acceptodds
Under review as a conference paper at ICLR 2027

CoBALT: Coupling Antigen–Antibody Binding-Site Prediction and Affinity Ranking through a Contact-Energy Bottleneck

Abstract

Antigen–antibody (Ag–Ab) recognition involves binding-interface organization and collective intermolecular energetics, captured from complementary perspectives by binding-site prediction (BSP) and binding-affinity ranking (BAR). Although intrinsically connected, the two tasks operate at different resolutions: BSP localizes residue-level interfaces, whereas BAR evaluates complex-level affinity. Unifying these perspectives requires reconciling their distinct signals under scarce joint annotations. We introduce CoBALT, a unified framework that treats BSP and BAR as complementary readouts of a shared Ag–Ab interaction landscape. From partner-conditioned residue representations, CoBALT constructs a contact–energy bottleneck that factorizes pairwise interactions into a contact field localizing binding interfaces and an energy field encoding affinity-relevant local contributions. To learn from disjoint annotations, we develop an alternating multitask scheme that trains the shared bottleneck on separate BSP pairs and BAR candidate sets, enabling cross-task transfer without dual-annotated pairs. In five-fold cross-validation, CoBALT achieves epitope and paratope AUPRCs of 65.60% and 88.58%, respectively, while improving BAR Spearman correlation from 0.65 to 0.70 and Precision@1 from 50.77% to 56.99% over the strongest baselines. These gains largely persist under antigen-homology-aware holdouts, demonstrating that the contact–energy bottleneck effectively connects binding-site detection with affinity ranking.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
antibody-antigen, binding site prediction, binding affinity ranking, multitask learning

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_4VVhOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

CoBALT jointly learns antigen–antibody binding-site prediction (BSP) and binding-affinity ranking (BAR). It constructs contact and energy-like fields from partner-conditioned residue representations combining sequence and geometric information. Alternating training uses separate BSP and BAR datasets without requiring both annotations for the same pair.

Critical strengths
  1. The framework addresses a meaningful data limitation: binding-site and affinity annotations are rarely available for the same pair. The shared contact–energy representation connects residue-level interface prediction with candidate-level affinity ranking.

  2. The reported improvements are promising, particularly for epitope prediction. In Table 1, epitope AUPRC increases from 56.22% to 65.60%, BAR Spearman from 0.65 to 0.70, and P@1 from 50.77% to 56.99%.

Critical weaknesses
  1. BSP performance with predicted structures remains unverified. BSP uses structures from experimentally determined complexes, whereas BAR uses independently predicted monomer structures from ESMFold and IgFold (Appendix A.3). Performance on new pairs whose structures must be predicted therefore remains unclear. Please re-evaluate the existing BSP test sets using the same structure-prediction pipeline as BAR, keeping model checkpoints and experimental interface labels fixed. Report changes in epitope and paratope AUPRC, AUROC, and F1 relative to experimental-structure inputs. Or you can re-train this task using predicted structures for both training and testing.

  2. The ablations do not establish the stability of component contributions. Table 2 reports only fold 1 of the random split. The single-task and component controls are useful, but their gains may depend on the split; the BAR Spearman improvement over the single-task model is only 0.68 to 0.70. Please report the key ablations across all five folds, including per-fold results and mean ± standard deviation, and repeat them under the existing 75%/80% antigen-homology-aware splits with comparable training budgets and model-selection procedures.

Questions for authors
  1. Why are only antigen-homology-aware splits reported? Could you evaluate unseen antibody families and a stricter split that holds out both antigen and antibody clusters? Please specify the antibody similarity criterion and coordinate these partitions across BSP and BAR.

  2. Why is AbLWR cited but omitted from the baselines? It also uses attention among candidates and ListMLE. Please explain the omission and, where feasible, provide a comparison with matched data, splits, and candidate-list settings. If implementation or protocol constraints prevent this, please clarify them.

8 comments
AuthorsOpenReview ↗

Response for Reviewer 3 (4VVh)

We sincerely thank the reviewer for the time and care devoted to evaluating our work. We appreciate your constructive feedback, which has prompted us to take a closer look at both the evaluation and presentation of the study. In the responses below, we address each point in detail and explain the revisions made to the manuscript.

AuthorsOpenReview ↗

Response for Question1

Q1: BSP performance with predicted structures remains unverified.

We thank the reviewer for raising this important point. We agree that BSP performance with predicted structures was not verified in the original manuscript. We have now conducted this evaluation using predicted unbound structures as BSP inputs. Structural graphs were constructed using the same pipeline as BAR (Appendix A.3), and CoBALT was re-evaluated under the random five-fold split. The results are tabulated in Table R1.

Table R1. BSP performance with original and predicted unbound structures under random five-fold cross-validation.

Structural inputE-AUPRC (%) ↑E-AUROC (%) ↑E-F1 (%) ↑P-AUPRC (%) ↑P-AUROC (%) ↑P-F1 (%) ↑
Original65.60 ± 1.2492.71 ± 0.8357.00 ± 1.7588.58 ± 0.5891.32 ± 0.3880.73 ± 0.75
Predicted unbound64.93 ± 1.6392.65 ± 0.6056.07 ± 2.0888.29 ± 0.5591.13 ± 0.4580.38 ± 0.50

Results are reported as mean ± standard deviation across five folds.

All six BSP metrics remain close to the original results, with numerical decreases of less than one percentage point. Epitope and paratope AUROC are nearly unchanged, decreasing by only 0.06 and 0.19 percentage points, respectively. These results directly verify BSP performance with predicted structures and indicate that CoBALT retains robust binding-site prediction performance without bound structural inputs.

AuthorsOpenReview ↗

Response for Question2

Q2: Please report the key ablations across all five folds.

We thank the reviewer for this suggestion. We have repeated the complete ablation study on all five random folds and under the existing 75% and 80% antigen-homology-aware splits. For each task, all variants use identical data partitions and training configurations.

The five-fold results are reported as mean ± sample standard deviation in Table R2, and the homology-aware results are presented in Tabel R3. Overall, the multi-fold experiments reproduce the principal ablation trends. CoBALT improves all five BAR ranking metrics in 5/5 folds relative to Single BAR and improves epitope/paratope AUPRC and F1 in 5/5 folds relative to Single BSP. Removing BAR→C causes the largest average BAR degradation, while removing C, E, or the interface evidence also reduces BAR ranking performance and BSP AUPRC. Under both homology-aware splits, CoBALT achieves the highest BAR Spearman and epitope/paratope AUPRC and outperforms the single-task counterparts on all twelve metrics.

We note that w/o BAR→C achieves a slightly higher mean epitope AUROC on the random folds (93.57 versus 92.71), w/o interface in BAR achieves marginally higher P-F1 at 75% homology (75.33 versus 75.27), and w/o E achieves lower MAE (0.872 versus 0.948). However, these differences are small and inconsistent. CoBALT therefore remains competitive. For completeness, we report the full fold-level results for all variants and twelve metrics in the appendix.

Table R2. Ablation results under random five-fold cross-validation

Model / VariantE-AUPRC ↑E-AUROC ↑E-F1 ↑P-AUPRC ↑P-AUROC ↑P-F1 ↑MAE ↓Spearman ↑Kendall ↑P@1 ↑Hit@3 ↑PairAcc ↑
CoBALT65.60 ± 1.2492.71 ± 0.8357.00 ± 1.7588.58 ± 0.5891.32 ± 0.3880.73 ± 0.750.65 ± 0.020.70 ± 0.010.61 ± 0.0156.99 ± 1.1378.36 ± 1.3084.42 ± 0.36
CoBALT-Single54.32 ± 1.7891.91 ± 0.4445.63 ± 3.0286.78 ± 0.4490.33 ± 0.2978.33 ± 0.390.67 ± 0.020.66 ± 0.020.57 ± 0.0253.44 ± 2.2473.99 ± 3.0982.23 ± 1.19
w/o BAR→C62.63 ± 2.2993.57 ± 0.6747.52 ± 3.3787.85 ± 0.9490.98 ± 0.6580.26 ± 0.760.84 ± 0.020.54 ± 0.000.47 ± 0.0041.07 ± 1.5861.44 ± 1.9076.21 ± 0.25
w/o C in BAR55.70 ± 2.8392.06 ± 0.7540.92 ± 4.4383.96 ± 1.1588.12 ± 0.8277.27 ± 0.780.71 ± 0.030.64 ± 0.010.55 ± 0.0150.59 ± 1.1971.57 ± 1.8381.15 ± 0.59
w/o E55.61 ± 2.8792.11 ± 0.7042.37 ± 3.3183.59 ± 0.6387.76 ± 0.4676.90 ± 0.480.72 ± 0.030.64 ± 0.010.55 ± 0.0049.99 ± 1.3772.09 ± 1.3380.92 ± 0.32
w/o interface in BAR55.91 ± 1.9992.09 ± 0.4840.60 ± 2.0584.30 ± 0.8888.37 ± 0.6377.58 ± 0.580.72 ± 0.030.63 ± 0.000.54 ± 0.0049.12 ± 2.2070.79 ± 0.9080.64 ± 0.16
<br>

Table R3. Ablation results under the 75% and 80% antigen-homology-aware splits

SplitModel / VariantE-AUPRC ↑E-AUROC ↑E-F1 ↑P-AUPRC ↑P-AUROC ↑P-F1 ↑MAE ↓Spearman ↑Kendall ↑P@1 ↑Hit@3 ↑PairAcc ↑
75%CoBALT53.5089.3849.6682.3087.8975.270.9480.3680.30231.7664.8968.76
CoBALT-Single44.6587.9830.6781.2787.2874.710.9500.3040.23828.4361.8864.95
w/o BAR→C37.5484.1734.5578.3085.4172.901.0320.2730.21223.1854.2363.24
w/o C in BAR49.5987.9243.2079.2885.9973.700.9040.3290.26726.1862.5766.50
w/o E48.8788.6040.0578.5285.1372.860.8720.3400.27828.4361.4167.14
w/o interface in BAR44.3686.5140.6780.7387.2875.330.8980.3650.29530.0463.7368.72
80%CoBALT58.3590.2951.8884.6589.6477.690.7830.4990.39832.0557.7575.38
CoBALT-Single46.0587.8430.7982.9388.3276.310.8690.4670.37127.6855.1473.61
w/o BAR→C51.4089.7744.2679.7586.6274.520.8440.3960.31026.2354.7370.17
w/o C in BAR43.5985.8331.9875.7182.1569.270.8480.4460.34827.2852.5472.71
w/o E37.1985.0332.4774.0481.9168.970.8430.4460.35124.9050.3472.86
w/o interface in BAR34.2985.0732.6576.6682.9370.710.9130.4370.34424.3751.8572.11
Reviewer_jwjaOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

CoBALT jointly trains binding-site prediction (BSP) and binding-affinity ranking (BAR) on datasets that need not annotate both tasks for the same sample. A shared model produces a contact field for epitope and paratope prediction and an energy-like field that helps rank antibody candidates. On random five-fold splits, it reports epitope and paratope AUPRCs of 65.60% and 88.58%. Compared with AbRank, BAR Spearman correlation rises from 0.65 to 0.70 and Precision@1 from 50.77% to 56.99% (Table 1). The paper also evaluates antigen-homology-aware splits and includes ablations and contact-field analyses.

Critical strengths
  • Joint training improves epitope AUPRC from 54.94% to 67.34% over the single-task variant on the reported ablation fold. Blocking BAR gradients to the contact field also reduces BAR Spearman from 0.70 to 0.54. These results provide evidence that the tasks benefit from sharing the contact representation.
  • The evaluation includes random and antigen-homology-aware splits, several task-specific baselines, and an analysis of predicted contacts. CoBALT leads the reported BSP AUPRC and BAR ranking results on the random split, while its contact field aligns better with structure-derived contacts than the single-task variants.
Critical weaknesses
  • The component ablations appear only on one random-split fold. The BAR gain from joint training over the single-task variant is modest (Spearman 0.68 to 0.70), while MAE slightly worsens (0.67 to 0.68). Results across folds and on homology-aware splits are needed to show that the benefit is consistent.
  • It is unclear whether AbRank was trained on the same FLAB partitions and labels as CoBALT. BSP also uses experimental complex structures, while BAR uses predicted monomer structures. Finally, BAR metrics are computed within fixed groups of 20 candidates, so their sensitivity to group composition and their relevance to ranking a full antibody pool remain uncertain.
Questions for authors
  1. Can you report the single-task and w/o BAR→C ablations across all random folds and at least one homology-aware split?

  2. Was AbRank trained with the same FLAB training folds, labels, and validation protocol as CoBALT, including in the homology-aware evaluation?

  3. How does BSP perform with predicted, unbound structures? Do the BAR gains persist across different 20-candidate groupings and when ranking all candidates for an antigen?

6 comments
AuthorsOpenReview ↗

Response for Reviewer 2 (jwja)

We sincerely thank the reviewer for the encouraging assessment and for the time and care devoted to reviewing our work. Our point-by-point responses are provided below.

AuthorsOpenReview ↗

Response for Question1

Q1: Can you report the single-task and w/o BAR→C ablations across all random folds and at least one homology-aware split?

We thank the reviewer for this suggestion. We have now evaluated the single-task counterparts and the w/o BAR→C ablation on all five random cross-validation folds, as well as on two antigen-homology-aware splits with 75% and 80% identity thresholds. For the single-task comparison, CoBALT-Single denotes Single BAR for the BAR metrics and Single BSP for the BSP metrics. The results are reported in the table below.

Table R1. Binding-site prediction (BSP) performance of CoBALT, its single-task counterpart, and the w/o BAR→C ablation under random and antigen-homology-aware splits.

SplitModel / VariantEpitope AUPRC (%) ↑Epitope AUROC (%) ↑Epitope F1 (%) ↑Paratope AUPRC (%) ↑Paratope AUROC (%) ↑Paratope F1 (%) ↑
Random, five foldsCoBALT65.60 ± 1.2492.71 ± 0.8357.00 ± 1.7588.58 ± 0.5891.32 ± 0.3880.73 ± 0.75
CoBALT-Single54.32 ± 1.7891.91 ± 0.4445.63 ± 3.0286.78 ± 0.4490.33 ± 0.2978.33 ± 0.39
w/o BAR→C62.63 ± 2.2993.57 ± 0.6747.52 ± 3.3787.85 ± 0.9490.98 ± 0.6580.26 ± 0.76
75% homology-awareCoBALT53.5089.3849.6682.3087.8975.27
CoBALT-Single44.6587.9830.6781.2787.2874.71
w/o BAR→C37.5484.1734.5578.3085.4172.90
80% homology-awareCoBALT58.3590.2951.8884.6589.6477.69
CoBALT-Single46.0587.8430.7982.9388.3276.31
w/o BAR→C51.4089.7744.2679.7586.6274.52

Table R2. Binding-affinity prediction and ranking (BAR) performance of CoBALT, its single-task counterpart, and the w/o BAR→C ablation under random and antigen-homology-aware splits.

SplitModel / VariantMAE ↓Spearman ↑Kendall ↑P@1 (%) ↑Hit@3 (%) ↑PairAcc (%) ↑
Random, five foldsCoBALT0.65 ± 0.020.70 ± 0.010.61 ± 0.0156.99 ± 1.1378.36 ± 1.3084.42 ± 0.36
CoBALT-Single0.67 ± 0.020.66 ± 0.020.57 ± 0.0253.44 ± 2.2473.99 ± 3.0982.23 ± 1.19
w/o BAR→C0.84 ± 0.020.54 ± 0.000.47 ± 0.0041.07 ± 1.5861.44 ± 1.9076.21 ± 0.25
75% homology-awareCoBALT0.9480.3680.30231.7664.8968.76
CoBALT-Single0.9500.3040.23828.4361.8864.95
w/o BAR→C1.0320.2730.21223.1854.2363.24
80% homology-awareCoBALT0.7830.4990.39832.0557.7575.38
CoBALT-Single0.8690.4670.37127.6855.1473.61
w/o BAR→C0.8440.3960.31026.2354.7370.17

Across all three splits, CoBALT outperforms CoBALT-Single on every reported metric. On the random folds, w/o BAR→C performs worse than CoBALT on 11 of the 12 metrics. It substantially degrades BAR performance and reduces epitope AUPRC and F1. The only exception is epitope AUROC, which increases modestly from 92.71 to 93.57. CoBALT nevertheless retains a clear advantage in epitope AUPRC (65.60 versus 62.63) and F1 (57.00 versus 47.52). Under both homology-aware splits, CoBALT outperforms both CoBALT-Single and w/o BAR→C on all twelve metrics, supporting the benefits of joint training and BAR→C supervision beyond the random-fold setting.

We will update Table 2 with these five-fold results and add the full twelve-metric ablations under both antigen-homology-aware splits, together with the complete fold-level results, to the appendix.

AuthorsOpenReview ↗

Response for Question2

Q2: Was AbRank trained with the same FLAB training folds, labels, and validation protocol as CoBALT, including in the homology-aware evaluation?

Thank you for raising this important point. We confirm that AbRank and CoBALT used identical FLAB training, validation, and test partitions,as well as the same affinity labels in both random five-fold cross-validation and the antigen-homology-aware evaluations. Regarding the final checkpoint selection, the two models used different criteria. AbRank selected the checkpoint with the lowest validation loss, following the official AbRank implementation and default configuration. CoBALT selected the checkpoint that maximized the sum of validation metrics, reflecting its joint objectives of affinity ranking and binding-site prediction. All checkpoint selections were performed using validation data only, without using test labels. Thus, the data partitions, affinity labels, and validation samples were identical between the two models, while the checkpoint-selection metrics were adapted to their respective learning objectives.

Reviewer_Wwh7OpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents CoBALT, a unified framework for antibody-antigen binding-site prediction and affinity ranking. Starting from partner-conditioned residue representations, the method constructs a shared contact-energy bottleneck: a contact field supports residue-level epitope and paratope prediction, while an energy field is aggregated for complex-level affinity ranking with candidate-set context. An alternating multitask procedure trains the shared components using separate binding-site and affinity-ranking datasets without requiring joint annotations for the same complexes. The empirical study covers five-fold random splits, antigen-homology-aware splits, component ablations, and an auxiliary evaluation of predicted residue-pair contacts.

Critical strengths
  1. The contact-energy bottleneck is a coherent way to connect residue-level localization and complex-level ranking. The alternating training scheme is particularly useful because it permits cross-task learning from disjoint datasets rather than assuming scarce joint annotations.

  2. The evaluation is unusually broad for this problem setting. It includes both random and homology-aware splits, multiple metrics for both tasks, component studies, and direct inspection of the learned contact field. This breadth makes the proposed modeling direction technically interesting and potentially useful.

Critical weaknesses
  1. The empirical evidence does not yet establish that the reported gains are statistically reliable or specifically attributable to the proposed bottleneck. The significance summary appears difficult to reconcile with at least one reported mean and standard deviation, the main ablation is limited to a single random-split fold, and uncertainty is not reported for small gains under homology-aware evaluation. These limitations weaken the central claim of consistent improvement.

  2. Several protocol choices leave the generalization and mechanistic claims ambiguous. Cross-task overlap under random splitting is not quantified, affinity scores can depend on candidate-group composition, and the contact metric may reward marginal interface localization without demonstrating specific residue-pair matching. Because these issues affect all three headline claims, I currently lean toward rejection despite finding the framework promising.

Questions for authors
  1. Please provide the five fold-level values behind each Table 1 comparison, the exact corrected p values, the testing unit, the correction procedure, and the statistical test code. Please also report test antigen-cluster counts and cluster-level confidence intervals for the homology splits, and repeat the principal Table 2 ablations across folds with matched data and model capacity.

  2. What exact antigen, antibody, and antigen-antibody-pair overlap exists between training data for one task and test data for the other under the random split? How are the HIV-1 and HEWL examples assigned? Would the conclusions change after removing every cross-task overlap or close homolog?

  3. How are size-20 candidate groups constructed, how are remainder candidates handled, and is grouping fully independent of affinity labels? Please report sensitivity to repeated regrouping or score each complex against several alternative candidate contexts. For the contact field, please also compare against a product-of-marginals baseline and report performance when both residues are restricted to known interface residues.

9 comments
AuthorsOpenReview ↗

Response for Reviewer 1 (Wwh7)

Thanks a lot for taking the time to carefully review our manuscript and for your thoughtful and constructive comments. Your suggestions have helped us identify areas for improvement and have been invaluable in strengthening the clarity and completeness of our work. We have carefully considered each of your comments and revised the manuscript accordingly. Below, we provide our detailed responses and describe the corresponding changes.

AuthorsOpenReview ↗

Response for Question1

Q1: Please provide the five fold-level values behind each Table 1 comparison, the exact corrected p values, the testing unit, the correction procedure, and the statistical test code.

Thank you for raising this important point. We have now clarified the statistical testing procedure, provided all fold-level results and analysis code, and corrected an error in the EpiFormer-related analysis.

Statistical test and multiple-testing correction. For each comparison in Table 1, we used fold-level metric values as the units of analysis. Specifically, for each evaluation metric, we paired the performance of our method with that of the corresponding baseline on the same five folds, yielding five paired observations. We then conducted a two-sided paired t-test using scipy.stats.ttest_rel. Finally, we applied the Benjamini–Hochberg (BH) procedure to adjust p-values across all baseline comparisons separately for each metric.

Fold-level results and reproducibility. The complete five-fold results underlying every Table 1 comparison, together with the exact BH-corrected p-values, are provided in Table1.xlsx. These corrected p-values have also been added to Appendix Table A2 and highlighted in red. The full statistical analysis code is available in ttest.py. Both files can be accessed through an anonymous public repository (https://anonymous.4open.science/r/CoBALT-341B).

Correction to the EpiFormer analysis. During this revision, we found that the original EpiFormer-related t-tests had mistakenly used fold-level results from an earlier, incomplete reproduction rather than the final reproduced version. We have therefore recomputed the corresponding paired t-tests and BH-corrected p-values using the correct five-fold results and updated the manuscript accordingly. Only the statistical conclusion for epitope AUROC changes: after correction, CoBALT’s advantage over EpiFormer on this metric is no longer statistically significant. The conclusions for all other metrics remain unchanged. All changes are highlighted in red. We apologize for this oversight.

AuthorsOpenReview ↗

Response for Question2

Q2: Please also report test antigen-cluster counts and cluster-level confidence intervals for the homology splits.

Thank you for this suggestion. We agree that point estimates may not fully reflect uncertainty here. We therefore added test antigen-cluster counts and cluster-level 95% confidence intervals (CIs) for both splits and assessed performance differences against the baselines.

The numbers of test antigen clusters used in each evaluation are tabulated in Table R1.

Table R1. Total antigen clusters and test-set statistics under antigen-homology-aware splits.

SplitTotal clustersBSPBAR
75%1,56022 / 1,57230 / 10,440
80%1,94910 / 1,41620 / 10,490

BSP and BAR entries denote the number of test clusters / test pairs.

Cluster-level CIs. We estimated 95% CIs using 2,000 antigen-cluster bootstrap replicates for each homology split (Table R2) and added them to the revised Figure 2 and Figure A1.

Table R2. Performance of CoBALT and comparison baseline methods with 95% CIs under antigen-homology-aware splits.

Split (BSP)MethodE-AUPRC (%) ↑E-AUROC (%) ↑E-F1 (%) ↑P-AUPRC (%) ↑P-AUROC (%) ↑P-F1 (%) ↑
75%EpiFormer / EpiScan36.55 [22.38, 46.49]84.76 [73.56, 89.71]34.67 [26.60, 38.94]75.86 [65.34, 87.01]85.54 [82.41, 88.54]73.03 [63.50, 82.82]
75%CoBALT53.50 [22.91, 72.19]89.38 [72.56, 95.06]49.66 [28.97, 63.10]82.30 [69.35, 89.91]87.89 [84.32, 89.58]75.27 [65.81, 83.43]
80%EpiFormer / EpiScan24.90 [14.98, 34.32]80.47 [70.36, 85.62]17.64 [9.23, 27.34]74.01 [61.82, 89.66]84.60 [80.72, 89.33]70.87 [59.27, 84.92]
80%CoBALT58.35 [22.79, 83.45]90.29 [78.03, 97.35]51.88 [29.66, 72.55]84.65 [66.12, 93.68]89.64 [83.21, 94.25]77.69 [64.80, 91.09]
Split (BAR)MethodMAE ↓Spearman ↑Kendall ↑P@1 (%) ↑Hit@3 (%) ↑PairAcc (%) ↑
75%AbRank0.90 [0.79, 1.14]0.36 [0.23, 0.47]0.29 [0.17, 0.38]31.44 [19.17, 35.59]61.18 [42.09, 68.57]67.74 [60.71, 73.56]
75%CoBALT0.95 [0.79, 1.34]0.37 [0.24, 0.52]0.30 [0.19, 0.41]31.76 [20.48, 35.45]64.89 [42.35, 73.07]68.76 [62.03, 76.13]
80%AbRank1.32 [0.75, 2.75]0.41 [0.31, 0.47]0.32 [0.24, 0.37]25.30 [21.41, 30.57]50.62 [43.58, 57.98]70.52 [65.31, 74.02]
80%CoBALT0.78 [0.75, 0.83]0.50 [0.38, 0.56]0.40 [0.31, 0.45]32.05 [25.81, 37.42]57.75 [48.22, 65.45]75.38 [69.91, 78.64]

Values in brackets indicate 95% CIs.

Uncertainty of the performance gains. To further assess the reliability of the improvements, we computed 95% CIs for the performance differences between CoBALT and the corresponding baselines using 2,000 paired antigen-cluster bootstrap replicates. The complete results are summarized in Table R3.

Table R3. 95% CIs for performance differences between CoBALT and comparison baseline methods antigen-homology-aware splits.

Split (BSP)E-AUPRC ↑E-AUROC ↑E-F1 ↑P-AUPRC ↑P-AUROC ↑P-F1 ↑
75%[−0.0035, +0.2603][−0.0113, +0.0597][+0.0122, +0.2441][+0.0219, +0.0714][+0.0064, +0.0268][+0.0012, +0.0292]
80%[+0.0540, +0.4936][+0.0282, +0.1293][+0.1031, +0.5057][+0.0368, +0.1131][+0.0238, +0.0604][+0.0331, +0.0730]
Split (BAR)MAE ↓Spearman ↑Kendall ↑P@1 ↑Hit@3 ↑PairAcc ↑
75%[−0.2638, +0.0669][−0.0288, +0.0592][−0.0202, +0.0581][−0.0192, +0.0311][+0.0070, +0.0652][−0.0031, +0.0350]
80%[−0.0175, +1.9842][+0.0362, +0.1378][+0.0341, +0.1211][+0.0049, +0.1041][+0.0224, +0.1170][+0.0191, +0.0738]

BSP: CoBALT vs. EpiFormer (epitope) / EpiScan (paratope). BAR: CoBALT vs. AbRank.

Overall, 16 of the 24 nominal 95% difference CIs lie entirely above zero. Under the 80% split, all six BSP comparisons and all five BAR ranking-metric comparisons have intervals above zero. Under the more stringent 75% split, positive intervals are retained for four BSP metrics and BAR Hit@3. The remaining intervals include zero, leaving uncertainty about the corresponding performance differences.

The uncertainty in several BAR comparisons under the more stringent 75% split highlights the importance of broader antigen coverage, consistent with our existing discussion of expanding dataset scale and diversity. Broader coverage of antigen families in training may improve generalization to dissimilar antigens, while evaluation on more independent held-out antigen clusters would allow more precise assessment of these gains. We have clarified this connection in the revised discussion.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.