acceptodds
Under review as a conference paper at ICLR 2027

ESFold: Adapting Biomolecular Structure Models with Experimental Signals

Abstract

Protein–RNA interactions are fundamental to gene regulation and RNA function, yet their structural modeling is limited by scarce experimentally resolved complexes. In contrast, enhanced crosslinking and immunoprecipitation (eCLIP) assays provide abundant interaction evidence but lack atomic resolution. Here we introduce ESFold, a two-stage post-training framework built on Protenix that integrates eCLIP signals into protein–RNA structure modeling. In the first stage, supervised post-training combines structural and eCLIP data to learn binding-aware representations. In the second stage, we adapt Diffusion Negative-aware Fine-Tuning (DiffusionNFT) to molecular structure generation, using eCLIP agreement and structural plausibility as feedback to optimize the distribution of generated conformations. We construct 265,070 candidate protein–RNA pairs spanning 53,014 RNA windows and 84 proteins. On a protein–RNA interface structure benchmark, ESFold improves the prediction success rate from 61.43% with Protenix to 74.29% and reduces interface RMSD from 7.61 Å to 5.02 Å. On a held-out test set of 190 experimental binding conditions, peak F1 averaged across RNA-binding proteins (RBPs) improves from 0.332 to 0.459. These gains are achieved while broadly maintaining Protenix’s performance on structural prediction tasks beyond protein–RNA complexes. Together, the findings highlight experimental-signal post-training as an approach to integrating functional genomics into three-dimensional biomolecular modeling.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Biomolecular structure prediction, Protein–RNA interactions, Experimental supervision, Diffusion models, Reinforcement learning, Post-training

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_S5CcOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

I think this paper presents an interesting approach to improving protein-RNA structure prediction using experimental binding signals. Compared with existing structure prediction models that mainly rely on structural data, this paper tries to use eCLIP data, which provides information about protein binding regions on RNA but does not contain atomic-level structural information. More detailed into the method, the authors propose ESFold, a two-stage post-training framework based on Protenix. In the first stage, the model learns binding-aware representations by aligning predicted RNA contact probabilities with experimental eCLIP signals, together with standard structural supervision from PDB. In the second stage, the authors adapt DiffusionNFT to optimize generated structures using experimental agreement and geometric constraints as rewards. The model does not require experimental signals during inference.

The experiments are conducted on the protein-RNA subset of FoldBench and a held-out eCLIP evaluation set. The method improves interface prediction success from 61.43% to 74.29% and experimental peak F1 from 0.332 to 0.459 compared with Protenix.

In all, I think the paper investigates a meaningful problem and reports some promising results. Using experimental data to improve pretrained structure prediction models is an interesting direction, especially when such data are much more available than experimentally resolved structures. However, after looking more carefully into the methods and experiments, I have several concerns regarding the evaluation metrics, experimental evidence, and the overall presentation of the paper. Some of these concerns are directly related to whether the reported results can fully support the main claims.

Critical strengths
  1. I think the paper investigates an important problem in biomolecular structure prediction. Protein-RNA complex structures remain relatively limited, while experimental interaction data such as eCLIP provide additional information that could potentially improve structure prediction. Exploring how to incorporate such experimental signals into pretrained structure models is a meaningful research direction.

  2. The proposed two-stage framework is reasonable, and the reported improvements are encouraging. The authors combine binding-aware representation learning with reward-based structure optimization and evaluate both structural accuracy and agreement with experimental signals. I think this is a useful direction, particularly because experimental binding information does not directly provide atomic-level structural supervision.

Critical weaknesses
  1. I am not fully convinced that the current experiments sufficiently support the main contributions. First, the definition of Peak F1 in Appendix C allows predicted regions to match experimental peaks through any overlap, without one-to-one matching. Together with percentile-based thresholding, this can potentially give high scores to predictions with poor localization. Since Peak F1 is one of the main evaluation metrics, I think a more careful evaluation is necessary. Moreover, Stage 1 combines PDB structural supervision with eCLIP signals, but the paper does not include a PDB-only fine-tuning baseline or shuffled eCLIP controls. Therefore, it remains unclear how much of the improvement can be attributed specifically to meaningful experimental supervision.

  2. I have significant concerns about the experimental analysis and overall presentation of the paper. Although Stage 2 introduces several reward components, the paper does not provide sufficient ablations to explain which components contribute to the final improvement. The structural evaluation set is also relatively small, making additional statistical analysis important. More importantly, I think the presentation quality is considerably below what I would expect for an ICLR submission. The main text does not include detailed quantitative result tables, with most important comparisons placed in the appendix. The figures are not vectorized, which affects readability, and some training details are not clearly reported. Given the complexity of the proposed framework, I think the paper needs a much clearer experimental presentation and more thorough analysis to support its conclusions.

Questions for authors
  1. Could the authors provide additional evaluation of Peak F1 using stricter region matching or position-level metrics? I think it would also be helpful to report the fraction of predicted positive positions to better understand the reliability of the current metric.

  2. Could the authors compare Stage 1 with PDB-only continued fine-tuning under a similar training budget? Also, would the improvements remain if the eCLIP signals were shuffled or replaced with mismatched profiles?

  3. Could the authors provide more detailed ablations of the Stage 2 reward components? I am also interested in whether the experimental reward correlates with structural accuracy, such as DockQ.

7 comments
AuthorsOpenReview ↗

Stricter localization metrics and positive-position coverage

We agree that overlap-based Peak F1 alone does not establish precise localization. We added nucleotide-level metrics, maximum-cardinality one-to-one peak matching, prediction coverage, and a threshold sweep. The comparison uses matched confidence-selected Top-1 predictions on 190 conditions from 82 RBPs.

MetricProtenix v1.0ESFoldPaired difference [95% CI]
Position precision0.1350.1640.030 [0.015, 0.045]
Position recall0.4750.6290.154 [0.087, 0.220]
Position F10.1990.2480.049 [0.027, 0.071]
Average precision0.2110.2960.086 [0.049, 0.127]
Trapezoidal PR-AUC0.1920.2730.080 [0.044, 0.121]
MetricProtenix v1.0ESFoldPaired difference [95% CI]
Legacy Peak F10.3320.4590.126 [0.089, 0.165]
One-to-one IoU ≥0.10 F10.2730.3710.098 [0.058, 0.142]
One-to-one IoU ≥0.25 F10.1980.2190.021 [-0.008, 0.053]
One-to-one IoU ≥0.50 F10.0850.072-0.013 [-0.037, 0.011]
Peak center ≤2 nt F10.1450.1680.023 [-0.011, 0.060]
Peak center ≤5 nt F10.2860.3360.049 [0.011, 0.087]

ESFold improves position-level metrics and one-to-one F1 at IoU ≥0.10. Gains are smaller at IoU ≥0.25 and do not extend to IoU ≥0.50, indicating stronger regional localization rather than uniformly sharper peak boundaries.

MetricProtenix v1.0ESFoldPaired difference [95% CI]
Positive-position fraction0.4030.4030.000 [0.000, 0.000]
Predicted regions / sample7.1635.705-1.457 [-1.968, -0.931]
Mean region width (nt)6.99811.0494.051 [2.674, 5.523]

Both models predict the same positive-position fraction. ESFold predicts fewer, wider regions, which motivates the additional one-to-one matching and threshold-sensitivity analyses below.

Prediction maskΔ position F1 [CI]Δ legacy F1 [CI]Δ IoU ≥0.25 F1 [CI]
Top 20%0.077 [0.044, 0.112]0.138 [0.093, 0.182]0.102 [0.064, 0.145]
Top 30%0.063 [0.034, 0.093]0.146 [0.104, 0.189]0.079 [0.046, 0.115]
Top 40%0.049 [0.027, 0.071]0.126 [0.089, 0.165]0.021 [-0.008, 0.053]
Top 50%0.040 [0.021, 0.059]0.120 [0.076, 0.164]0.003 [-0.026, 0.033]

We additionally evaluate an absolute contact definition: minimum RNA–protein heavy-atom distance ≤8 Å, without percentile thresholding for the binary mask. Binary contact metrics and RNA-position PR-AUC are reported below. PR-AUC is the trapezoidal area under the RNA-position-level precision–recall curve; it is distinct from average precision (AP).

MetricProtenix v1.0ESFoldPaired difference [95% CI]
Position F10.1910.2550.064 [0.038, 0.090]
Position PR-AUC (trapezoidal)0.1790.2410.062 [0.038, 0.088]
One-to-one IoU ≥0.25 F10.1750.2760.101 [0.063, 0.141]
Contact fraction0.4550.361-0.094 [-0.116, -0.070]

Together, these analyses show that the improvement persists in position-level evaluation, across prediction thresholds, and under an absolute distance-based contact definition. All signal metrics are averaged within each RBP and then across the 82 RBPs; paired intervals are 95% bootstrap CIs.

Manuscript revision. Section 4.4 and Appendix M, Tables 24–28, now give strict matching, coverage, threshold sensitivity, and distance-based results. Appendix Q defines AP and trapezoidal PR-AUC, including the ranking score and threshold direction.

AuthorsOpenReview ↗

PDB-only fine-tuning and signal-control experiments

Under the same training budget, true-eCLIP Stage-1 supervision achieves higher structural accuracy and signal agreement than PDB-only fine-tuning and both signal controls.

  • Shuffled-eCLIP: randomly permute the measured signal values across valid RNA positions, preserving each example’s signal-value distribution but disrupting positional correspondence.
  • Random-signal control (labeled “mismatched-eCLIP” in the table): replace the measured profile with a random signal. This tests uninformative supervision rather than a real profile from another protein–RNA pair.

True-eCLIP and both signal controls retain the same PDB structural supervision; only the eCLIP labels differ. PDB-only excludes eCLIP supervision, and Protenix v1.0 is the pretrained baseline.

Stage-1 supervisionDockQ successiRMSD (Å)lRMSD (Å)lDDT
Protenix v1.061.43%7.6118.660.74
PDB-only64.29%6.4616.860.76
Shuffled-eCLIP62.86%7.0617.410.76
Mismatched-eCLIP61.43%7.3417.820.75
True-eCLIP (ESFold_rep)70.00%5.3713.240.78

Structural results use the same 70 interfaces from 49 complexes. Signal-agreement results below use the 190-condition, 82-RBP eCLIPBench.

Stage-1 supervisionPeak F1 [95% CI]JaccardSpearman ρEnrichment
Protenix v1.00.332 [0.305, 0.361]0.1260.0511.732
PDB-only0.339 [0.297, 0.379]0.1190.0351.797
Shuffled-eCLIP0.338 [0.304, 0.371]0.1040.0561.763
Mismatched-eCLIP0.315 [0.281, 0.344]0.1130.0471.772
True-eCLIP (ESFold_rep)0.417 [0.388, 0.445]0.1550.1443.044

Manuscript revision. Section 4.5 and Table 2 present the supervision controls; Appendix N, Tables 29–30, provides the remaining metrics.

AuthorsOpenReview ↗

Stage-2 ablations and reward–DockQ association

The ablations indicate that reference loss provides the strongest stabilization, geometry helps preserve physical plausibility, and signal reward gives a modest Peak F1 gain in the evaluated run.

Ablation settingMean Peak F1 ↑Mean physical penalty ↓
All rewards + reference loss0.3530.086
No signal reward0.3450.089
No geometry0.3550.089
No reference0.2730.171
No reference + no geometry0.1680.605

Values average nine validation checkpoints from steps 8,000–10,000, sampled every 250 steps, from one run per setting. Geometry terms are ablated as a group.

For reward–DockQ association, the available diagnostic uses native-derived pseudo-eCLIP profiles, not real experimental profiles paired with native structures:

Native-derived diagnostic, 49 complexesPearson r [95% CI]Spearman ρ [95% CI]
Enhanced eCLIP reward vs. mean interface DockQ0.473 [0.239, 0.662]0.394 [0.121, 0.615]
Peak F1 component vs. DockQ0.282 [−0.050, 0.553]0.202 [−0.105, 0.463]
Profile-rank component vs. DockQ0.631 [0.465, 0.766]0.655 [0.504, 0.764]

Interface scores are averaged within each of the 49 complexes, with 2,000 bootstrap resamples. The enhanced reward shows a positive association with DockQ in this native-derived diagnostic; its relationship to DockQ using real experimental eCLIP profiles remains untested.

Manuscript revision. Section 4.5 and Table 3 report the five ablations. Appendix O defines the ablations and reports the native-derived reward–DockQ diagnostic in Table 31; the paper includes the trajectories in Figure 8.

Reviewer_FeopOpenReview ↗
Rating
4
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper presents ESFold, a post-training framework for diffusion-based protein complex prediction models aligned with experimental signals such as single-nucleotide-level protein-RBP contact signals obtained from eCLIP. The first stage aligns the model's Pairformer output to the experimental binding signal. The second stage uses DiffusionNFT on the coordinate space targeting alignment between predicted and eCLIP contact signals and geometric validity. ESFold shows superior interface structural accuracy and eCLIP cotact retrieval compared to the baseline methods.

Critical strengths
  1. The proposed method adequately addresses the main challenges of protein-RNA structure predictions. Specifically, the incorporation of eCLIP signals moves beyond conventional static structural data-driven probabilistic modeling, and demonstrates substantial improvements in predictive accuracy and generalizability. The work also performs adequate ablation studies (rep and nft baselines) justifying the proposed two-stage framework.
Critical weaknesses

None.

Questions for authors
  • Since the core evaluation is based on eCLIP signals on RNA sequences, how much of the model's performance comes from protein-specific interactions, and how much from RNA sequence-dependent interaction hotspots? In the meantime, does the train/test split sufficiently address potential leakage from sequence similarity?
  • How much do the geometry rewards contribute to the model performance, considering that the pre-trained model should possess the capability to produce high-quality interfaces?
  • How much does the adapted contact map in stage 1 agree with the final structure output in stage 2?
  • Since the reported results are on the highest-confidence predictions, how to disentangle the benefit from better structure generation and that from better ranking/selection? For example, what is the average performance on all 25 candidates for each model?
6 comments
AuthorsOpenReview ↗

Data separation and the contribution of localized experimental signals

For PDB data, we follow an AlphaFold 3-style temporal/homology split; for eCLIP, we use RBPNet’s source partitions and hold out its test partition. The signal controls show that correct signal positions matter, although they do not separately quantify protein-specific recognition and RNA-sequence effects.

1. Data partitioning and sequence-overlap analysis.

For PDB data, we use structure-release dates and sequence-homology filtering to reduce training–evaluation overlap. For eCLIP, the 25,529 training pairs combine RBPNet’s source training partition (21,792 pairs) and validation partition (3,737 pairs), while the source test partition remains held out.

We additionally compared all 190 eCLIPBench RNA inputs against the actual training RNA sequences:

Minimum sequence identityTest RNAs with a detected training match / 190Fraction
10%1 / 1900.53%
20%1 / 1900.53%
30%1 / 1900.53%
40%1 / 1900.53%
50%1 / 1900.53%
60%1 / 1900.53%
70%1 / 1900.53%
80%1 / 1900.53%
90%0 / 1900.00%

The MMseqs2 search requires ≥90% coverage of both sequences. We found no shared RNA IDs, exact sequence matches, or detected matches at ≥90% identity; short shared motifs are not excluded by this analysis.

2. Protein-specific interactions versus RNA-sequence effects.

Correctly localized eCLIP supervision improves Stage-1 Peak F1 relative to both controls:

Stage-1 supervisionPeak F1
PDB-only0.339
Within-RNA shuffled eCLIP0.338
True eCLIP0.417

Because within-RNA shuffling preserves protein identity while disrupting signal positions, this comparison supports the value of correctly localized experimental supervision. Separating protein-specific recognition from RNA-sequence effects would require an additional protein-swap or matched cross-RBP control.

Manuscript revision. Section 4.1 and Appendix J describe source partition use and the RNA sequence comparison (Table 19); Appendix Q gives search parameters and interpretation limits.

AuthorsOpenReview ↗

Contribution of geometry rewards

Geometry rewards mainly help maintain physical plausibility during post-training. Their contribution is modest with reference loss retained and larger when reference loss is removed; they do not consistently improve Peak F1.

Ablation settingMean Peak F1 ↑Mean Physical penalty ↓
All rewards + reference loss0.3530.086
No signal reward0.3450.089
No geometry0.3550.089
No reference0.2730.171
No reference + no geometry0.1680.605

Without reference loss, adding geometry reduces the physical penalty from 0.605 to 0.171. With reference loss retained, its effect is smaller and Peak F1 does not improve. Values average nine validation checkpoints from steps 8,000–10,000, sampled every 250 steps, from one training run per setting.

Manuscript revision. Section 4.5, Table 3, and Appendix O report these ablations and the training trajectories.

AuthorsOpenReview ↗

Consistency between Stage-1 contacts and Stage-2 structures

Using matched chain mappings, representative atoms and an 8 Å contact definition, Stage-1 predictions and Stage-2 contacts show stronger pair-level agreement on FoldBench than on eCLIPBench.

DatasetEvaluated interfaces/conditionsPair AUPRCPair top-k IoURNA-profile Spearman
FoldBench700.6460.4900.383
eCLIPBench1900.1330.0940.358

Pair AUPRC and top-k IoU use the 64/70 FoldBench interfaces and 188/190 eCLIPBench conditions with nonempty Stage-2 contacts. With equal numbers of contacts selected from each stage, pooled contact retention is 60.0% on FoldBench and 17.1% on eCLIPBench.

Consistency vs. outcomeSpearman ρ95% cluster-bootstrap CI
FoldBench pair AUPRC vs. DockQ0.509[0.234, 0.705]
FoldBench pair IoU vs. DockQ0.461[0.185, 0.665]
eCLIPBench pair AUPRC vs. Peak F10.092[-0.048, 0.238]
eCLIPBench pair top-k IoU vs. Peak F10.111[-0.040, 0.259]
eCLIPBench RNA-profile Spearman vs. Peak F10.251[0.115, 0.381]

On eCLIPBench, RNA top-40% IoU is 0.408, indicating stronger consistency at the RNA-region level than at the exact residue-pair level. Pair-level consistency correlates with DockQ on FoldBench, while its association with eCLIPBench Peak F1 is weak. These are descriptive associations for the selected predictions.

Manuscript revision. Appendix K, Tables 20–21, now reports contact consistency, valid counts, retention, and outcome associations.

Reviewer_wR71OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

In this submission, the authors are focusing on a promising problem in the field of biology called "Protein-RNA binding interaction". Specifically, a type of experimental data (i.e., eCLIP assay, widely used strategy to identify the interactions among Proteins and RNAs) was applied to conduct post-training (fine-tuning) towards Protenix, with the aim of improving the performance specifically in the prediction task of Protein-RNA complexes. In general, this is a clean submission with informative data. However, there still remain several concerns which need to be addressed (see blow).

Critical strengths
  1. Introduce a strategy by integrating experimental data (i.e., eCLIP assay) to post-train a public folding pipeline (i.e. Protenix). Specifically, the authors are focusing on an important problem in the field of biology called 'Protein-RNA binding interactions'.
  2. Construct a dataset, including 265,070 candidate Protein–RNA pairs spanning 53,014 RNA windows and 84 proteins, which might be used for further efforts to conduct precision RNA design.
Critical weaknesses
  1. [Major] DockQ: I assume it refers to the standard tool for evaluating the quality of interfaces in 3D structures (i.e., Bioinformatics, 2024). If the answer is yes, please disclose the distribution of DockQ towards each predicted structures from each pipeline (i.e., ESFold and baselines). Specifically, according to the official repo of Dock, please include: (1) [0.00, 0.23): Incorrect (2) [0.23, 0.49): Acceptable quality (3) [0.49, 0.80): Medium quality (4) [0.80, 1.0]: High quality. Besides the interface success rate in general, the overall quality of the 'success' predictions is also essential.
  2. [Major] Additional Filtering (i.e., Lines 884-885): Why did you set the cutoff towards the length of amino acids as 600 AA? Have the authors ever tested the performance of ESFold on those 'longer' protein-RNA complexes? If the answer is yes, please consider disclosing some more results. Additionally, I'm also curious about the generalizability of ESFold across different types of human cell lines, except for K562 and HepG2 (i.e., line 870). In other words, have the authors found any other available eCLIP data then assessed the performance of ESFold?
  3. [Minor] Case Studies (i.e., Figures 4, 7 and 8): How are they selected? Would the authors please disclose some more details (i.e., workflow of selecting representatives)?
Questions for authors

Please refer back to "Critical Weaknesses".

7 comments
AuthorsOpenReview ↗

DockQ quality tiers and score distributions

Yes, we use standard DockQ, computed with DockQv2. The tables report confidence-selected Top-1 predictions, with success defined as DockQ ≥0.23. The tiers are Incorrect [0, 0.23), Acceptable [0.23, 0.49), Medium [0.49, 0.80), and High [0.80, 1]. ESFold increases high-quality interfaces from 4/70 to 22/70 relative to Protenix v1.0.

ModelInterfaces / assembliesIncorrectAcceptableMediumHigh
AlphaFold 369 / 4826 (37.7%)16 (23.2%)21 (30.4%)6 (8.7%)
Boltz-158 / 4325 (43.1%)19 (32.8%)14 (24.1%)0 (0.0%)
Chai-155 / 4227 (49.1%)21 (38.2%)3 (5.5%)4 (7.3%)
HelixFold 358 / 4130 (51.7%)16 (27.6%)10 (17.2%)2 (3.4%)
Protenix v1.070 / 4927 (38.6%)20 (28.6%)19 (27.1%)4 (5.7%)
ESFold_rep70 / 4921 (30.0%)12 (17.1%)23 (32.9%)14 (20.0%)
ESFold_nft70 / 4922 (31.4%)17 (24.3%)24 (34.3%)7 (10.0%)
ESFold70 / 4918 (25.7%)7 (10.0%)23 (32.9%)22 (31.4%)
ModelMean DockQMedian [Q1, Q3]Success
AlphaFold 30.3690.363 [0.062, 0.570]62.32%
Boltz-10.2800.281 [0.038, 0.428]56.90%
Chai-10.2690.240 [0.045, 0.389]50.91%
HelixFold 30.2630.223 [0.021, 0.413]48.28%
Protenix v1.00.3500.357 [0.037, 0.567]61.43%
ESFold_rep0.4850.523 [0.187, 0.778]70.00%
ESFold_nft0.4080.449 [0.058, 0.627]68.57%
ESFold0.5670.710 [0.232, 0.835]74.29%

External-baseline results are taken from FoldBench; percentages use the interface counts shown in the table. Protenix v1.0 and all ESFold variants are evaluated on the same 70 interfaces.

Manuscript revision. Appendix G, Tables 13–14, now report DockQ tiers, denominators, and score distributions; Table 1 summarizes the main structural results.

AuthorsOpenReview ↗

Training length cutoff and generalization to longer proteins

The 600-aa cutoff reflects our training memory budget, not an inference limit. With approximately 800 tokens available, longer proteins would require cropping; because eCLIP labels RNA regions but does not locate the protein-side interface, such cropping could remove binding residues. The cutoff retained 84 of 150 proteins in the source pool.

We added two long-protein evaluations: 100 conditions from 50 RBPs with 601–1000-aa proteins, and 16 conditions from 16 RBPs with proteins longer than 1000 aa. Samples were selected independently of model performance. Top-1 uses each model’s native confidence ranking; oracle scores are per-metric maxima over the sampled candidates using evaluation labels.

601–1000-aa proteins

ModelConditions / RBPsTop-1 Peak F1 [95% CI]Top-1 JaccardOracle best Peak F1Oracle best Jaccard
Protenix v1.0100 / 500.340 [0.313, 0.366]0.1140.5400.225
AlphaFold 3100 / 500.363 [0.327, 0.394]0.1220.6530.233
Boltz-1100 / 500.334 [0.312, 0.357]0.1170.5210.208
Chai-1100 / 500.342 [0.315, 0.369]0.1180.5930.232
ESFold_rep100 / 500.383 [0.355, 0.410]0.1370.6360.245
ESFold_nft100 / 500.356 [0.328, 0.383]0.1330.5640.226
ESFold100 / 500.391 [0.368, 0.415]0.1450.6580.248

>1000-aa proteins

ModelConditions / RBPsTop-1 Peak F1 [95% CI]Top-1 JaccardOracle best Peak F1Oracle best Jaccard
Protenix v1.016 / 160.280 [0.204, 0.351]0.1190.5450.227
AlphaFold 316 / 160.332 [0.279, 0.394]0.1240.5830.218
Boltz-116 / 160.326 [0.263, 0.398]0.1190.5530.214
Chai-116 / 160.316 [0.252, 0.387]0.1120.5360.209
ESFold_rep16 / 160.355 [0.282, 0.419]0.1300.5690.229
ESFold_nft16 / 160.296 [0.233, 0.357]0.1240.5590.224
ESFold16 / 160.360 [0.303, 0.422]0.1360.6110.238

These sets evaluate experimental-signal agreement. We additionally assessed structural accuracy using FoldBench complexes with reference structures.

Complementary structural evaluation on FoldBench.

Protein lengthInterfacesProtenix Top-1 DockQESFold Top-1 DockQProtenix / ESFold Top-1 successProtenix Oracle best DockQESFold Oracle best DockQ
≤300 aa380.3810.65573.7% / 89.5%0.4790.744
301–600 aa160.2660.30643.8% / 37.5%0.3720.481
>600 aa160.3600.61650.0% / 75.0%0.4480.717

For the 16 FoldBench interfaces with proteins longer than 600 aa, ESFold improves mean DockQ from 0.360 to 0.616 and success rate from 50.0% to 75.0%, supporting generalization beyond the training cutoff in this subset.

Manuscript revision. Appendix B explains the length filter; Appendix H, Tables 15–17, reports both long-protein evaluations and the FoldBench length strata. Dataset construction is described in Appendix Q.

AuthorsOpenReview ↗

Evaluation beyond K562 and HepG2

Yes. We added an external-cell-type benchmark containing 50 conditions from five RBPs: FUS and U2AF1 in HEK293, DDX6 in HEK293T, and G3BP1 and TIA1 in motor neurons, with ten conditions per RBP. All seven methods cover all 50 conditions. The labels are reproducibly enriched genomic windows in 300-nt contexts, so this benchmark measures coarse regional localization and is evaluated separately from the original eCLIPBench.

ModelConditions / RBPsTop-1 Peak F1Top-1 JaccardOracle best Peak F1Oracle best Jaccard
Protenix v1.050 / 50.4760.2060.6750.361
AlphaFold 350 / 50.4950.2320.8410.478
Boltz-150 / 50.4900.2320.6550.337
Chai-150 / 50.4730.1860.6980.380
ESFold_rep50 / 50.5230.2480.7780.420
ESFold_nft50 / 50.4820.2270.6850.372
ESFold50 / 50.6160.3160.8780.497

Scores are averaged within each RBP and then equally across RBPs. Top-1 uses native confidence ranking; oracle scores maximize each metric over the sampled candidates using evaluation labels. ESFold achieves the highest scores in this five-RBP benchmark, supporting transfer of regional signal agreement beyond K562 and HepG2.

Manuscript revision. Appendix I (Table 18) reports the unified 50-condition benchmark; Appendix Q (Table 34) gives source accessions and regional-label construction.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.