acceptodds
Under review as a conference paper at ICLR 2027

Beyond Structure Generation: Co-Folding Representations for Binding Affinity and ADMET Prediction

Abstract

Binding affinity and enzyme-dependent metabolism are distinct molecular outcomes that both involve protein–ligand recognition. We introduce PairSense, a framework for predicting binding affinity and CYP substrate status from protein-conditioned co-folding representations without explicit structure generation. PairSense combines Boltz-2’s Pairformer embeddings with predicted pairwise distance distributions, retaining probabilistic geometric information while replacing distances derived from diffusion-generated coordinates. With the co-folding trunk frozen, we fine-tune only the affinity module on public data curated to remove benchmark overlap. PairSense improves over the original Boltz-2 by approximately 0.09 in Pearson correlation on the four-target FEP+ benchmark and 0.16 in Spearman correlation on OpenBind, with additional gains on CASP16 and OpenFE876. Keeping the affinity-adapted interaction representation fixed, we jointly fine-tune its native binary readout with molecular fingerprints on three cytochrome P450 substrate tasks. The joint model achieves higher mean scores than the published TDC leaderboard references. Controlled comparisons show that additional binding supervision improves CYP prediction and that protein context adds value beyond ligand features and task identity. These findings support co-folding representations as a reusable basis for predicting binding affinity and protein-mediated molecular properties beyond structure generation.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
co-folding representations, protein-ligand interactions, molecular property prediction, binding affinity prediction, CYP substrate prediction, multi-task learning, ADMET prediction

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_bUN7OpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
2(+1)
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes PairSense, a framework that uses pair representations from the Boltz-2 trunk for binding affinity and CYP substrate. In addition, the distrogram is used both for pocket cropping and as a probability-weighted input to the affinity module. With the Boltz-2 trunk kept frozen and only the affinity module fine-tuned, PairSense is validated on binding affinity benchmarks and three TDC CYP substrate tasks.

Critical strengths

None

Critical weaknesses

1. Missing comparison with the baselines

[Binding affinity]\textbf{[Binding affinity]} The binding affinity comparison in Figure 2 considers FEP+4, OpenFE876, and CASP16, but is missing conventional baselines reported in concurrent works, such as physics-based baselines (FEP+, ABFE, OpenFE) and recent models such as Nesso-1 (open source) and IsoDDE. Without these baselines, the claim of the highest correlation is not supported. Nesso-1 should be compared directly, given its released code and weights; for the closed-source IsoDDE, a discussion would be sufficient.

[CYP substrate]\textbf{[CYP substrate]} Jang et al. (2026), cited in the paper and the source of its ligand-only trunk control, evaluates Boltz-2 representations on the same TDC tasks under the official splits. While Table 2 includes MiniMol, I cannot see Jang et al.’s results on CYP substrates. This leaves the main CYP claim insufficiently supported.

2. Overstatement of ADMET in the title for three CYP substrate tasks

While the title and introduction present the work as ADMET prediction, the evaluation only covers three CYP substrate tasks. The authors note that they have considered eight protein-related TDC tasks and excluded five because of conformer generation failure. However, three small tasks do not support an ADMET-level claim, and therefore I think the title should be changed to protein-mediated metabolism rather than ADMET.

Questions for authors

1. CYP tasks without conformer generation failing cases

Could the authors present the results on the test data with conformer generation failing cases excluded? Though excluding them would break comparability with the official TDC leaderboard, comparison with models including MiniMol and FP2048 on the reduced test split would provide insight on the proposed method, informative for more than three CYP substrate tasks.

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

The authors propose PairSense, a framework built upon Boltz-2 that introduces predicted-distance-based pocket cropping and probability-weighted distogram embeddings while reusing and fine-tuning the existing affinity prediction module. The model outperforms Boltz-2 on several binding affinity benchmarks, achieving SOTA performance. The authors further extend the framework to CYP substrate prediction, demonstrating its potential utility for drug metabolism prediction.

Critical strengths
  • The experimental results are promising, with improvements over several existing methods.
  • Exploring the application of co-folding representations to CYP substrate prediction is interesting and potentially valuable for drug discovery.
Critical weaknesses

I'm not fully convinced of the methodological novelty of this work, as the proposed framework largely builds upon Boltz-2 with relatively minor modifications.

Questions for authors
  1. The text in Figure 1 is much smaller than the main body text, making it difficult to read. Could the authors increase the font size to improve readability?
  2. Given that the title emphasizes ADMET prediction, could the authors clarify why the evaluation is limited to CYP substrate prediction? Have the authors considered evaluating other protein-dependent ADMET endpoints, such as drug transporter substrate recognition, plasma protein binding (e.g., human serum albumin), or substrate recognition by other metabolic enzyme families, to better demonstrate the generalizability of the proposed framework?
Reviewer_w2gQOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

PairSense focuses on the limited reuse of protein-ligand interaction representations for binding affinity and CYP substrate prediction. It combines interaction features and predicted distance distributions from a frozen Boltz-2 trunk to predict binding affinity without generating complex coordinates, then reuses the affinity-adapted representations with molecular fingerprints for joint CYP prediction. It improves affinity prediction over Boltz-2 and achieves higher mean scores than TDC on all three CYP tasks, and additional affinity training improves CYP prediction further.

Critical strengths
  1. The figures and tables are clear and well organized. They make the main results and ablation studies easy to follow and compare.
  2. The experiments show that additional affinity training improves all three CYP tasks, with a notable gain on CYP2D6. This provides useful empirical evidence that affinity-trained representations can support CYP substrate prediction. The paper also includes controlled comparisons of protein context, joint training, and molecular fingerprints.
Critical weaknesses
  1. The methodological novelty of PairSense appears limited. Prior work such as TerraBind [1] already uses Pairformer features and predicted distance distributions for binding affinity prediction without relying on generated coordinates. PairSense follows a similar approach using a frozen Boltz-2 trunk.
  2. The positive transfer from affinity training to CYP prediction is valuable and supported by the experiments. However, the paper provides limited insight into what drives this transfer or why the gains differ across tasks. This limits the broader research value of the finding.

[1] Rossi et al. (2026). TerraBind: Fast and Accurate Binding Affinity Prediction through Coarse Structural Representations.

Questions for authors
  1. Given the similarities between PairSense and TerraBind [1], could the authors explain what motivates the design changes discussed in Sec. 2? Which limitations do these changes address, and what additional benefits or new insights do they provide that are sufficient to support the paper’s novelty and research value? Alternatively, do other aspects of PairSense offer additional benefits or new insights?
  2. Could the authors explain why affinity training improves CYP prediction and why the gains differ across the three tasks? What new insights do these findings provide about the transferability of affinity-trained representations?
  3. Does the proposed geometric input offer benefits over the original coordinate-based input under the same affinity training conditions?
  4. Could the authors report the extent of molecule and scaffold overlap between each task’s test set and the other CYP tasks’ training sets? If overlap exists, do the joint-training gains remain when the splits are aligned across tasks?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.