acceptodds
Under review as a conference paper at ICLR 2027

PACT: A Self-Explaining Scoring Function for Protein-Ligand Binding Affinity with Context-Dependent Interaction Terms

Abstract

Structure-based affinity predictors return one number, and an explanation can be added afterward using post-hoc attributions or attention weights that are not expressed in affinity units and whose reliability is contested. Empirical scoring functions instead sum named physical terms but give each term one global weight, pricing a solvent-exposed hydrogen bond like a buried one. We introduce PACT, a predictor whose output is exactly decomposed into signed ligand-atom-residue contributions in pK over twelve interaction channels, with local chemistry and pocket context rescaling each contribution. Additivity determines how affinity is distributed over contacts, not whether that distribution can be trusted, so we evaluate the stability of the contributions under model retraining and contact removal. On a leakage-reduced split, PACT matches the accuracy of the strongest baselines we evaluated while producing faithful and reproducible contributions.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
interpretability, XAI, scoring functions, cheminformatics, computer-aided drug design

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_An95OpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
3(+1)
Read review
Paper summary

PACT is a structure-based affinity scoring function whose prediction can be decomposed into signed contributions over twelve interaction channels (in pK units), with bounded local corrections from chemistry and pocket context. The main claim is that the explanation is what the model actually computes, not a post-hoc attribution.

Critical strengths
  1. The prediction adds up exactly from contact-level contributions in affinity units, which is more direct than attention or post-hoc methods.
  2. The contributions are clearly more stable and reproducible under retraining and small geometric perturbations than most baselines.
  3. When high-ranked contacts are removed, the change in the score lines up well with the assigned contributions, so the explanation is largely faithful to the model.
Critical weaknesses

The contributions are reported in pK units, but the paper never checks whether they correlate with real or approximate interaction energies. There is also no test of whether these contributions are actually useful for downstream design decisions. As written, the evidence mainly shows faithfulness to the model, not physical meaning or practical value.

Questions for authors

Could the authors report the correlation between several key channel-level contributions and external energy estimates such as MM/PBSA on a subset of complexes?

5 comments

We thank the Reviewer for their time and for the questions and suggestions. We ran new experiments and included them in the revision.

Q1. Correlation between channel contributions and external energy estimates.

We thank the Reviewer for suggesting we evaluate our method against MM/PBSA. We note that PACT’s interaction channels are predefined physical forms, and its learned contributions are fitted to experimental binding affinity, not specific interaction energies. We test these predicted interactions against a subset of the PLAS-20k dataset (new Section 5.5), which reports MM/PBSA components computed from molecular dynamics trajectories. There are 888 complexes in our test set that appear there with stable component values. We assigned each of the twelve channels to one component, then we applied this mapping and computed the correlation (Appendix D.8). The correlation values in every row have the expected sign, which does not change after removing the size effect. All values have p<10−3p < 10^{-3}.

MM/PBSA componentChannel groupρ\rhosize-controlled ρ\rho
VDWAALSvdW, repulsion, aromatic-0.779-0.504
ENPOLARhydrophobic, apolar burial-0.595-0.406
EPBpolar burial+0.459+0.400
EELsix directional polar channels-0.220-0.124

MM/PBSA reports only four components, so four groups are all this method provides. While the correlation is not perfect, the direction is consistent across all components, even though PACT predicted its channels from a static PDB frame rather than an MD trajectory.

W1. Practical value for design decisions.

In the original submission, the matched molecular pairs experiment best demonstrates our model's practical utility. We first used an external benchmark, the 3D-MMP set of Kalinowsky et al. [1], which contains 99 matched pairs, 178 structures and 33 target clusters. For each matched pair, we recorded what the edit added or removed. We read it from the same input atom features the channels use without checking any channel’s value. Labels count input atom features and, based on them, assign classes. We then tested whether the responsible channel moves as the chemistry implies. It does in 0.649 [0.576, 0.728] of 171 edits. It is important to add that for some complexes, the protein changes, so it might affect the prediction. Across different heavy-atom counts, the result stays flat (0.618, 0.771 and 0.618). It also achieves 0.652 on the complexes that are only in the test set. This shows that PACT performance is neither from size nor from training subset membership. New Appendix B.4 gives the protocol and per-class results.

In the current revision, to further address the question of practical usefulness of the model, we retrospectively analyzed two published lead optimization papers. Each describes a series of compounds from one study. The authors made one modification at a time and solved a crystal structure for every step. They also state which interaction each change was meant to create or remove. We have performed two experiments. The first is based on a study in which the authors report a series of modifications, one of which introduces an interaction that disappears when an additional functional group is attached. The channel corresponding to this interaction should fire only for the modification that forms the new interaction. The second experiment is based on a study that analyzed a group of BACE1 inhibitor analogs. Potency rises about ninety-fold across the series, but the interactions with the molecular core should remain the same. Both are described in detail in the new appendix D.11.

(continued in the next response)

Experiment 1 - uPA [3]. The starting compound carries a chlorine atom at position 6, and the authors replaced it with larger groups that reach into a side pocket. A pyrimidine ring forms a hydrogen bond to Gly219 and keeps it. A second contact to Arg217 depends on how the ligand sits: it is present in one compound and disappears in the next one. We assigned every contribution only to the added group, so the unchanged part of the molecule cannot produce them.At Gly219, a hydrogen bond appears with the pyrimidine ring and stays. At Arg217, it appears with the first methoxy substitution and then drops by 95.8% for the second substitution, as the interaction distance increases from 3.425 to 4.262 A. The molecule grows by two heavy atoms and gains two contacts at Arg217. Size and occupancy therefore both increase, and the named channel decreases. The test is which channel carries the change, and whether the direction is correct.

Compound / PDB6-substituentGly219 H-bond, pKArg217 H-bond, pK
5 / 5ZA7Cl00
25 / 5ZAGpyrimidine0.05380
26 / 5ZAHmethoxypyrimidine0.05450.0250
27 / 5ZAJdimethoxypyrimidine0.05000.0011

Experiment 2 - BACE1 [4]. In this experiment, all molecules in the series should bind BACE1 by forming the same key interactions with aspartic acids. Across the six held-out complexes, the ionic term on Asp93 and Asp289 varies by only 1.4% (stable across molecules), while measured potency resulting from other substitutions spans 1.95 pK. On the other hand, the aromatic contribution calculated by PACT, which was supposed to change, moves from 0.026 to 0.151 pK. The same holds outside the series. In all 56 BACE1 test complexes, the ionic term is non-zero only on a catalytic aspartate. It is not fixed to aspartate, though: arginine and lysine carry 27% of the ionic signal in the test set.

PACT behaves correctly in both experiments, showing that the model can provide valuable insights about binding interactions in real drug design campaigns. We hope that this addresses the Reviewer's concerns, and we are glad to answer any further questions.

[1] Kalinowsky, L. et al. A diverse benchmark based on 3D matched molecular pairs for validating scoring functions. ACS Omega 3, 5704-5714 (2018).

[2] Korlepara, D. B. et al. PLAS-20k: extended dataset of protein-ligand affinities from MD simulations for machine learning applications. Scientific Data 11, 180 (2024).

[3] Buckley, B. J. et al. 6-Substituted hexamethylene amiloride (HMA) derivatives as potent and selective inhibitors of the human urokinase plasminogen activator for use in cancer. J. Med. Chem. 61, 8299-8320 (2018).

[4] Mandal, M. et al. Design and validation of bicyclic iminopyrimidinones as beta amyloid cleaving enzyme-1 (BACE1) inhibitors: conformational constraint to favor a bioactive conformation. J. Med. Chem. 55, 9331-9345 (2012).

Dear Reviewer An95,

Thank you for your focused and constructive review. Your suggestion to compare our contributions with MM/PBSA components became a new section, and your point about practical value led us to test the method on an external benchmark and on two published medicinal chemistry campaigns. All changes are marked in blue in the revised paper.

As the discussion period closes soon, we would greatly value your thoughts on these additions. If anything is unclear or you would like more detail, we would be glad to respond promptly.

Thank you again,

The Authors

Reviewer_nNzMOpenReview ↗
Rating
3
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies structure-based protein-ligand affinity prediction with intrinsic, chemically labeled explanations. It proposes PACT, an additive contact-term model that reweights twelve predefined interaction bases with bounded local context, yielding exact signed ligand-atom-residue contributions in pK. Experiments on leakage-reduced LP-PDBbind assess predictive performance, cross-retraining reproducibility, removal-based contribution-effect agreement, and chemical interaction recovery.

Critical strengths

The model has a genuinely intrinsic and auditable decomposition. Equation 1 makes the prediction exactly equal to an intercept, bounded ligand-level term, and signed group-channel contributions, while the construction in Sections 3.1-3.2 ensures that learned contextual factors cannot activate an absent chemical basis or reverse its prescribed sign. This is substantially more informative than merely attaching an attribution map to an opaque affinity model, because the displayed ledger is the actual computation.

Critical weaknesses

The broad baseline claim should be interpreted more cautiously because the paper itself acknowledges non-uniform baseline tuning and reproduction protocols. Section 6 states that several baselines "were developed and tuned on other splits" and are run with published settings, so their accuracy may be below their best. Since Table 1 is used to support the claim that PACT "matches the strongest baselines," the comparison is informative but not a fully controlled estimate of the best achievable performance of each architecture. This does not undermine PACT’s intrinsic decomposition or its within-model intervention results, but it weakens comparative accuracy claims. Reporting which baseline hyperparameters, early-stopping rules, and tuning budgets were revalidated on the LP-PDBbind validation split, and separating fully retuned from published-setting results, would resolve the concern.

Questions for authors

For the Table 1 comparison, which of the reported baselines had their hyperparameters and checkpoint-selection rules revalidated on the LP-PDBbind validation split, rather than being evaluated under transferred published settings? If all leading baselines received comparable validation tuning, this would materially increase my confidence in the "matches the strongest baselines" claim.

2 comments

We thank the Reviewer for their time and for the questions and suggestions. We have made revisions, and we describe them below under W1.

Q1. Equal tuning budget on the validation set

As stated in Appendix E.1, we retrained every baseline from scratch on the recommended LP-PDBbind training split. We did not use any pretrained checkpoints. We only reused the published hyperparameter settings and training strategies. For every baseline, we selected the checkpoint on the recommended LP-PDBbind validation split, except for two baselines that keep their original upstream selection procedure. Both are described in Appendix E.1 and neither favors PACT. We did not perform a hyperparameter search for any baseline, so each one uses its published configuration. We did not perform an exhaustive search for PACT neither. Achieving the lowest error was not the main goal of this work. As we reported in Appendix C.3, we did not even choose the variant with the lowest RMSE, since model interpretability and faithfulness were our main priorities. Following the Reviewer’s suggestion to state which hyperparameters and selection rules were revalidated on this split, we have revised Section 6 (Discussion and limitations) to describe it explicitly.

W1. Equal tuning budget on the validation set

Following the Reviewer's suggestion, we ran hyperparameter tuning for the two strongest baselines, CheapNet and IGN, which were tied with PACT. For each of the two models, we evaluated 12 configurations by modifying the learning rate and weight decay by one order of magnitude and the width by a factor of two. Each configuration was trained with three seeds (0-2). In both cases, the published settings turned out to be the best, especially for CheapNet. Although a few configurations achieved a slightly lower validation RMSE, none improved on the results in Table 1, so we kept them.

We revised the caption of Table 1 so that it clearly states that models were trained using their published parameters. We also ran 10,000 paired bootstrap samples to compare the best baselines statistically. It was computed on the 2,133 test samples shared by all models, so every model is compared on exactly the same data. The RMSE difference from PACT is +0.006 for IGN (95% CI [-0.019, +0.030]), +0.014 for CheapNet (95% CI [-0.010, +0.038]) and +0.021 for RF-Score (95% CI [-0.005, +0.048]). All of those confidence intervals contain zero. This means that we cannot say that the difference is statistically significant. For a baseline to become significantly better than PACT, its RMSE would need to improve by more than 0.030 for IGN, 0.038 for CheapNet and 0.048 for RF-Score.

As suggested by the Reviewer, we tuned the leading baselines using a budget comparable or greater than PACT, which has not resulted in higher scores on the test set. We hope this resolves all of the Reviewer’s concerns, and we are glad to answer any further questions.

Dear Reviewer nNzM,

Thank you for your careful review and for recognizing the value of PACT’s intrinsic decomposition. Your question on baseline tuning shaped our revision: we retuned the two strongest baselines, CheapNet and IGN, on the LP-PDBbind validation split, added a paired bootstrap comparison, and clarified the protocol in Section 6 and Appendix E.1 (marked in blue).

We hope this addresses your concern. Before the discussion period closes, we would be glad to clarify any other aspect of the paper or provide further detail where helpful.

Thank you again,

The Authors

Reviewer_DLu9OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces PACT, a scoring function for protein–ligand binding affinity prediction. It computes binding affinity as a weighted sum of predefined chemical basis functions, where the weights are learned and depend on atomic and interaction features.

Critical strengths

The paper provides an explicit decomposition of predicted binding affinity into individual chemical contributions.

Critical weaknesses
  1. The implementation of PIGNet2 differs from the original paper, which makes it unclear whether the comparison fairly reflects the performance of PIGNet2.
  2. The paper calculates its final score by adding up individual interaction terms, with limited interactions between them. Therefore, the consistency between the predicted contributions of individual interactions and the changes caused by contact removal may largely result from the model design, making this experiment less informative.
Questions for authors

See above weaknesses.

2 comments

Thank you for reading the paper carefully. Both weaknesses are fair questions, and we can answer both from results already in the paper.

W1. The PIGNet2 comparison.

The reviewer is right that the submission did not make this explicit. Section 6 mentions the retraining protocol, but Appendix E.1 listed the differences without saying what they cost. We have now measured each one.

Two differences are specific to PIGNet2: we used the repository’s default split 6-6 Lennard-Jones van der Waals term rather than the Morse potential of upstream’s training scripts and released checkpoints, selecting it because it scored better on validation, and we did not apply upstream’s xavier_normal_ re-initialisation of parameters of rank ≥ 2. Everything else: the scoring-only task, checkpoint selection on the LP-PDBbind validation split, and the shared preprocessing, applies identically to every baseline in Table 1.

To establish the reference point we ran upstream’s repository end to end, unmodified, on the same cohort. Three seeds per variant, one graph cache, the same 1,862-complex LP-PDBbind test set. The implementation in the original repository gives RMSE of 1.8915 log units (the table below).

VariantTest RMSE (pK)
Upstream’s own trainer and model, unmodified (Morse)1.8915 ± 0.0387
Morse, our pipeline1.8707 ± 0.0552
Lennard-Jones, w/o xavier (submitted baseline, reproduced)1.8463 ± 0.0579
Lennard-Jones, w/ xavier (strongest PIGNet2)1.7623 ± 0.0332

What this shows: the vdW substitution, the single largest difference and the one the reviewer would most reasonably suspect, moves PIGNet2 the wrong way, so our baseline was already stronger than published PIGNet2. The initialiser is the one difference that did cost PIGNet2, restoring it yields 1.7623, better than any configuration upstream publishes, and we now report that in Table 1. The rest are protocol-wide or immaterial.

The ranking in Table 1 is unchanged. PACT stays 0.258 pK ahead of the strongest PIGNet2 available. We will update all affected results in the paper.

W2. Does the additive design make the contact-removal test trivial?

We thank the reviewer for this question. No, the interaction terms are there so that affinity is routed through protein-ligand structure, because models can score well on dataset-level patterns that do not describe binding at all (Durant et al., 2025; Volkov et al., 2022; Sieg et al., 2019). That design decides how affinity is distributed over contacts. It does not establish that the distribution can be trusted, which is why the paper measures it rather than asserting it: “additivity determines how affinity is distributed over contacts, not whether that distribution can be trusted, so we evaluate the stability of the contributions under model retraining and contact removal" (abstract). Your objection would be decisive if the removal agreement were an arithmetic consequence of the sum. It is not, and that can be measured:

An additive decomposition can fail this test (Table 3). PIGNet2 is the comparison that bears on your point, because it is also a physics-informed model whose interaction pairs sum to the prediction, scored on its own native contributions. Yet its discrepancy is 1.591 pK against effects of 1.133 pK (140.4%), with slope 0.490. Additivity therefore does not by itself deliver agreement.

PACT’s agreement is close but not exact (Table 3). Sparsemax allocation makes the groups sharing a ligand atom compete, and the group update is an MLP reading all twelve pooled channels of its own group, so deleting one contact changes what the surviving contacts are worth. The removal test therefore has room to disagree with the assignment. If the result followed from the sum, the discrepancy would be zero. It is 0.191 pK, about 17% of the removal effects, and the slope is just below 1. The gap comes from removing residues redistributing van der Waals attraction among those that remain, which the decomposition does not anticipate.

The comparison is conservative for PACT (Table 3). The other baselines rank residues by their own measured deletion effects, so they agree with the smallest removals by construction. PACT ranks by its assigned contributions and has no such advantage. It still has the lowest relative discrepancy (17% against 24% for the next-best model) and the slope closest to 1.

The additive form does not determine the explanation (Table 6). All ablations share the same sum over the same twelve channels and have nearly identical accuracy (within about 0.09 pK RMSE). Their top-five agreement across retraining nevertheless ranges from 0.60 to 0.91. The constraints, not the additive form, drive PACT’s behaviour.

We will modify Sections 3, 6, and Appendix E.1 to state this more clearly.

Dear Reviewer DLu9,

We have now run all the experiments described in our earlier response and revised the paper accordingly.

We trained the PIGNet2 variants from the earlier response on one shared graph cache and the same test set, five seeds per variant. The strongest is the Lennard-Jones model with upstream’s Xavier initialisation, at 1.764 ± 0.025 pK. We report it in Table 1, and every result that depends on that checkpoint has been recomputed: Tables 2, 3 and 4, the selection fidelity and sparsity tables, and the attributed-mass and share numbers. The rankings and conclusions remain unchanged.

Appendix E.1 now describes how each difference affects the performance of PIGNet2 instead of simply listing them. We also revised Sections 3, 5.3 and 6 to state more clearly what the additive form determines and what we measure.

All changes introduced during the discussion period are marked in blue in the revised paper. Thank you again for your time. We hope our responses and the revised paper address your concerns.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.