Molecular Graph Convolutions: Moving Beyond Fingerprints

• Drug-design models need molecular representations related to biological activity or experimental endpoints, and more complete, general representations may reduce reliance on application-specific encodings.Source: §1 · p. 1
• Deep learning enables models to use less application-specific raw representations and learn suitable features from input data.Source: §1 · p. 1
• The molecular graph represents atoms as nodes and bonds as edges, with properties such as atom type and bond order, but omits quantum-mechanical structure and may not capture everything medicinal chemists infer.Source: §1 · p. 1
• Molecular graph convolutions represent small molecules as undirected atom graphs and use atom and bond properties plus graph distances to form molecule-level features for conventional machine-learning applications.These representations can be used in place of fingerprint descriptors.Source: §1 · pp. 1–2

2 Related Work

• 2D graph-derived descriptors are inexpensive and interpretable; ECFPs expand heavy-atom fragments for 2–3 steps and hash identifiers, while alternatives encode paths, subtrees, or atom-pair distances [Rogers and Hahn, 2010; OpenEye GraphSim Toolkit; Carhart et al., 1985].ECFPs are also called circular or Morgan fingerprints; atom-pair descriptors encode atom types and graph distances for all atom pairs.Source: §2 · p. 2
• 3D methods represent molecular shape and electrostatics: ROCS aligns pre-generated conformers for shape and chemical similarity, while USR compares intramolecular-distance distributions without alignment.ROCS alignments can also support electrostatic-field similarity calculations.Source: §2 · p. 2
• Multitask neural networks trained on fingerprints and other molecular descriptors consistently outperformed random-forest and logistic-regression baselines [Dahl et al., 2014; Ramsundar et al., 2015].Interest in these networks grew after the Merck Molecular Activity Challenge.Source: §2 · p. 2
• Prior graph-learning approaches update atom states with atom–bond-specific weights [Merkwirth and Lengauer, 2005], transfer local information across edges [Micheli, 2009; Scarselli et al., 2009], or ensemble directed recursive conversions [Lusci et al., 2013].Scarselli et al. find fixed points across local graph functions during training; Micheli applies local information transfer to alkane boiling-point prediction.Source: §2 · p. 2
• A learned generalization of fingerprint computation improved solubility and photovoltaic-efficiency predictions but not binding-affinity prediction [Duvenaud et al., 2015].The architecture learns the fingerprint computation via backpropagation.Source: §2 · p. 2
• Spectral graph convolutions require fixed graph structure [Bruna et al., 2013], whereas manifold convolutions describe deformation-invariant point shapes and combine local descriptors for shape classification [Masci et al., 2015].Masci et al. discuss shape descriptors for points on manifolds such as the surface of a human being.Source: §2 · p. 2

3.1 Deep neural networks

• Neural-network architectures specify neuron operations, parameter sharing, and connections; multitask outputs can share lower-layer computation before task-specific prediction.Training minimizes a chosen loss using back-propagation and stochastic gradient descent.Source: §3.1 · p. 3

3.2 Desired invariants of a model

• The model requires order-invariant outputs for arbitrarily ordered molecular inputs, while atom and pair layer values permute consistently with the input ordering.This makes each atom’s or pair’s representation independent of the ordering of other atoms or pairs.Source: §3.2 · pp. 3–4

3.3 Invariant-preserving operations

• Weave modules combine atom-to-atom, pair-to-pair, pair-to-atom, and atom-to-pair operations, applying shared learned functions and commutative aggregation before modules are stacked.The operations preserve the layer permutation properties; stacking can progressively incorporate longer-range information when pair distances are restricted.Source: §3.3 · p. 4, Figure 4 · p. 5

3.4 Molecule-level features

• Molecule-level features are fuzzy histograms of top-level atom features, with a final atom convolution expanding feature depth before pooling and no ReLU to avoid zero-valued inputs.The histograms use eleven Gaussian bins; pair features influence final atom features through Weave operations but are not directly summarized.Source: §3.4 · p. 5
• The implementation caps initial featurization at 60 heavy atoms, so larger molecules use an arbitrary traversal-dependent subset of atoms and pairs.Only 814 of the 1,442,713 unique dataset molecules exceeded the limit.Source: §3.4 · p. 6

3.5 Input featurization

• Atom and pair inputs use broad, simple descriptors generated with RDKit; experiments omitted explicit hydrogens to conserve memory and emphasize heavy atoms.The feature set was intended to be broad rather than exhaustive, and some features may be derivable from or correlated with others.Source: §3.5 · p. 6

3.6 Datasets

• The benchmark collection comprised 259 datasets from PCBA, MUV, DUD-E, and Tox21, totaling over 38 million data points across many biological target classes.The collection was nearly identical to Ramsundar et al. [2015], with changes to processing and cross-validation fold divisions.Source: §3.6 · p. 8

3.7 Model training and evaluation

• Models were evaluated by ROC AUC using five-fold stratified cross-validation, with each fold divided into 60% training, 20% validation, and 20% test data.Comparators included PMTNN, random forest, and logistic regression on radius-2 Morgan fingerprints, plus maximum Tanimoto similarity to training actives.Source: §3.7 · p. 8
• Models are compared using median 5-fold mean AUC, median per-dataset ∆AUC versus PMTNN, and 95% Wilson intervals for a sign test that assesses whether AUC exceeds PMTNN.Intervals excluding 0.5 indicate significantly different performance; DUD-E results are omitted because all models had median 5-fold mean AUC > 0.98.Source: §3.7 · p. 8

3.8 Comparisons to other methods

• For comparison with neural fingerprints (NFP) [2015], the authors trained multitask graph convolution models with 5-fold cross-validation on aqueous solubility, drug efficacy, and photovoltaic efficiency datasets.The published NFP models were single-task and reported mean squared errors.Source: §3.8 · p. 9
• On the HIV dataset used to validate IRV [2009], graph convolutions were evaluated with 10-fold stratified cross-validation and ROC AUC and BEDROC [2007] at α = 20.RDKit processed 41 476 of 42 678 SMILES strings; each round used one fold for testing, one for validation, and the rest for training.Source: §3.8 · p. 9
• Comparisons with published results are imperfect because fold assignments differ and featurization failures can change dataset composition.The authors describe the resulting comparisons as reasonable, but not exact.Source: §3.8 · p. 9

4.1 Proof of concept

• W2N2 was statistically indistinguishable from the PMTNN on PCBA, MUV, and Tox21; DUD-E was omitted because all models had extremely high median AUC.Several graph convolution models that performed worse than PMTNN had only small median ∆AUC differences.Source: §4.1 · p. 9, Table 4 · p. 10
• Graph convolution models generally performed worse than or comparably to PMTNN on ROC enrichment at 1%, 5%, 10%, and 20% false-positive rates.Subsequent performance analysis and hyperparameter optimization used ROC AUC scores only.Source: §4.1 · p. 9
• Graph convolution results were comparable to published NFP results, significantly improved on photovoltaic efficiency, and had slightly lower HIV AUC and BEDROC than published IRV values.Graph convolution models were multitask-trained on all three NFP datasets; on HIV, they used fewer examples because one fold was held out for validation.Source: §4.1 · p. 9, Table 5 · p. 12

4.2 Input featurization

• The simple featurization—atom type, bond type, and graph distance—achieved similar median 5-fold mean AUC to the full featurization.The additional features may be ignored during training or derivable from the simpler graph representation; assessing individual features may require datasets sensitive to them.Source: §4.2 · p. 11, Figure 7 · p. 11
• Weave modules made atom features more heterogeneous; after two modules, simple and full featurizations produced similarly diverse atom-feature distributions.Initial pair features mostly encoded graph distance, and pair features showed similar evolution.Source: §4.2 · p. 11, Figure 8 · p. 13

4.3.1 Number of Weave modules

• Models with one Weave module were outperformed by deeper models; more than two modules offered little benefit on PCBA and Tox21, while three gave the best median AUC on MUV at increased training time.The comparison used models with two to four modules against a single-module model.Source: §4.3.1 · p. 11, Figure 9 · p. 12

4.3.2 Alternative feature reductions

• Gaussian histogram reductions consistently improved scores over sum reductions; RMS was less robust across datasets, though it had a slightly higher median AUC than Gaussian histograms on MUV.RMS and Gaussian histogram reductions had similar distributions of absolute AUC values.Source: §4.3.2 · p. 11, Figure 10 · p. 12

4.3.3 Distance-dependent pair features

• Pair distance had dataset-dependent effects: N2 improved over N1 on PCBA, while N1 had the best median AUC on MUV and Tox21, and unlimited distance performed clearly worse on PCBA.These results suggest graph convolution models did not effectively preserve or emphasize distance-dependent information; distance-specific weights were explored, but those preliminary results were not reported.Source: §4.3.3 · pp. 12–13

5 Discussion

• Graph convolutions perform comparably to state-of-the-art multitask neural networks trained on traditional fingerprints, neural fingerprints [Duvenaud et al., 2015], and influence relevance voter.The comparison is with models trained on traditional molecular fingerprints and with the two named alternative methods.Source: §5 · p. 13
• For the tested datasets, graph-convolution performance is relatively insensitive to Weave-module count and maximum atom-pair distance; two modules, distance 2, and Gaussian histogram reductions are a suggested starting point.The models also perform well with simple atom-type, bond-type, and graph-distance features.Source: §5 · p. 13, Figure 11 · p. 14
• Because graph convolutions encode the complete molecular graph and tune parameters through backpropagation, they can learn task-specific features rather than act as fixed descriptors.The paper frames graph-based methods as integrated virtual-screening approaches, with learned molecule-level features potentially useful for other tasks or molecular similarity.Source: §5 · pp. 13–14
• Current graph-convolution models do not consistently outperform state-of-the-art fingerprint-based models, but the authors expect optimization could let them exceed the best available methods.Proposed improvements include tuning convolution depths and the P→A reduction, and reducing memory use through more efficient operations or by avoiding all atom pairs.Source: §5 · p. 14
• Molecular graphs omit much biologically relevant information, including 3D shape, electrostatics, quantum effects, and conformational flexibility [Nicholls et al. [2010]].Extending deep learning to 3D biology [Wallach et al. [2015]] requires addressing multiple-instance learning in a relatively small-data regime.Source: §5 · p. 14

Version information

• v2 changed the cross-validation scheme to use a held-out validation set.Source: Version information · p. 14
• v2 added comparisons to additional models and more background for the methods in response to reviewer comments.Source: Version information · p. 14
• v3 added ROC enrichment metrics and changed baseline model training to use sample weights.Source: Version information · p. 14
• v3 added a BEDROC comparison to IRV models.Source: Version information · p. 14
• v3 corrected an error in the logistic regression model training protocol.Source: Version information · p. 14
• v3 updated the method for counting unique molecules, and some AUC values changed slightly after model retraining and/or reevaluation.Source: Version information · pp. 14–15

A Appendix: Model comparison

• The box-plot median confidence intervals do not necessarily correspond to the sign-test confidence intervals reported in Table 4.The box-plot intervals are calculated as ±1.57×IQR/√N [McGill et al., 1978].Source: A Appendix: Model comparison · p. 17
• PCBA performance is shown with full box plots and difference box plots versus the PMTNN baseline.The plots represent absolute 5-fold mean AUC scores and differences in those scores.Source: Figure A.1 · p. 17, A Appendix: Model comparison · pp. 17–18
• MUV performance is shown with full box plots and difference box plots versus the PMTNN baseline.The plots represent absolute 5-fold mean AUC scores and differences in those scores.Source: A Appendix: Model comparison · pp. 17–18, Figure A.2 · p. 18
• Tox21 performance is shown with full box plots and difference box plots versus the PMTNN baseline.The plots represent absolute 5-fold mean AUC scores and differences in those scores.Source: A Appendix: Model comparison · p. 17, Figure A.3 · p. 18

B Appendix: ROC enrichment

• ROC AUC and enrichment metrics were optimized separately on each model’s held-out validation set, so scores at different FPRs may use different training checkpoints.The reported enrichment scores are described as following Jain and Nicholls (2008).Source: B Appendix: ROC enrichment · p. 19
• At 1% and 5% FPR, the tables report median 5-fold mean ROC enrichment, median difference from the PMTNN baseline, and 95% Wilson intervals for sign tests.Bold values mark intervals that exclude 0.5; the captions do not provide the underlying values.Source: Table B.1 · p. 20, Table B.2 · p. 21
• At 10% and 20% FPR, the tables report median 5-fold mean ROC enrichment, median difference from the PMTNN baseline, and 95% Wilson intervals for sign tests.Bold values mark intervals that exclude 0.5; the captions do not provide the underlying values.Source: Table B.3 · p. 22, Table B.4 · p. 23

• For each Section 4.2 experiment, figures show absolute 5-fold mean AUC box plots and differences against a baseline, without y-axis cropping.Source: C Appendix: Input featurization · p. 24
• Figure C.1 compares models with “simple” and “full” input featurizations.Source: Figure C.1 · p. 24

D Appendix: Hyperparameter sensitivity

• For each Section 4.3 experiment, figures show absolute 5-fold mean AUC scores and differences against a baseline model, without y-axis cropping.The figures use box plots for both absolute scores and score differences.Source: D Appendix: Hyperparameter sensitivity · p. 25

D.1 Number of Weave modules

• Figure D.1 compares models with different numbers of Weave modules.The supplied passage does not report which module count performs best.Source: Figure D.1 · p. 25
• The module-count comparison examines variation in the number of Weave modules across models.No specific module counts or results are provided in the supplied passage.Source: Figure D.1 · p. 25

D.2 Alternative feature reductions

• Figure D.2 compares models using different feature reduction methods.The supplied passage does not identify a best-performing reduction method.Source: Figure D.2 · p. 26

D.3 Distance-dependent pair features

• The distance-dependent pair-feature section includes a full box plot.The passage labels this display “(a) Full box plot.”Source: §D.3 · p. 26
• A second distance-dependent pair-feature view is a difference box plot versus the N1 model.The passage labels this display “(b) Difference box plot vs. N1 model.”Source: §D.3 · p. 26
• Figure D.3 compares models with different maximum atom-pair distances.The supplied passage does not provide the comparison results.Source: Figure D.3 · p. 26

E Appendix: Atom pair feature evolution

• Initially similar atom-pair featurizations diverge as they pass through graph convolution Weave modules.The initial features describe ibuprofen and largely encode graph distance between atoms.Source: E Appendix: Atom pair feature evolution · p. 27
• Pairs beyond the maximum atom-pair distance are masked, leaving blank rows; only unique pairs, not both pair orders, are represented.For example, (a, b) is shown without (b, a).Source: E Appendix: Atom pair feature evolution · p. 27
• The full-featurization evolution is shown for a W3N2 architecture, with unique atom pairs arranged by row.Initial pair features appear on the left, followed by Weave module outputs.Source: Figure E.1 · p. 27
• The simple-featurization evolution is shown for a W2N2 architecture, with unique atom pairs arranged by row.Initial pair features appear on the left, followed by Weave module outputs.Source: Figure E.2 · p. 28

F Appendix: Gaussian histogram membership functions

• Appendix F includes Table F.1, titled “Gaussian membership functions.”The supplied passage gives the table title but no membership-function values.Source: Table F.1 · p. 29
• Appendix F also includes a visualization of the Gaussian membership functions.The supplied caption identifies the visualization but provides no further details.Source: Figure F.1 · p. 29