MolLedger: An Additive Graph Neural Network with Chemically Grounded ADME Attributions
cs.LG
Quick verdict
Additive GNN with anchor loss yields exact, chemically grounded ADME attributions at matched accuracy; promising, but faithfulness evaluation is circular.
Summary
The paper introduces MolLedger, a graph neural network for ADME property prediction whose output is exactly the sum of per-atom scores, making its atom-level attributions exact by construction. Two components enable this: an additive head in which each atom's score is computed from its message-passing representation concatenated with an 8-dimensional global context vector (Eq. 1), and an auxiliary loss (Eq. 2) that anchors atom scores, via normalized deviations, to Crippen atomic lipophilicity contributions or topological polar surface area, with task-specific signs. Experiments merge 11 ADME tasks from ExpansionRx, Mollipo, Therapeutics Data Commons, and ESol (14,539 molecules, Murcko scaffold splits). Across three random seeds, MolLedger matches a pooled GNN, GNAN, LigandFormer, descriptor-based GNNs, and a gradient-boosted tree on MAE, Spearman correlation, and matched-pair ΔMAE, whereas GNAN degrades markedly; anchoring sharply improves faithfulness, keeps the exactness gap at zero, and leaks less than IG, Grad-CAM, LIME, and WISP, with case studies supporting chemically sensible explanations. Limitations include the faithfulness metric's partial circularity (it correlates scores with the same anchors used in training), the absence of significance testing with only three seeds, very small test sets for several tasks (as low as n=17), and worse per-molecule MAE for clearance tasks under anchoring. Code is released.
Review complete
Assessment at a glance
6
risk flags identified
● high 1
● medium 3
● low 2
4 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- An additive head that computes each atom's score from message-passing representations plus a global context vector makes predictions exactly additive over atoms without the performance loss suffered by feature-space additive networks like GNAN (§2.1, §4.1).
- An auxiliary loss anchoring per-atom scores to Crippen contributions or TPSA produces attributions faithful to chemical properties at little to no cost in ranking or matched-pair accuracy (§2.2, §4.2).
- Post-hoc methods (IG, Grad-CAM, LIME, WISP) and attention-based interpretation yield conflicting, chemically less sensible attributions than MolLedger's native scores on matched molecular pairs (§4.3).
- Localizing property changes to swapped fragments (low leakage) is likely impossible for most ADME properties because they depend on cross-fragment interactions (§4.2, §5).
Strengths
Exact atom-level attributions by construction without the usual accuracy penalty
The additive head with a size-8 global context vector (§2.1, Eq. 1) matches the pooled GNN on the five largest tasks (Figure 1), whereas the no-context additive variant degrades on kinetic solubility and Caco2 A-to-B and GNAN degrades on three tasks — cleanly demonstrating the context vector carries the capacity lost to additivity (§4.1, Figure 4).
Well-designed ablation separating architecture from the anchor loss
Applying the anchor loss to GNAN (§3.2, Figure 2) isolates the contribution of the auxiliary loss: anchoring helps MolLedger far more than GNAN, attributed to neighbor-identity information from message passing that GNAN's topology-only weighting lacks (§4.2).
Three-axis interpretability evaluation with matched molecular pairs
Unusually candid failure analysis
Weaknesses
Faithfulness metric is partially circular
No statistical testing; only 3 seeds
Very small test sets for six of eleven tasks
No duplicate/conflict policy for the multi-source merged registry
Abstract overclaims 'no cost to performance'
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
MolLedger (arXiv 2608.30636, cs.LG; single author, Hamilton College) proposes a GNN for ADME regression whose prediction is exactly the sum of per-atom scores. Three central claims are made: (1) an additive head equipped with a global context vector achieves additivity "at no cost to performance" (Abstract, §2.1); (2) an auxiliary loss anchoring atom scores to Crippen atomic contributions or topological polar surface area (TPSA) yields attributions "more faithful to chemical properties" than post-hoc and attention-based interpretability (§2.2, §4.2); (3) matched-molecular-pair case studies show MolLedger's explanations are chemically sensible where competing methods conflict (§4.3). The scope is regression-only: the binary human intestinal absorption task was excluded because "binary tasks do not fit with the additive anchored framework" (Appendix A).
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 4 strengths & 5 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is MolLedger?
MolLedger is a graph neural network that predicts ADME (absorption, distribution, metabolism, excretion) properties of small molecules as the exact sum of per-atom scores, so each atom's contribution to the prediction can be read directly without a separate post-hoc explanation tool. It combines an additive output head with a global context vector and an auxiliary loss tying atom scores to chemical quantities like Crippen lipophilicity contributions and topological polar surface area.
How is MolLedger different from GNAN or post-hoc explainability methods?
GNAN restricts each atom's contribution to a function of that atom's own features only, which this paper finds substantially hurts accuracy on ADME tasks. Post-hoc methods (integrated gradients, Grad-CAM, LIME, WISP) explain an already-trained model and, in these experiments, produce conflicting and chemically implausible attributions. MolLedger keeps full message passing plus a global context vector for accuracy while remaining exactly additive, and it trains the atom scores to align with physicochemical anchors.
Does MolLedger improve ADME prediction accuracy?
No large accuracy gain is claimed. Across 11 ADME tasks and three random seeds, MolLedger performs on par with a pooled GNN, LigandFormer, descriptor-based GNNs and a gradient-boosted tree; its selling point is exact, chemically grounded atom-level attributions at matched accuracy. The paper does note that anchoring worsens per-molecule MAE on the two intrinsic clearance tasks.
Is the code and data available?
Model and experiment code are available at github.com/cxji/molledger. The datasets combine the public Mollipo, Therapeutics Data Commons and ESol sets with the ExpansionRx registry (7,608 molecules) described in a 2026 OpenADMET blind-challenge paper; the paper does not explicitly state how to obtain the ExpansionRx portion.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Numerical results inside Figures 1, 2, 4, 5, 8 and 9 were extracted as garbled axis-tick text streams; exact per-method values could not be read, so quantitative comparisons rely on the paper's narrative plus the readable n=/m= counts and Figure 7 runtimes.
- The molecular structures in Figures 3, 10 and 11 are images; only their textual annotations (labels, per-pair predictions, atom Crippen values) were available, so case-study verification rests on captions and narrative.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.