why?

I will not comment on the multi-omics and biomedical side of things being a computer scientist myself, hopefully other reviewers will cover the applicative aspect.

Questions for authors

Same as the weakness section, repeated here.

I have two major comments here from the methodology point of view that I feel strongly about, that is regarding robustness of information flow and reliability of downstream predictions -

  1. regarding robustness of information flow, though you have showed that information flows through biologically relevant concepts and pathways, does the work measure any information leakage between these, which would confound the apparent interpretability (think of concept leakage as an indirect analogy)... is transparency overestimation due to information leakage a possibility here?

  2. regarding reliability of downstream predictions, you have mentioned yourself "Reported fold SD cannot be used as the uncertainty of a difference between two models evaluated on a single fold. Such a comparison requires paired measurements for the competing models across the same folds and seeds." - so what is the solution then, as a semi-rigorous UQ is extremely crucial when using AI in this kind of scenario, perhaps a calibration based frequentist approach (like conformal prediction as an example, but could be something else), can be used to provide a marginal coverage, is that possible, and if not, explain why?

I will not comment on the multi-omics and biomedical side of things being a computer scientist myself, hopefully other reviewers will cover the applicative aspect.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.