PaperlayerView exact revision on arXiv

Source-linked AI summary

Fragment-Level Graph Encoding for Interpretable Molecular Property Prediction

Rafael Lopes Almeida, Ina Pöhner

arXiv:2610.29589v1cs.LGstat.ML

TL;DR

Atom-level GNN explanations are difficult for medicinal chemists to use because they reason in functional groups and pharmacophores. The paper encodes molecules as BRICS fragment graphs so existing GNNs and explainers produce substructure-level attributions. Across tasks and architectures, fragment encoding retains accuracy on much smaller graphs, and Integrated Gradients recovers known pharmacophores above chance on two enzyme targets.

  • Problem

    Atom-level explanations assign importance to individual atoms, while medicinal chemists reason in functional groups, pharmacophores, and ring systems.

    Source: §1 · p. 1

  • Method

    The method replaces atom-level inputs with BRICS fragment graphs whose nodes use the element-wise sum of their atoms’ features.

    Source: §1 · p. 1, §3.1 · p. 2

  • Results

    Fragment encoding matches atom-level test accuracy on roughly five-fold smaller graphs, while IG pharmacophore hit rates reach 1.5 times chance on BACE and 3.2 times on PTR1.

    Source: §1 · p. 2, §6 · p. 8

  • Takeaways & Limitations

    Fragment training offers a convention-free alternative to unstable post-hoc aggregation without a predictive-accuracy cost, but is not uniformly superior.

    Source: §6 · p. 9

    Molecules without BRICS-cleavable bonds are excluded, and PTR1’s negative test R2 means its comparison is not evidence that either model predicts potency.

    Source: §4.2 · p. 5, §5 · p. 8

Abstract

from arXiv · show

Graph neural networks (GNNs) are among the standard tools for molecular property prediction, but the atom-level explanations of their predictions are hard for medicinal chemists to act on. We propose a change of input representation rather than a new model: molecules are encoded as fragment-level graphs from BRICS decomposition, where each node is a synthetically meaningful fragment whose features are the element-wise sum of its atoms' features. Because only the representation changes, any message-passing GNN and any post-hoc attribution method produces explanations at the resolution of functional groups instead of single atoms. Across four architectures, four prediction tasks, and two explainers under scaffold splits, fragment-level encoding matches atom-level test accuracy on roughly five-fold smaller graphs. Integrated-Gradients attributions recover known pharmacophores above chance on two enzyme targets, namely BACE at 1.5 times the chance rate, and PTR1 at 3.2 times. A like-for-like comparison against an aggregated atom-level baseline shows that post-hoc aggregation is not a stable substitute: its BACE hit rate range both exceeds and falls below the fragment model's hit rate depending on two arbitrary aggregation choices, whereas fragment-level training yields a single convention-free attribution at no predictive cost. The code is available at: https://github.com/4D-Lab/frame

1 Introduction

Atom-level explanations are difficult for medicinal chemists to act on because they reason in functional groups, pharmacophores, and ring systems. The paper addresses this resolution mismatch with BRICS fragment graphs, using the same GNNs and explainers to produce substructure-level attributions.

Motivation

  • Atom-level heatmaps rank individual atoms, while medicinal chemists reason in functional groups, pharmacophores, and ring systems.The paper frames this as a mismatch between explanation resolution and the units used in lead optimization.Source: §1 · p. 1

Proposed representation

  • BRICS fragments become graph nodes with summed atom features, shifting attribution from individual atoms to chemically meaningful substructures.The representation changes without modifying the architecture or explainer.Source: §1 · p. 1

Evaluation protocol

  • The evaluation pairs predictive metrics with pharmacophore hit rate, attribution concentration, and cross-explainer agreement under scaffold splits, including an aggregated atom-level control.The control tests representation change against post-hoc aggregation.Source: §1 · pp. 1–2

Results

  • Across four tasks, four architectures, and two explainers, fragment encoding preserves accuracy on smaller graphs and recovers pharmacophores above chance for localized drug-target interactions.The evidence also delineates where the approach does not help: globally driven properties.Source: §1 · p. 2

2 Background & Related Work

Prior work includes atom-centered GNNs, fragment-aware representations, and post-hoc or architecture-level approaches to fragment explanations. This paper distinguishes its approach by changing the input representation while retaining unmodified architectures and attribution methods.

Molecular GNNs

  • Existing GNN families such as GAT, GIN, PNA, and AttentiveFP use atoms as nodes and bonds as edges, despite differing in message aggregation.Their aggregation strategies include attention, sum aggregation, multiple aggregators with degree scalers, and graph-level attention pooling.Source: §2 · p. 2

Molecular representations and fragment-level methods

  • Other fragment-level methods use custom architectures or modified message passing, whereas this approach feeds fragment representations to unmodified architectures.The passage contrasts the approach with FragNet and fragment-aware representation methods.Source: §2 · p. 2

Interpretability: the two closest works

  • SME masks substructures on an atom-level model, while SEAL modifies message passing; this method instead changes the input representation.The aggregated atom-level baseline is an SME-style control for testing the representation change against post-hoc aggregation.Source: §2 · p. 2

Explainability methods and evaluation

  • Integrated Gradients and GNNExplainer represent gradient-based and perturbation-based post-hoc attribution methods used to test whether findings depend on one explainer.Earlier studies cited in the passage report that IG can highlight features consistent with known pharmacophores.Source: §2 · p. 2

Scaffold splits

  • Murcko scaffold splits keep molecules sharing a core ring system together, and all experiments use them as a proxy for novelty-based evaluation.The passage notes that these splits remain generous relative to truly dissimilar test sets.Source: §2 · p. 2

3 Method

The method replaces atom-level molecular graphs with BRICS fragment graphs and evaluates them under scaffold splits using four GNN architectures and two attribution methods. Its evaluation quantifies plausibility, concentration, and agreement, and includes an aggregated atom-level control.

3.1 Fragment-level graph construction

  • BRICS decomposition creates one node per fragment with the element-wise sum of its atoms’ features; cleaved bonds connect fragments using uniform edge attributes.The graph is deterministic in the stated RDKit implementation, and fragment edges carry no bond-type information.Source: §3.1 · pp. 2–3
  • Sum aggregation preserves fragment-size information, while molecules without BRICS-cleavable bonds are excluded; drug-like graphs shrink to roughly 4–8 nodes.The cited example describes molecules with 20–40 heavy atoms.Source: §3.1 · p. 3

3.2 Data splitting

  • An 80/10/10 Murcko scaffold split keeps molecules sharing a Bemis–Murcko scaffold in the same partition to prevent scaffold leakage.Scaffold groups are assigned deterministically, with larger groups sorted first.Source: §3.2 · p. 3

3.3 Architectures and training

  • Four architectures test whether the encoding works across message-passing approaches, with identical training pipelines for atom and fragment graphs.The architectures are AttentiveFP, GAT/GATv2, PNA, and GIN; reported test metrics average five training seeds.Source: §3.3 · p. 3

3.4 Explainability and its evaluation

  • The aggregated atom-level control sums atom attributions within each fragment, enabling comparison with native fragment attribution but requiring choices about signs, magnitudes, and normalization.The control uses the same architecture and training to isolate representation change from post-hoc aggregation.Source: §3.4 · p. 3
  • The evaluation measures pharmacophore hit rate, attribution concentration via the Gini coefficient, and per-molecule agreement between IG and GNNExplainer rankings.Hit rate credits a fragment only when a matched pharmacophore class on the intact molecule falls wholly inside it.Source: §3.4 · pp. 3–4

4 Experiments & Results

Across scaffold-split experiments, fragment encoding remained competitive with atom-level prediction while producing smaller graphs and pharmacophore-level explanations. Explanation quality depended on target and decomposition: recovery exceeded chance on both enzyme targets, but was limited for BBBP and for motifs split by BRICS.

4.1 Setup

  • The four tasks cover BACE classification and regression, BBBP permeability, and literature-compiled PTR1 regression; PTR1’s 25-molecule test split makes model selection unreliable.The PTR1 dataset contains 250 compounds compiled from 16 medicinal-chemistry resources.Source: §4.1 · pp. 4–5

4.2 Predictive performance

  • Fragment encoding won 11 and lost 5 of 16 dataset–architecture pairs, with a mean primary-metric change of +0.037 and comparable seed-to-seed variance.The BACE classification and regression results were level or favored fragments, PTR1 favored fragments, and BBBP favored atoms.Source: §4.2 · p. 5, Table 2 · p. 5, Table 3 · p. 5
  • Comparable accuracies came from graphs with roughly five times fewer nodes, reducing attribution targets from dozens of atoms to a handful of substructures.For BACE, graph size fell from 34.1 to 7.2 nodes at comparable training cost.Source: §4.2 · p. 5

4.3 Explanation quality

  • Pharmacophore recovery exceeded chance on both enzyme targets: native fragment IG reached 1.5-fold enrichment on BACE and 3.2-fold on PTR1.On PTR1, all three approaches hit the same 18 of 23 test molecules, reaching the BRICS-imposed ceiling rather than separating approaches.Source: §4.3 · pp. 5–6, Figure 3 · p. 6
  • BRICS boundaries constrain which motifs can register as hits: the BACE hydroxyethylamine mimic spans a cut and survives inside one fragment in none of 148 test molecules.In PTR1, the pterin head survives BRICS intact in 88.0% of molecules in which it occurs.Source: §4.3 · p. 6
  • Post-hoc aggregation is convention-sensitive: on BACE, its hit rate exceeds native fragment attribution under two conventions and falls below it under the other two.The choices are signed versus magnitude sums and whether to normalize by fragment size; the authors draw no directional conclusion.Source: §4.3 · pp. 6–7
  • Corrected concentration rankings vary by target, while IG–GNNExplainer ranking agreement spans −0.03 to 0.84 across dataset–approach cells.Native attribution is most concentrated on BBBP, exceeds atom-level attribution on BACE, and trails both alternatives on PTR1; agreement is highest on PTR1.Source: §4.3 · p. 7

4.4 Where it does not help: BBBP

  • For BBBP, directional TPSA hits are modestly above chance, but explanations are less concentrated and 12.9% of molecules lack a cleavable BRICS bond.The authors conclude fragment explanations are most useful for targets with localized SAR.Source: §4.4 · p. 7

5 Discussion

Fragment explanations are most useful when localized pharmacophoric interactions align with BRICS fragments, but data scarcity and evaluation choices constrain what the experiments establish. The discussion also connects fragment attributions to synthesis planning and identifies limits from decomposition, feature aggregation, and explainer dependence.

When fragment-level encoding works best

  • BACE shows 1.5-fold enrichment over 148 molecules and PTR1 3.2-fold over 23, but PTR1’s small test set cannot distinguish the three approaches.BRICS splits BACE’s defining motif in every test molecule and preserves PTR1’s in 88%, limiting the plausibility ceiling according to motif survival.Source: §5 · p. 8

What data scarcity does to the evaluation

  • At PTR1’s sample size, model selection and calibration fail even though attributions remain pharmacophorically sensible, making fragment-level explanation comparisons unreliable.The authors distinguish data-scarcity effects on evaluation from the representation question.Source: §5 · p. 8

Limitations

  • The aggregated-baseline comparison is convention-dependent in both directions, and the authors do not claim a pharmacophore-recovery advantage for native attribution.Establishing a fidelity advantage would require a causal faithfulness study.Source: §5 · p. 8
  • Atom-resolution rescoring against intact-molecule pharmacophore matches removes fragment boundaries, while the plausibility metric itself bounds what the experiments can settle.On BACE, AUC spans 0.26–0.61 with a median of 0.47 across twelve class–approach cells, and 67% lie within 0.1 of chance; PTR1’s two tail classes each occur in only one test molecule.Source: §5 · p. 8
  • BRICS excludes molecules without cleavable bonds, summed features discard intra-fragment connectivity, and individual fragment rankings remain explainer-dependent.The authors suggest coarser or protected-group decompositions as next steps; the two explainers’ agreement has median ρ = 0.33 and spans −0.03 to 0.84.Source: §5 · p. 8

Practical implications

  • Fragment maps can guide lead optimization by preserving high-attribution recognition elements and identifying low-attribution caps as replacement candidates.Because caps are BRICS building blocks, proposed replacements are synthetically accessible; averaged fragment attributions can also form SAR tables connected to synthesis planning.Source: §5 · p. 8

6 Conclusion

The conclusion presents BRICS fragment encoding as a convention-free route to chemically meaningful explanations at competitive predictive accuracy. The appendices document the hyperparameter search space and winning configurations.

6 Conclusion

  • Across four tasks, four architectures, and two explainers under scaffold splits, fragment encoding matches atom-level test accuracy on roughly five-fold smaller graphs.Integrated Gradients recover known pharmacophores above chance on BACE and PTR1; the conclusion reports rates of 1.5 and 3.2 times chance, respectively.Source: §6 · p. 8
  • Post-hoc aggregation is not a stable substitute: its BACE hit rate moves from below to above the fragment model’s under two arbitrary aggregation choices.The conclusion describes fragment encoding as an additional convention-free strategy, not a uniformly superior one; predictive wins are traded with atom-level encoding across four targets.Source: §6 · p. 9

A Hyperparameter search spaces

  • The shared Optuna/TPE search used 150 trials, batch size 32, and three tuning seeds, except PTR1 used batch size 16 and five seeds.PTR1’s smaller training set and validation split motivated the exception; every winning PTR1 configuration fell within the tabulated ranges.Source: A Hyperparameter search spaces · p. 13

B Best hyperparameter configurations

  • Table A.1 lists the hyperparameter search space, while Table B.1 lists winning configurations for each experiment.The supplied passages identify the tables’ scope but do not provide their individual settings.Source: Table A.1 · p. 13, Table B.1 · p. 14

C Complete metric tables

The appendix tables compile classification and regression test metrics for atom-level and fragment-level encodings, with accompanying metric definitions and evaluation conventions.

C Complete metric tables

  • Tables C.1 and C.2 give test metrics for every architecture/encoding cell across classification and regression tasks.The supplied passages identify the tables’ scope but do not provide their cell values.Source: C Complete metric tables · p. 15, Table C.1 · p. 15, Table C.2 · p. 15
  • Classification metrics include accuracy, balanced accuracy, F1, precision, recall, MCC, average precision, and ROC AUC.Most are computed at a fixed 0.5 probability threshold; average precision and ROC AUC use continuous scores and are threshold-free.Source: Table C.1 · p. 15
  • For the listed regression metrics, higher is better except for RMSE, MAE, and delta; the stated conventional thresholds are r2_m > 0.5 and delta r2_0 < 0.3.The passage also defines delta r2_0 as the absolute difference between r2_0 and its inverse.Source: C Complete metric tables · p. 15

D Sensitivity of the aggregated baseline

The sensitivity comparison tests four ways to aggregate atom-level Integrated Gradients attributions against the convention-free native fragment model.

D Sensitivity of the aggregated baseline

  • On BACE, the aggregated baseline’s top-1 pharmacophore hit rate spans 0.223–0.426, bracketing the native fragment model.On PTR1, the baseline spans 0.696–0.783, and the native model ties the upper end because three of the five values coincide.Source: D Sensitivity of the aggregated baseline · p. 15
  • Table D.1 reports hit rates for all four aggregation conventions against the convention-free native fragment model.The conventions vary whether signed attributions or magnitudes are summed and whether results are divided by fragment atom count.Source: D Sensitivity of the aggregated baseline · p. 15, Table D.1 · p. 16

E Matching pharmacophores to fragments

Pharmacophore labels are assigned by matching patterns on intact molecules and crediting fragments that contain every atom of a match. The appendix also identifies the scaffold and graph statistics tabulated.

E Matching pharmacophores to fragments

  • Reference labels come from SMARTS matches on the intact molecule, credited to a fragment only when the full match lies inside it.Matching on isolated fragment SMILES can change hydrogen-count-constrained environments after bond cleavage.Source: E Matching pharmacophores to fragments · p. 16
  • The two labeling protocols disagree on 18.5% of isolated-protocol BACE test labels, while PTR1 is largely unaffected.The parent-molecule protocol is reported throughout; PTR1’s ring-system classes typically remain intact under BRICS decomposition.Source: E Matching pharmacophores to fragments · p. 16

F Scaffold-split and graph statistics

G Sources of the PTR1 dataset

The PTR1 dataset was curated from 16 medicinal-chemistry publications, retaining 250 compounds and 287 numeric Ki values. Some IC50 measurements were converted to Ki under a competitive-inhibition assumption.

G Sources of the PTR1 dataset

  • 250 PTR1 compounds were curated from 16 primary medicinal-chemistry publications.Table G.1 lists sources by the number of compounds contributed after curation; its Reported and Curated columns distinguish original measurements from those retained.Source: G Sources of the PTR1 dataset · p. 17, Table G.1 · p. 17
  • 287 of 303 PTR1 entries had reportable Ki values after curation.The 53 dropped measurements were qualitative or unavailable, including 16 entries with no numeric Ki; no compounds were removed as duplicates, and every source remained represented.Source: G Sources of the PTR1 dataset · p. 17
  • [S] is substrate concentration during IC50 determination and Km the PTR1 Michaelis-Menten constant without inhibitor; Ki could therefore be reported for 287 of the 303 entries.Source: G Sources of the PTR1 dataset · p. 17

H PTR1 under data scarcity

PTR1 attributions recovered the pterin head at 3.2 times chance even though the models did not predict absolute potency reliably. The small, scaffold-split test set limits both predictive evaluation and the resolution of explanation comparisons.

H PTR1 under data scarcity

  • Every PTR1 test R2 was negative, so no configuration beat a constant predictor in squared error.CCC remained positive because it rewards rank-and-scale agreement; absolute potency prediction was not available at this sample size.Source: H PTR1 under data scarcity · p. 17
  • Validation-to-test CCC gaps ranged from 0.24 to 0.60, with 150 trials selecting on only 25 validation molecules.The authors attribute this model-selection limitation to selection on noise as much as signal.Source: H PTR1 under data scarcity · p. 17
  • 3.2 times chance, PTR1 attributions recovered the pterin head even where models were not predictive.The authors caution that plausibility metrics measure agreement with prior chemistry, not fidelity to the model.Source: H PTR1 under data scarcity · pp. 17–18
  • 23.3 to 4.4 nodes per molecule, while fragment encoding won on three of four architectures.The five-fold node reduction held at this scale without predictive cost, according to the authors.Source: H PTR1 under data scarcity · p. 18
  • All three approaches selected the same 18 of 23 PTR1 molecules, limiting the explanation metrics’ ability to distinguish them.The authors say a larger or more diverse test set would be needed to establish equivalence.Source: H PTR1 under data scarcity · p. 18
  • The 0.783 ceiling fell to 0.481 across all 239 decomposable molecules under a different split scope.With a permissive aromatic-ring registry, chance was 0.490 and hit rate was 0.957, still separated at 95% confidence.Source: H PTR1 under data scarcity · p. 18

I Additional attribution maps

Figure I.1 shows native fragment-level IG maps for additional BACE and PTR1 molecules. The PTR1 pterin head remains a single BRICS fragment and ranks highest in all four illustrated molecules, unlike BACE’s cut-spanning hydroxyethylamine motif.

I Additional attribution maps

  • Figure I.1 maps BACE actives and PTR1 test molecules; the PTR1 pterin head is top-attributed in all four molecules, while BACE’s hydroxyethylamine spans a BRICS cut.The figure titles give predicted probability for BACE and predicted pKi for PTR1, alongside the top fragment’s pharmacophore class.Source: Figure I.1 · p. 18, I Additional attribution maps · p. 18
Loading 2610.29589v1…