PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Christina X. Ji

arXiv:2608.30636v1cs.LG

TL;DR

Existing ADME models can predict properties useful for lead optimization, but their atom-level explanations may not be exact or chemically meaningful. MolLedger introduces an additive graph neural network whose per-atom scores sum exactly to predictions and are anchored to physical properties. It achieves comparable predictive performance while producing more chemically faithful explanations, although faithfulness and performance involve a trade-off with anchor strength.

  • Problem

    ADME prediction supports lead optimization and virtual screening, but existing interpretation methods may not provide exact atom contributions or align with physical properties.

  • Method

    MolLedger combines message passing, a global context vector in an additive head, and an auxiliary loss anchoring per-atom scores to chemical properties.

  • Results

    MolLedger achieves similar predictive performance to existing models while its attributions align with physical properties much better than other interpretability methods.

  • Takeaways & Limitations

    MolLedger provides exact, chemically grounded explanations that can help chemists assess model predictions during virtual screening and lead optimization.

  • Takeaways & Limitations

    Faithfulness to chemical anchors trades off against predictive performance as the anchor-loss strength changes, and matched-pair changes may be difficult to localize because ADME properties involve interactions across molecules.

Abstract

from arXiv · show

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 meaningful per-atom attributions. Our model MolLedger outputs predictions that are the sum of per-atom scores. MolLedger's additive framework obtains exact interpretability at no cost to performance because the global context vector gives the additive head enough context to produce good per-atom scores. Furthermore, MolLedger produces attributions that are more faithful to chemical properties than other interpretability methods because the auxiliary loss in MolLedger anchors the atom scores to chemical properties. 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.

1 Introduction

ADME models can accelerate lead optimization, but existing interpretation methods may not reveal exact atom contributions or chemically meaningful explanations. MolLedger addresses this gap with additive, chemically grounded atom-level attributions while maintaining predictive performance.

  • Motivation: Machine learning models can predict ADME properties and support virtual screening during lead optimization.This can reduce the need to synthesize and assay every candidate experimentally.
  • Prior work: Existing post-hoc, attention-based, and additive approaches provide atom-level interpretations but have important limitations.Their attributions may not equal exact atom contributions, align with physical properties, or preserve performance.
  • Contribution: MolLedger produces predictions as exact sums of per-atom scores using message passing, a global context vector, and chemical-property anchoring.The global context vector supplies molecule-level information, while the auxiliary loss anchors scores to physical properties.
  • Contribution: MolLedger achieves similar predictive performance to existing models and produces atom-level attributions more faithful to chemical properties than existing interpretability methods.The paper evaluates these explanations through comparisons across molecular pairs and other interpretability methods.

2 MolLedger

MolLedger preserves exact additive atom-level predictions by conditioning per-atom scores on global molecular context. An auxiliary loss anchors these scores to chemically relevant properties, making them more meaningful for ADME interpretation.

  • 2 MolLedger: MolLedger introduces an additive formulation with a global context vector and an auxiliary loss that anchors per-atom scores to physical properties.These components are designed to provide faithful attributions without sacrificing predictive accuracy.
  • 2.1 Additive Atom Scores with Global Context: Generalized additive models sum functions of individual features, but their restricted form can reduce performance relative to multilayer perceptrons.MolLedger applies additivity to atoms rather than individual features, allowing each score to depend on all atom representations.
  • 2.1 Additive Atom Scores with Global Context: Message passing creates local atom representations, while pooled representations produce a global context vector concatenated to each atom representation before scoring.A second two-layer MLP then produces each per-atom score.
  • 2.2 Auxiliary Loss for Anchors: Prediction-only mean squared error does not uniquely determine sensible atom scores, so MolLedger adds an auxiliary loss toward chemical anchors.The anchors are Crippen score and topological polar surface area.
  • 2.2 Auxiliary Loss for Anchors: Crippen-based anchors support tasks related to lipophilicity, clearance, solubility, and plasma protein binding, whereas TPSA measures polarity for polarity-related tasks.The paper describes task-specific positive or negative relationships between these properties and the anchors.
  • 2.2 Auxiliary Loss for Anchors: MolLedger matches normalized per-molecule deviations between predicted atom scores and anchors, with sign and loss strength adjusted by task.The anchor coefficient is tuned over λ_anchor ∈ {0.1, 0.3} and compared with λ_anchor = 0.

3 Experiment Set-up

The experiments assess ADME prediction and interpretation across diverse tasks, scaffold-based splits, matched molecular pairs, and multiple attribution methods. Evaluation covers predictive quality, chemical faithfulness, exactness, and localization, while explicitly examining the trade-off between anchoring and performance.

  • Datasets and splits: The study compiles 11 ADME tasks from ExpansionRx, Mollipo, and Therapeutics Data Commons datasets.The tasks include properties such as logD, solubility, clearance, permeability, efflux, and plasma protein binding.
  • Datasets and splits: The merged registry contains 14,539 unique molecules split by Murcko scaffold into roughly 80% training, 10% validation, and 10% test sets.Scaffold grouping keeps molecules sharing a scaffold in the same split.
  • Evaluation design: ADME models are evaluated by prediction accuracy, virtual-screening ranking, and change prediction for matched molecular pairs.The corresponding metrics are MAE, Spearman correlation, and matched-pair evaluation.
  • Evaluation design: Matched molecular pairs share a common core and differ by a single swapped side group resembling a medicinal-chemistry refinement.Pairs are constructed with an RDKit fragment-and-index algorithm based on one acyclic, non-conjugated, single-bond cut.
  • Interpretability metrics: Interpretability evaluation asks whether atom scores align with chemical ground truth, sum to predictions, and localize matched-pair changes to the changed R-group.These axes are termed faithfulness, exactness, and localization.
  • Evaluation caveat: Faithfulness to anchors trades off against predictive performance, so the paper examines both as anchor-loss strength changes.This trade-off is analyzed in Appendix B rather than treated as automatically eliminated by anchoring.
  • Interpretability metrics: The exactness gap is zero by construction for both MolLedger and GNAN.Exactness is measured as the MAE between the sum of atom attributions and the predicted label.
  • Baselines and comparisons: The study compares anchored and unanchored MolLedger and GNAN with LigandFormer attention and several post-hoc methods applied to a pooled GNN.The comparisons include integrated gradients, LIME, Grad-CAM, and occlusion-based WISP.

4 Results

MolLedger preserves predictive performance while providing exact, chemically grounded atom-level attributions. Across interpretability evaluations and molecular-pair case studies, its explanations better reflect physical properties and relevant chemical changes than competing methods.

  • 4.1 Model Performance: MolLedger’s additive formulation does not reduce performance compared with the pooled GNN, while GNAN significantly degrades performance on logD, kinetic solubility, and Caco2 A-to-B permeability.Removing MolLedger’s global context vector also harms kinetic solubility and Caco2 A-to-B permeability, indicating that molecule-wide context supports predictive capacity.
  • 4.1 Model Performance: The anchor loss does not hurt performance for logD, kinetic solubility, or Caco2 A-to-B permeability, although intrinsic-clearance MAE for individual molecules worsens in mouse and human liver cells.Pairwise-difference prediction remains similar for intrinsic clearance, and the ability to rank molecules is preserved.
  • 4.1 Model Performance: MolLedger achieves similar or better performance than LigandFormer, the pooled GNN with descriptors, and a gradient-boosted tree with descriptors.The results support the authors’ conclusion that interpretability-oriented additive and anchor constraints do not negatively affect performance.
  • 4.2 Evaluation of Atom-level Attributions: Anchoring significantly improves MolLedger’s faithfulness to physical properties, whereas post-hoc interpretations of the pooled model, LigandFormer attention, and GNAN scores are often not sensible for ADME tasks.MolLedger’s nonlinear message passing and global context vector let atom scores incorporate neighboring-atom identities while matching anchor properties.
  • 4.2 Evaluation of Atom-level Attributions: MolLedger and GNAN have an exactness gap of 0 by construction, while Grad-CAM and WISP show large gaps and integrated gradients trades greater exactness for increased runtime.MolLedger is much faster than the other interpretability methods, GNAN, and LigandFormer; Grad-CAM has similar runtime but lower exactness.
  • 4.2 Evaluation of Atom-level Attributions: Core–substituent leakage is difficult for most tasks because interactions across fragments prevent clean separation, creating a performance–leakage trade-off.MolLedger nevertheless has less leakage than other interpretability methods besides GNAN on logD and the two intrinsic-clearance tasks.

5 Discussion

MolLedger combines exact additive atom-level attributions with chemical anchoring for ADME predictions, retaining performance while producing more meaningful explanations. Its analyses also show that localizing property changes solely to swapped fragments is generally difficult because ADME properties involve interactions across molecules.

  • MolLedger combines an additive GNN head with a global context vector and an auxiliary loss anchoring atom scores to physical properties.
  • MolLedger's attributions sum exactly to ADME predictions and align with physical properties better than other interpretability methods.
  • Localizing matched-pair attribution changes to swapped fragments is likely impossible for most ADME properties because interactions span the molecule.
  • MolLedger retains performance comparable to existing models while providing chemically grounded atom-level explanations.

A Model and Dataset Details

The paper uses a message-passing GNN architecture with multiple-layer perceptrons and harmonizes heterogeneous ADME datasets onto shared label conventions. Half-life is treated as an unanchored task because its relationship to hydrophobicity is not directionally clear.

  • Model: MolLedger uses four GINEConv message-passing blocks with 2-layer MLPs, batch normalization, ReLU, and 128-dimensional outputs.
  • Dataset: Merged datasets are converted to common label scales, including log transformations for several ADME properties.
  • Dataset: Half-life measures the time required for blood or plasma drug concentration to decrease by 50%.
  • Dataset: No anchor is applied to half-life because hydrophobicity accelerates both distribution and clearance, leaving correlation direction unclear.

B Tuning the Context Vector and Anchor Loss in MolLedger

The tuning analyses show that global context preserves additive-model performance, while anchor loss substantially improves attribution faithfulness with little performance cost. A context width of about 8 provides a practical saturation point.

  • Global context vector: Without global context, additive constraints substantially increase MAE on kinetic solubility and Caco2 A-to-B permeability relative to pooled baselines.
  • Global context vector: Adding global context improves performance and kinetic-solubility faithfulness, with benefits saturating around context size 8.
  • Anchor loss: Anchor loss significantly increases faithfulness without affecting model performance across three additive-head variants.
  • Global context vector: The MolLedger head computes per-atom scores from message-passing representations together with a global context vector derived from summed and mean atom representations.
  • Anchor loss: Increasing anchor strength yields a small faithfulness increase with little effect on performance.

C Interpretability Baseline Methods

The paper compares MolLedger with attention, perturbation, gradient-based, and occlusion interpretability methods. These baselines differ in how they construct atom contributions, their exactness guarantees, and their runtime trade-offs.

  • Attention-based methods: LigandFormer interprets atoms using attention averaged across query atoms, heads, and attention blocks, but its weights are shared across tasks in the multitask setting.
  • Evaluation: MolLedger and GNAN have zero exactness gap by construction, whereas exactness is not evaluated for LigandFormer because its attention weights sum to one rather than to the prediction.
  • Gradient methods: Integrated gradients follows a path from an all-zero atom-embedding graph to the molecule, trading exactness against runtime as path steps increase.
  • Gradient methods: Grad-CAM forms a linear approximation in message-passing activation space, and its atom contributions can be positive or negative for regression.
  • Perturbation methods: LIME perturbs molecular inputs, obtains predictions, and fits a linear surrogate to estimate contributions.
  • Perturbation methods: WISP estimates an atom's attribution from the average prediction change caused by valid substitutions of that atom with alternative elements.

D Results on Additional Tasks

Additional matched-pair case studies show that MolLedger’s chemical attributions can capture property-relevant structural changes, while also exposing limitations of its anchors. Comparisons across ADME tasks highlight both successful explanations and cases where competing methods or anchor information are imperfect.

  • Intrinsic clearance: In intrinsic clearance, anchored MolLedger assigns negative scores to the azetidine ring but also to polar groups, whereas unanchored MolLedger misses the ring’s negative contribution.The flexible aminopropyl chain is associated with oxidation risk, while the rigid azetidine is less accessible for oxidation.
  • Caco2 A-to-B permeability: For Caco2 A-to-B permeability, replacing pyrimidine with methoxy removes a P-gp-binding Seelig motif, and TPSA-guided attributions are chemically well-motivated.The pyrimidine contributes two hydrogen-bond acceptors, whereas the ether is less polar and the compound tends to remain inside the cell more.
  • Caco2 efflux ratio: For Caco2 efflux ratio, TPSA anchors are only partially informative because they do not strongly indicate donor strength, causing anchored methods to emphasize both nitrogens in each fragment.The methylimidazole nitrogen can anchor P-gp binding, whereas the substituted pyrimidine nitrogens cannot play the same role.
  • Plasma protein binding: In plasma protein binding examples, replacing lipophilic groups with methylprazole introduces two stronger hydrogen-bond acceptors in both mouse plasma and brain pairs.The original fluorines or weakly polar ring oxygen are described as less effective hydrogen-bond acceptors than the introduced methylprazole nitrogens.
  • Aqueous solubility: For aqueous solubility, adding a hydroxyl in place of hydrogen is attributed by all methods to the hydroxyl or its affected carbon.Because the original substituent has no heavy atom, the change is interpreted through the newly introduced hydroxyl-bearing structure.
Loading 2608.30636v1…