acceptodds
Under review as a conference paper at ICLR 2027

MolLedger: An Additive Graph Neural Network with Chemically Grounded ADME Attributions

Abstract

Optimizing absorption, distribution, metabolism, and excretion (ADME) is an important part of small molecule drug discovery. Many machine learning models have been built to predict ADME properties to facilitate this optimization process, but explaining model predictions is challenging. We propose a new graph neural network architecture with built-in atom attributions. Our model MolLedger learns a global context vector for each molecule and a per-atom head to output atom scores that sum to the predicted property. The atom scores are regularized to align with relevant chemical properties. We prove that MolLedger is a universal approximator and demonstrate empirically that the new architecture obtains explainability with little effect on performance. We show that the interpretations from MolLedger are faithful, concordant with held-out physical properties, and align with the changes between matched molecular pairs. Our case studies comparing interpretations from multiple methods on molecular pairs reveal that MolLedger is much better at producing sensible explanations for predicted property changes.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Machine learning for drug discovery, Molecular machine learning, Interpretability, Explainable AI, Graph neural network

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_tzr9OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes MolLedger, a graph neural network for ADME property prediction with atom-level attributions. The model combines atom representations with a global molecular context to produce additive atomic scores and uses Crippen and TPSA contributions as regularization anchors to encourage agreement with chemical priors. The authors evaluate predictive performance on 11 ADME tasks and assess attribution quality through concordance with physicochemical properties, attribution completeness, and matched molecular pair analysis.

Critical strengths

Understanding the contributions of individual atoms and fragments is an important challenge in molecular property prediction. In drug optimization, molecule-level predictions alone often provide insufficient guidance for structural modifications; understanding how particular atoms or functional groups relate to property changes is equally valuable. MolLedger incorporates atom-level attribution directly into the prediction process and uses global context to account for the influence of the molecular environment on local contributions. This approach addresses a relevant need in medicinal chemistry: interpreting structure–property relationships and formulating hypotheses for molecular modifications while retaining useful predictive performance.

Critical weaknesses
  1. The interpretability evaluation relies heavily on the chemical priors used during training. Although the authors use Crippen and TPSA as training anchors and deliberately evaluate different physicochemical properties, these properties are not independent of the anchors. More importantly, the loss function encourages atom attributions to align with these priors, thereby restricting the model's flexibility to learn other potential chemical effects. The authors aim to improve interpretability by encouraging agreement with the anchors, but if the model mainly follows these priors, what independent explanatory information do its attributions provide? Appendix E (Figure 10) even shows that the Crippen prior can lead to misleading attributions. 2.MolLedger shows improvements on some atom-level attribution metrics, but its performance in chemical property concordance and attribution localization varies across tasks. These results do not yet provide convincing evidence that its attributions are more accurate than those of existing methods. Moreover, the experiments mainly evaluate predictions of ADME properties, including LogD, solubility, clearance, permeability, and plasma protein binding, as well as how these properties change when a single side chain or local functional group is replaced. The model has not been sufficiently tested on complex cases, such as strongly non-additive property changes or interactions among multiple fragments. Whether it can preserve predictive accuracy and provide reliable atom-level explanations in these cases remains unclear.
Questions for authors
  1. What additional explanatory information do MolLedger’s attributions provide compared with directly using Crippen/TPSA atomic contributions? Do the interpretability gains persist after controlling for the information contained in the anchors?
  2. When the chemical priors conflict with experimental observations, can the model reliably correct those priors? Could the authors provide a systematic analysis of such cases rather than individual examples?
1 comment
AuthorsOpenReview ↗

Significance of MolLedger's contributions

Thank you for reviewing our paper. Please see the comment we added above to demonstrate quantitatively how MolLedger is able to learn relevant information beyond the provided anchors. Our results show that MolLedger is significantly better at predicting and explaining the differences between pairs of molecules compared to other interpretable models or posthoc interpretability methods applied to a GNN. As medicinal chemists typically swap one fragment at a time during the lead optimization process and providing per-atom attributions for ADME predictions can help significantly with this process, the focus on evaluating the ADME properties in this paper and how to interpret the effects of fragment swaps is a significant contribution to the field. We hope that addresses your questions.

Reviewer_1bVYOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

MolLedger predicts ADME properties with a message-passing GNN whose atom scores depend on pooled molecular context and sum to the prediction. A correlation-based auxiliary loss aligns these scores with signed Crippen or TPSA contributions. The paper proves expressivity of the readout on a finite collection of node-representation multisets and evaluates prediction, chemical concordance, matched-pair attribution, and runtime across 11 tasks, using scaffold splits and three training seeds.

Critical strengths
  • The context and anchor ablations establish a useful practical tradeoff. Adding context reduces kinetic-solubility MAE and Caco2 permeability MAE, approaching the pooled GNN. Anchoring improves HBD/HBA concordance, and the explanations are inexpensive to obtain.
  • The matched-pair evaluation connects the attributions to measured property changes.
  • The method and writing is easy to understand.
Critical weaknesses
  • The paper interprets agreement with chemical properties that differ from the training anchors as evidence that MolLedger learns chemical intuition beyond those anchors. However, being different properties does not make them independent. In particular, the TPSA anchor already uses polar atom types and attached hydrogens, which are closely related to the hydrogen-bond donor/acceptor labels used for evaluation. A model trained to reproduce TPSA-related preferences could therefore improve on this evaluation without learning the additional chemical relationships claimed.
  • Relatedly, a direct control is to score the same concordance tests using the signed anchor values themselves. This would better establish how much agreement is already present in the training prior. If MolLedger exceeds that reference, it would support learning beyond the anchors on these tests. If it matches that reference, these tests would not establish that the learned scores capture chemical information beyond the supplied prior.
  • IG explains a prediction relative to a reference prediction, which can vary with molecular graph structure (Appendix D; IG completeness result). However, the matched-pair evaluation compares IG fragment-score changes with the full measured property change while discarding changes in those reference predictions, as confirmed by the code. This omission can inflate apparent attribution error even when IG correctly accounts for each prediction relative to its reference. The large attribution-error gap for Caco2 permeability, despite similar prediction error, therefore does not by itself establish that IG's explanations are worse. The authors should quantify the omitted reference changes and check whether MolLedger's advantage persists under a comparison with consistent references.
Questions for authors
  • How do the signed Crippen/TPSA anchor values themselves perform on the same chemical-concordance tests, and does MolLedger improve on that reference? This would clarify whether these results support learning chemical relationships beyond the supplied anchors.
  • Can you report the magnitude of the matched-pair IG reference-prediction differences and show whether MolLedger's attribution-error advantage persists when those differences are accounted for?
1 comment
AuthorsOpenReview ↗

Calculation of explained delta MAE for IG updated

Thank you for taking the time to review our paper and share helpful suggestions for improving the evaluations.

Please see the comment we added above to demonstrate quantitatively how MolLedger is able to learn relevant information beyond the provided anchors.

Regarding the second question about integrated gradients, we appreciate your sharp eye regarding the effect of discarding the difference between the reference molecules in the matched pair evaluation of explained Δ\Delta MAE. Thank you for taking the time to check our code. The calculation has been fixed, and Figures 4 and 16 have been updated in the paper. The explained Δ\Delta MAE for IG has decreased. After this change, MolLedger still performs better than IG on explained Δ\Delta MAE for 10 of 11 tasks. (The only exception is plasma protein bound in mouse muscle, the smallest task).

We hope that improves the soundness of our evaluation of the attributions of MolLedger compared to IG and addresses your questions.

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

The paper proposes MolLedger, an interpretable graph neural network that produces atom-level attributions when predicting molecular properties. Instead of the typical global pooling followed by a prediction head, MolLedger constructs a global context vector and uses it with the node activations to compute per-atom scores. The prediction is then the sum of these scores. Additionally, an auxiliary loss function is defined to anchor per-atom properties to relevant atom descriptors for a given task. The model is evaluated on 11 ADME tasks using both performance-based and attribution-based metrics, demonstrating promising interpretability without degrading performance. A case study presents specific examples of molecular structure replacements scored by the model and compared with other explainability techniques.

Critical strengths
  1. The model proposed in the paper is interpretable and expressive. Theorem 1 shows this expressiveness by proving MolLedger is a universal approximator. This gives the proposed model an advantage over other interpretable molecular models, such as GAMs. Interpretability comes from additive atom-level scores, which guarantee that the attributions sum exactly to the prediction (the exactness property). The experiments confirm performance comparable to other GNN methods.
  2. The authors evaluate model interpretability with multiple metrics and, even more interestingly, visualize attributions in a qualitative case study on matched molecular pairs. These examples show how the model assigns positive and negative attributions to fragments that change the property value upon replacement. This study shows that the model predictions and attributions can be actionable for medicinal chemists.
Critical weaknesses
  1. Unfortunately, the quantitative evaluation of attributions is insufficient to show they are more useful than per-atom Crippen scores. The anchoring introduced with the auxiliary loss causes predicted scores to replicate Crippen scores. As a result, the concordance metric is artificially elevated for the anchored model because the arbitrarily chosen proxy descriptors correlate with logP (semi-circular evaluation + the Authors even admit that hydrophobicity and aromaticity are weak predictors of the evaluated tasks). The unanchored model shows the same level of concordance as other baselines. Suggested changes: As an additional baseline, measure concordance for the atom-level Crippen scores. Show that predicted atom scores are significantly better than the anchor to prove the usefulness of the signal provided by training with the auxiliary loss beyond just replicating anchor results (positive transfer). Discuss for which tasks anchoring does not help the model.
  2. The impact of the global context vector on the attributions should be studied more. Because the global context vector sees the whole molecular graph, it can ignore node information completely, in which case all atoms produce the same scores. Is anchoring the only mechanism that directly prevents this from happening? Other strategies could also reduce the impact of the global context vector and minimize leakage. For example, in SEAL [1], a conceptually similar interpretable GNN, the authors introduce regularization on some edges in the graph. Could a similar mechanism be applied to the global context vector construction without compromising the model expressivity?
  3. In the qualitative results in Figure 5, the anchored model is very close to the Crippen baseline, which further confirms Weakness 1. Moreover, some conclusions may be misleading or overclaimed. For example, the interpretation of LigandFormer's attention weights could be different. The secondary alcohol is already oxidized, so CYPs will not target it, which is why it is highlighted. LigandFormer is also much closer to the true labels, but unsigned attention weights are difficult to interpret. You also claim that MolLedger explains both reasons for the decreased clearance, including a furan being “a metabolic soft spot because CYP450 tends to epoxidize the ring.” There is no proof that the model learned this mechanism from the training data. A more likely explanation is that the model, like the Crippen scores, assigns positive scores to the furan ring as a lipophilic fragment that tends to increase clearance.

[1] Musiał, S., Zieliński, B., & Danel, T. (2025). Fragment-wise interpretability in graph neural networks via molecule decomposition and contribution analysis. arXiv preprint arXiv:2508.15015.

Questions for authors
  1. What λanchor\lambda_{anchor} was used for each dataset?
  2. How much of the prediction can be attributed to the global context vector?
  3. Does leakage increase with the number of message-passing layers?
2 comments
AuthorsOpenReview ↗

Differentiating the effect of the global context vector and message-passing for the atom embeddings

Thank you for reviewing our paper. We are glad you appreciate the interpretable and expressive design and how our evaluation of interpretability with metrics and visualizations shows the actionability for medicinal chemists. We have made revisions to incorporate new analyses around your helpful suggestions into the paper.

Please see the comment we added above to demonstrate quantitatively how MolLedger is able to learn relevant information beyond the provided anchors.

The effect of ablating the global context vector is studied in Figure 6 in Appendix C. Without a context vector, performance is worse. However, even with a very small context dimension of 2 that cannot carry that much information, model performance significantly improves. The results we report in the main paper use a very small context dimension of 8. Unanchored MolLedger achieves concordances significantly above 0.5, so the anchor is not the sole mechanism for learning relevant per-atom scores. We do not use the fragment-based approach in SEAL because there are multiple ways to fragment a single molecule and we prefer to give chemists the flexibility to interpret the attribution on any fragment by summing the scores of the constituent atoms.

Regarding the questions:

  1. The same λanchor\lambda_{\text{anchor}} was applied across all tasks and datasets. We tuned this value in Figure 7 in Appendix C and set it to 0.1. We found that it did not matter if we set lower lambda values for tasks that are less relevant. For instance, we set λ\lambda to 0.1 for LogD and ppb mouse plasma; 0.05 for aqueous solubility, ppb mouse brain, ppb mouse muscle, and Caco2 A-to-B; 0.025 for kinetic solubility, clearance in mouse liver, clearance in human liver, Caco2 efflux ratio; and 0 for half-life. Figure 10 added in Appendix C shows there is no difference in performance between using a constant 0.1 across all tasks and this tiered system based on relevance.

  2. To assess how much of the prediction is attributed to the global context vector instead of the per-atom scores, we applied integrated gradients with 128 steps to the inputs to the per-atom heads and computed ∣ϕG∣∣ϕG+ϕH∣\frac{\lvert \phi_G \rvert}{\lvert \phi_G + \phi_H \rvert}, where ϕG\phi_G is the sum of the attributions to the global context vector and ϕH\phi_H is the sum of the attributions to the per-atom scores. This proportion of the prediction attributed to the global context vector GG versus the atom embeddings HH is shown in Figure 7 added in Appendix C. Most tasks rely primarily on the per-atom embeddings. Only clearance in mouse liver cells relies heavily on the global context vector. This may be because the clearance mechanism involves more atoms than the neighborhood reached by message passing. For other tasks, the per-atom scores mostly rely on the atom embeddings and are thus informative of the attribution to that particular atom.

  3. The question about whether leakage increases with the number of message-passing layers is interesting as it reflects how much leakage comes from the global context vector versus the message-passing process. Figure 8 added in Appendix C shows there is a large decrease in leakage when going from no message-passing layers to 1 message-passing layer, likely because the atom scores now provide some context about the neighborhood and the prediction is less reliant on the global context vector, so less leakage is introduced.

We hope that addresses your questions and demonstrates how MolLedger is able to learn more specific information about the atoms than provided by either the anchors or the global context vector.

Thank you for replying to my comments. Given the late rebuttal responses, I will keep my score for now and consider all authors' responses after the discussion period.

Comments on the paper

1 comment
AuthorsOpenReview ↗

MolLedger learns more than its anchors

Thank you to the reviewers for the thoughtful comments and questions. Regarding the shared question around what MolLedger learns on top of the provided anchors, we have added the additional requested baseline measuring concordance of the atom-level Crippen scores in Figure 3. As the reviewers hypothesized, MolLedger and the anchor have similar concordance as Crippen is related to hydrophobicity and aromaticity. However, we note that our original plot shows unanchored MolLedger achieves concordances significantly above 0.5, so MolLedger learns some of the physicochemical properties even without the anchor.

As an additional evaluation of how much MolLedger learned on top of the anchors, we considered ionization as that is the correction factor in LogD that is not reflected in Crippen. We focused particularly on the 31 carboxylic acids in the test dataset. The ionization factor on the acidic oxygen in a carboxylate group is not reflected in the Crippen anchor. As a sanity check, if we measure the probability the Crippen anchor is lower on a random carboxylic acid oxygen than on a random neutral polar atom, the AUROC is 0.28. This verifies Crippen and ionization factor are not related at all for this atom. Then, we fitted OLS regressions between the anchors and the atom scores and computed the residuals. We would expect these residuals to be very negative on carboxylic acid oxygens to account for the ionization factor. For anchored MolLedger, the average residual on the carboxylic acid oxygen is 3.47 times the average residual on neutral polar atoms. All of the carboxylate atoms have negative residuals. Zeroing the residuals on the carboxylic acid oxygens (using the anchors instead of the predicted atom scores from MolLedger) results in LogD MAE increasing by 0.314, whereas zeroing the residuals on the same number of random neutral polar atoms in the same molecule only results in MAE increasing by 0.048. The AUROC for residuals being lower on carboxylic acid oxygens than neutral polar atoms is 0.997.

This is in contrast to anchored GNAN, which fails to learn about ionization. For anchored GNAN, the average residual on the carboxylic acid oxygen atoms is exactly the same as the average residual on neutral polar atoms. Only half the carboxylate atoms have negative residuals. Zeroing the residuals on the carboxylate atoms results in LogD MAE increasing by only 0.177, whereas zeroing the residuals on the same number of random atoms in the same molecule decreases MAE by 0.005. The AUROC is 0.48. Therefore, MolLedger is much better at learning information not captured in the anchor than GNAN.

We added this evaluation to Section 3.3 with the numbers provided in Table 1. For ease of reference, the table is reproduced here:

ModelEnrichment factor (↑\uparrow)Prop negative (↑\uparrow)Zeroing carboxylic oxygen MAE (↑\uparrow)Zeroing random MAE (↓\downarrow)AUROC (↑\uparrow)
Anchored MolLedger3.471.000.3140.0480.28
Anchored GNAN1.000.500.1770.0050.48

We hope this demonstrates how MolLedger learns more than just the physicochemical anchors it is given.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.