acceptodds
Under review as a conference paper at ICLR 2027

BE-CORR: Learning to Correct Source-Library Outcomes for Endogenous Base-Editing Prediction

Abstract

DNA base editors enable targeted single-nucleotide changes and are promising tools for functional genomics and therapeutic genome engineering. Machine learning has become an effective approach for predicting base-editing outcomes, helping prioritize editor-guide designs early in development. Many predictors are trained on integrated or synthetic target-library assays (source libraries), where many targets are measured in controlled, non-native contexts. However, practical editing occurs at endogenous genomic loci, where chromatin and regulatory state can substantially shift outcomes relative to source libraries. Existing approaches typically target either source-library prediction or direct endogenous prediction: the former does not directly transfer to native-locus behavior, while the latter is constrained by scarce endogenous training data. We therefore reformulate endogenous outcome prediction as source-to-endogenous correction: obtain a source outcome profile as a source-library prior, predict a context-induced shift from endogenous features such as chromatin accessibility and transcriptional activity, and recover the endogenous profile as the source profile plus the shift. We introduce BE-CORR (Source-to-Endogenous Correction of Base-Editing Outcomes), an interpretable and robust correction model that represents the source profile as nonnegative editing hazard, rescales overall editing opportunity, and redistributes per-position opportunity through constrained hazard transport with reliability-gated blending. Experiments show that BE-CORR consistently outperforms strong sequence-only, context-aware, and source-using baselines in endogenous endpoint accuracy, ranking, calibration, matched source-to-endogenous transfer, and correction gain. Code and data are available at https://anonymous.4open.science/r/BE-CORR.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
DNA base editing, endogenous outcome prediction, biological sequence modeling, transfer 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_BSBCOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The main motivation of this paper is that large-scale training data are collected in a source-library setting, whereas applications occur in an endogenous setting. BE-CORR formulates endogenous prediction as source-to-endogenous correction, it first obtains a source-library prior, then predicts context-dependent correction terms for the endogenous context, and finally combines them to form the endogenous prediction. BE-CORR is evaluated on endogenous endpoint prediction and matched source-to-endogenous transfer benchmarks.

Critical strengths
  1. The motivation is practical and clear. The chromatin context is critical to endogenous base-editing outcomes and could be emphasized to make accurate endogenous predictions.

  2. I find the measured source–endogenous benchmark is also valuable. The reconstruction of matched pairs by concrete editor construct and exact 40-nt context is a good experimental contribution. It allows the evaluation of a setting in which the source outcome is actually observed, rather than inferring transfer quality only from a predicted source profile. Also, BE-CORR includes hazard scaling, local transport, and reliability gating to model overall editing strength, location changes, and source reliability, which provide interpretable endogenous outcomes to some degree.

I also checked the main-text mathematical claims and all of them are coherent and looks sound to me.

Critical weaknesses

Due to the 1-2 critical weakness limit of this block, I moved few weaknesses to the Question section.

  1. Limited contribution of problem reformulation: The previous study BE_Endo already systematically investigated the differences between source-library and endogenous editing outcomes and showed the influence of endogenous genomic context on editing outcomes. BE-CORR recasts this widely known problem as source-to-endogenous correction and lists it as one of its main contributions. This reformulation provides limited additional insight beyond adding the source profile as an additional input.

    Moreover, Sec. 5 describes the adaptations largely as retaining released encoders or feature and replacing task-specific heads. Some adaptation is understandable because the methods have different native outputs. However, the resulting comparisons establish performance against the specified adaptations; they do not automatically establish superiority over the complete published methods.

  2. Unclear Advantages Over Simpler Residual Correction Methods: The central claim of BE-CORR is to incorporate source-library profiles into context-aware prediction to learn source-to-endogenous correction, which is essentially a form of residual correction of source-library predictions. Although BE-CORR includes hazard scaling, local transport, and reliability gating, the current experiments do not prove that adding these modules improves prediction accuracy over simple source-conditioned residual correction.

Questions for authors
  1. Table 9 shows that the model without correction context retains approximately 71% of the absolute improvement over direct prediction. This does not mean endogenous context is not important, since the direct branch still receives endogenous features, and the full model performs better.

    A convincing study would remove or permute context separately in the correction and direct branches, and report the results on the measured-source benchmark.

  2. I also have reservations regarding the hazard representation claim. The expected number of edited bases is the sum of marginal editing probabilities. BE-CORR instead conserves the sum of transformed hazards during transport. These are different quantities.

  3. The biological assumption of Hazard transport is questionable: BE-CORR assumes that changes in editing activity at different positions can be represented by mass-conserving hazard transport, which means that reduced editing activity at one position can be redistributed to other positions. However, in reality, editing activity at multiple positions may increase or decrease simultaneously, and such changes do not necessarily follow a mass-conserving transfer mechanism. Additional biological evidence should be provided to support the assumption of mass-conserving transport.

  4. Figure 1B is not sufficient to prove the usefulness of the reliability gate: The results show that when the model relies more on source predictions, the average prediction gain is higher. This only shows a correlation between source utilization and model performance; it does not prove that the gate can determine whether the source prediction is more accurate than the direct prediction for an individual sample. For example, some samples may be easier to predict and achieve higher performance regardless of source reliability.

  5. Missing FORECasT-BE baseline: FORECasT-BE (Pallaseni et al., 2022) is a relevant method for predicting position-specific base-editing outcomes but is missing from the baseline comparison.

  6. Potentially unfair CRISPRon-BE comparison: In Appendix G, auxiliary features such as CRISPRoff energy and Cas9 activity are set to zero, which may weaken the baseline's performance. Why adopt this setting?

  7. Unclear source-prior generation protocol: Does the correction model use in-sample or out-of-fold predictions from the source predictor during training? Differences in source-prior quality between training and testing may lead to a distribution mismatch.

1 comment

Dear Reviewer BSBC,

As the author–reviewer discussion period is nearing its end, we would be grateful if you could take a look at our rebuttal and revised manuscript when your schedule permits.

We would welcome any further questions or feedback and would be happy to continue the discussion. We understand that this is a busy period and sincerely appreciate your time and thoughtful consideration.

Best regards,

The Authors

Reviewer_L8njOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

Most base-editing predictors are trained on library assays, where many targets are measured in an artificial setting. In real use, though, editors work on endogenous loci, where chromatin changes the result. The paper treats this as a correction problem: start with the library profile, predict how the cellular context shifts it, and add the two together. BE-CORR does this in three steps. First, it converts the library profile into a measure of editing “exposure” at each editable position, and a context model predicts how much that exposure should go up or down overall and where it should move. Second, it shifts the exposure only to nearby editable positions, keeping the total the same. Third, it blends the result with a separate predictor that doesn’t use the library data. The more the two disagree, the less weight the library-based version gets. The paper also gives a proof with three results. It limits how much the library-based branch can hurt accuracy. It also shows when that branch helps: when its errors differ from the direct predictor’s errors, even if it is less accurate overall. The tests use 9,116 endogenous records and 8,450 matched library–endogenous pairs. They compare against several types of baseline, including adapted genomic foundation models, over five splits.

Critical strengths

• The design fits the problem, and the checks show it behaves as described. Splitting the correction into “how much editing happens here” and “where in the window it lands” matches how editors actually go wrong. Restricting the second part to small local moves between chemically editable positions builds in prior knowledge instead of hoping the model finds it. The authors also tested whether that constraint actually binds. Almost all of the redistribution is a shift of one position or less. That diagnostic could have gone against them, and it didn’t. • The theory earns its place. The most useful result is the condition for when leaning on the library measurement helps. It depends on the two estimates making different kinds of mistakes, not on the library estimate being more accurate. That is a somewhat counterintuitive point to establish. It justifies the blending step over simply picking whichever predictor looks better, and I think it would carry over to other problems with the same structure.

Critical weaknesses

• The paper’s own ablation suggests chromatin features add far less than the framing implies, and this goes undiscussed. The paper argues that native context shifts editing outcomes and must be modelled. Yet an appendix table shows that removing the chromatin features keeps most of the gain, and shuffled features land in about the same place. Most of the improvement seems to come from the hazard representation, local redistribution and fallback gate, not from context. The numbers are reported honestly but left unremarked. This matters because someone deciding whether to generate chromatin data for a new cell type could take the wrong message from the abstract. • The data is narrow in exactly the place that matters here. The study covers one adenine editor, one cytosine editor and one cell line. The locus features are the cell-type-specific ones, so they are the least likely to transfer to another setting. That means the part least tested for generalisation is also the part contributing least. The authors do acknowledge this limitation. Separately, several comparison methods are run in modified form because the benchmark doesn’t supply everything they were built to use. This is set out clearly and in good faith, but the comparisons against those methods shouldn’t be read as strongly as the abstract presents them.

Questions for authors
  1. The appendix ablation shows that most of the gain remains when the endogenous features are removed. Please clarify how the authors interpret this result. Is the contribution of context concentrated in particular feature types or at particular classes of locus? The abstract may also direct readers toward the wrong component. The result does not weaken the method. It indicates that the constrained correction accounts for most of the improvement, and the paper should state this explicitly.
  2. The methods receiving a library profile also benefit from the library data behind it, which the sequence-only and context-aware rows never see. Your comparison against the nearest competitor is unaffected, since it has the same inputs, so this is only about how the wider table reads. Would you consider noting the asymmetry where those cross-group comparisons are drawn?
  3. One suggestion: the low-data result in the appendix, a clear advantage when only a small share of endogenous labels is available that narrows as more arrive, is the best answer in the paper to the question of when someone should reach for this method. It speaks directly to the motivation you open with, and it would serve readers well in the main text.
Reviewer_PJ7dOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper predicts endogenous base-editing outcomes by correcting source-library profiles. BE-CORR scales and redistributes source-derived editing hazards, then combines the result with a direct endogenous predictor through a reliability gate. The authors evaluate the method on BE_Endo data using either predicted or experimentally measured source profiles. They report improvements in prediction error and ranking, supported by ablations and low-data experiments.

Critical strengths

(1) The paper addresses a relevant problem. Editing outcomes measured in artificial target libraries may differ from those at endogenous sites. Learning to correct this difference is a reasonable approach. The distinction between predicted and experimentally measured source profiles also makes the two use cases clear. (2) The reported improvements are consistent across both settings. Compared with FT-Transformer, window MAE decreases from 2.76 to 2.44 percentage points with predicted source profiles and from 2.54 to 2.16 with measured profiles. The nonlinear control, no-transport ablation, and low-data experiment provide useful support for these results.

Critical weaknesses

(1) Both benchmarks in Tables 1 and 2 come from the same BE_Endo experimental system and cover only two editor constructs in HEK293T cells (Appendix B). The results mainly support prediction within this system. I am not convinced that they establish generalization to new cell types, editors, or experimental conditions. This limits the scope of the conclusions. (2) The comparisons do not fully establish the need for the proposed complexity. Table 6 shows that several methods use replacement output heads and two auxiliary input channels of CRISPRon-BE are set to zero. These adaptations need validation. BE-CORR also includes a separately trained direct predictor and several gating components. Table 9 includes an MLP and a no-transport variant, but not a simple residual correction model using the same direct branch and gate. This makes it difficult to isolate the contribution of transport. (3) The architecture and interpretability evidence are not presented clearly enough. Section 3 and Appendix C provide equations and pseudocode, but no overall architecture diagram or complete network configurations. Figure 3 mainly shows correlations and binned trends, with no concrete target-level example comparing the source profile, corrected prediction, and measured outcome. I cannot tell from these results whether the learned redistribution has a convincing biological interpretation. (4) Eq. 12 shows that the source branch can help when its errors complement those of the direct predictor. However, Eq. 8 reduces the source weight as the two branches disagree more. This may suppress useful corrections when the direct predictor is wrong. Figure 1A tests corrupted source inputs, but does not show what happens when the source branch is more accurate despite strong disagreement. Table 9 also lacks an ablation of this attenuation term.

Questions for authors

(1) Section 3 and Appendix C describe the main operations. What are the architectures and parameter counts of the source predictor, direct predictor, correction encoder, and gate? In Algorithm 1, is the direct predictor frozen during correction training? Are training-set source predictions generated through cross-fitting? (2) Do the information-matched baselines in Tables 1 and 2 use the same editable-position masks, window-weighted loss, and comparable tuning budgets? What implementation and training checks were performed for the baseline adaptations listed in Table 6? (3) In addition to the controls in Table 9, how does a simple residual correction model perform when combined with the same direct predictor and gate? What are the separate contributions of learned reliability, disagreement attenuation, and global shrinkage in Eq. 8? (4) Can you report the results in Tables 1 and 2 separately for ABE and CBE and restrict evaluation to editable positions? Have you evaluated the method on an endogenous dataset independent of BE_Endo? If not, what prevents that evaluation?

1 comment

Dear Reviewer PJ7d,

As the author–reviewer discussion period is nearing its end, we would be grateful if you could take a look at our rebuttal and revised manuscript when your schedule permits.

We would welcome any further questions or feedback and would be happy to continue the discussion. We understand that this is a busy period and sincerely appreciate your time and thoughtful consideration.

Best regards,

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.