PaperlayerView exact revision on arXiv

Source-linked AI summary

FP-GNN: a versatile deep learning architecture for enhanced molecular property prediction

Hanxuan Cai, Huimin Zhang, Duancheng Zhao, Jingxing Wu, Ling Wang

arXiv:2205.03834v1q-bio.MNcs.LG

TL;DR

Accurate molecular-property prediction and selecting important molecular descriptors are research challenges. FP-GNN couples molecular graphs with fingerprints and performed outstandingly compared with four recently published methods.

  • Problem

    The paper addresses accurate molecular-property prediction and the challenge of selecting important molecular descriptors.

    Source: Introduction · pp. 2, 4

  • Method

    FP-GNN combines molecular-graph and fingerprint information in a deep-learning architecture.

    Source: Introduction · p. 5, Conclusions · p. 25

  • Results

    FP-GNN performed outstandingly compared with four recently published methods.

    Source: Conclusions · p. 25

  • Takeaways & Limitations

    Ablations found that molecular graphs and mixed molecular fingerprints contributed to FP-GNN prediction performance.

    Source: Conclusions · p. 26

    The authors identify insufficient data quantity as a limitation and suggest pre-trained methods as a potential future direction.

    Source: Conclusions · p. 26

Abstract

from arXiv · show

Deep learning is an important method for molecular design and exhibits considerable ability to predict molecular properties, including physicochemical, bioactive, and ADME/T (absorption, distribution, metabolism, excretion, and toxicity) properties. In this study, we advanced a novel deep learning architecture, termed FP-GNN, which combined and simultaneously learned information from molecular graphs and fingerprints. To evaluate the FP-GNN model, we conducted experiments on 13 public datasets, an unbiased LIT-PCBA dataset, and 14 phenotypic screening datasets for breast cell lines. Extensive evaluation results showed that compared to advanced deep learning and conventional machine learning algorithms, the FP-GNN algorithm achieved state-of-the-art performance on these datasets. In addition, we analyzed the influence of different molecular fingerprints, and the effects of molecular graphs and molecular fingerprints on the performance of the FP-GNN model. Analysis of the anti-noise ability and interpretation ability also indicated that FP-GNN was competitive in real-world situations.

Introduction

The introduction presents FP-GNN as a model combining molecular graphs and fingerprints, motivated by mixed comparisons of graph-based and conventional learning methods. It reports competitive or superior performance across public benchmarks and screening datasets, alongside noise-robustness and interpretability findings.

Introduction

  • Jiang et al. found descriptor-based RF and XGBoost outperformed graph-based deep learning in prediction accuracy and computational efficiency on 11 public datasets.This comparison illustrates the disputed relative performance of graph-based deep learning and conventional descriptor-based methods.Source: Introduction · p. 5
  • The introduction identifies insufficient training data as a potential limitation because GNNs may not learn robust graph representations from small datasets.It also frames accurate prediction of physicochemical, bioactive, and ADME/T properties as a fundamental challenge for molecular design and drug discovery.Source: Introduction · pp. 2, 3, 5
  • FP-GNN combines molecular graphs and fingerprints, using attention to propagate node information and fixed fingerprints as complementary prior information.Its task-specific encoding learns local atomic environments by propagating information from nearby to more distant nodes.Source: Introduction · pp. 5–6
  • FP-GNN achieved comparable or superior performance to all baseline models in 11 out of 16 experiments across 13 public datasets.The paper describes this as strong out-of-the-box performance across a broad range of molecular properties.Source: Introduction · p. 6
  • On unbiased LIT-PCBA, FP-GNN showed comparable or superior performance to fingerprint-based classical ML and graph-based deep learning methods, and performed well on 14 breast-cell-line screening datasets.The breast-cell-line comparisons included XGBoost and graph-based methods such as GAT, GCN, MPNN, and Attentive FP.Source: Introduction · p. 6
  • Noise testing found FP-GNN superior to Attentive FP, XGBoost, and HRGCN+ while retaining predictive power; it also identifies significant graph fragments and fingerprint substructures.The authors suggest these interpretable features can help chemists design molecules with desired functions or properties.Source: Introduction · pp. 6–7

Methods and materials … Molecular fingerprints

The methods encode molecules as graphs for attention-based neural processing, initialize atoms with molecular properties, and use three complementary molecular fingerprints in FP-GNN.

Graph neural networks with attention mechanism

Initial molecule featurization

  • Molecular properties were used to initialize molecular-graph nodes before input to the GNN.The atomic features are provided in Supplementary Table 1.Source: Initial molecule featurization · p. 8

Molecular fingerprints

  • FP-GNN used MACCS, Pharmacophore ErG, and PubChem fingerprints because they complement and holographically express molecular characteristics.Source: Molecular fingerprints · p. 9
  • The study used the short 1+166-bit MACCS fingerprint, a SMARTS-pattern substructure-key representation.MACCS covers atomic and bond properties and atomic neighborhoods at diverse topological separations.Source: Molecular fingerprints · p. 9
  • PubChem was represented by an 881-bit substructure-key fingerprint with extensive coverage of chemical structures.Source: Molecular fingerprints · p. 9
  • Pharmacophore ErG is a 2D fingerprint that uses an extended reduced graph and pharmacophore-type node descriptions to encode molecular properties.Source: Molecular fingerprints · p. 9

FP-GNN network architecture · Hyperparameter optimization and training protocol · Benchmark datasets and performance evaluation metric

FP-GNN processes molecular graphs and three complementary fingerprints through parallel paths, then combines their representations. The study describes Bayesian hyperparameter optimization and evaluates the model across public, unbiased, and phenotypic screening datasets using task-appropriate metrics.

FP-GNN network architecture

  • SMILES molecules enter two parallel FP-GNN paths, one for fingerprints and one for molecular graphs.The path outputs are fitted together and passed through fully connected layers to produce the final output.Source: FP-GNN network architecture · pp. 9–10
  • The fingerprint path concatenates MACCS, PubChem, and Pharmacophore ErG fingerprints into mixed fingerprints, which an ANN represents.The resulting fingerprint vector is input to the artificial neural network.Source: FP-GNN network architecture · pp. 9–10
  • The graph path uses a GNN to capture molecular-graph information, aggregates node representations with attention, and averages nodes for a graph representation.Source: FP-GNN network architecture · p. 10

Hyperparameter optimization and training protocol

  • Hyperopt Bayesian optimization [39] selected six hyperparameters spanning GNN, attention, fingerprint networks, and the GNN-to-FP-GNN ratio.The selected parameters were GNN dropout, number and hidden size of multi-head attentions, fingerprint-network hidden size and dropout, and the ratio of GNN in FP-GNN.Source: Hyperparameter optimization and… · p. 10

Benchmark datasets and performance evaluation metric

  • FP-GNN was tested on 13 commonly used public drug-discovery datasets: three physicochemical, six bioactivity and biophysics, and four physiology and toxicity datasets.Source: Benchmark datasets and performance… · p. 11
  • Additional evaluation used LIT-PCBA and 14 phenotypic screening datasets for breast cell lines.LIT-PCBA covers 15 targets and 7,844 confirmed active and 407,381 confirmed inactive compounds.Source: Benchmark datasets and performance… · p. 11
  • Regression tasks were evaluated by RMSE, while classification tasks used ROC-AUC or PRC-AUC.The supplied passage spells out ROC-AUC as the area under the receiver operating characteristic curve and PRC-AUC as the area under the precision recall curve.Source: Benchmark datasets and performance… · p. 11

Results and discussion · Performance of the FP-GNN network architecture on the public benchmark

Across 13 public drug-discovery benchmarks, FP-GNN achieved the best performance on seven of 16 tasks and consistently matched or outperformed the evaluated baselines. Its strongest results included all three scaffold-split classification tasks.

datasets

  • Datasets were randomly divided into training, validation, and test sets at an 8:1:1 ratio; BACE, BBBP, and HIV also used scaffold-based splits, and FP-GNN metrics were averaged over 10 random seeds.Source: datasets · pp. 12–13
  • FP-GNN performed best on three of eight bioactivity and biophysics tasks: scaffold-split BACE and HIV classification, and PDBbind-C regression.Chemprop led on four tasks, while MoleculeNet's graph-based weave models led on MUV; FP-GNN ranked second on random-split HIV, MUV, PDBbind-F and PDBbind-R.Source: datasets · p. 13
  • FP-GNN achieved the best classification results on both BBBP splits and SIDER, while Chemprop led on Tox21 and XGBoost on ClinTox.FP-GNN also outperformed MoleculeNet's Weave models on ClinTox.Source: datasets · p. 14
  • FP-GNN performed best on FreeSolv; HRGCN+ led on ESOL and Attentive FP on Lipophilicity, where FP-GNN outperformed the other graph-based methods listed.Those methods included GCN, MPNN and Weave.Source: datasets · p. 14
  • Scaffold-split classification performance was lower than random-split performance, but FP-GNN ranked best on BACE, BBBP and HIV under scaffold splitting.The results indicate that scaffold splitting was more challenging and that FP-GNN performed well on molecules with new scaffolds.Source: datasets · p. 15
  • Across 16 tasks, FP-GNN led on seven, versus five for Chemprop and one each for MoleculeNet, Attentive FP, HRGCN+ and XGBoost, while matching or outperforming baselines overall.The authors attribute the improved generalization to coupling molecular graphs with fingerprints.Source: datasets · p. 15

Performance of the FP-GNN network architecture on an unbiased and realistic · LIT-PCBA dataset · Performance of FP-GNN compared to the SOAT graph-based and fingerprint-

FP-GNN showed strong performance on both phenotypic cell-line screening and the unbiased LIT-PCBA benchmark, achieving the highest average AUC across 14 cell lines and the best results on six LIT-PCBA targets.

LIT-PCBA dataset

  • LIT-PCBA contains 7,844 confirmed active and 407,381 confirmed inactive compounds across 15 targets, addressing obvious and hidden chemical biases in artificially constructed public benchmark datasets commonly used by the community.Source: LIT-PCBA dataset · p. 16
  • On LIT-PCBA, FP-GNN performed best on six targets, while NB led on two, DNN on two, GCN on two, and SVM, XGBoost, and GAT on one each.FP-GNN’s leading targets were ADRB2, ALDH1, ESR1_ago, MAPK1, PPARG, and TP53.Source: LIT-PCBA dataset · p. 17
  • FP-GNN outperformed fingerprint-based models and showed comparable or superior performance to GCN and GAT on the challenging LIT-PCBA benchmark.Fingerprint baselines included NB, SVM, RF, XGBoost, and DNN; the fingerprint models used Morgan and mixed fingerprints, and performance was evaluated by ROC-AUC.Source: LIT-PCBA dataset · pp. 16–17

based models on cell-based phenotypic screening datasets

  • The phenotypic-screening evaluation covered 13 breast cancer cell lines and one normal breast cell line, using the same data split and ROC-AUC metric as He et al. and averaging results over 10 random seeds.FP-GNN was compared against published results for four graph-based models and the fingerprint-based XGBoost model.Source: based models on cell-based phenotypic… · pp. 17–18
  • FP-GNN performed best on eight of 14 cell lines; Attentive FP led on three, XGBoost on two, and GCN on one.FP-GNN’s eight leading cell lines were MDA-MB-453, SK-BR-3, T-47D, MCF-7, BT-474, BT-20, BT-549, and HBL-100.Source: based models on cell-based phenotypic… · p. 18
  • 0.849 was FP-GNN’s highest average AUC across 14 cell lines, where it achieved the overall best performance against Attentive FP, GCN, GAT, MPNN, and XGBoost.The FP-GNN models were second-best on HS-578T, MDA-MB-231, and MDA-MB-468.Source: based models on cell-based phenotypic… · p. 18

The Ablation Experiment of FP-GNN. · The influence of different types of fingerprints

FP-GNN combines molecular graph and fingerprint information, and the ablation and fingerprint comparisons indicate that these inputs can complement one another. Mixed fingerprints outperformed Morgan fingerprints in FP-GNN on classification and regression datasets.

The Ablation Experiment of FP-GNN.

  • Among optimal settings across 13 public datasets, 54.3% of FP-GNN GNN ratios were between 0.4 and 0.6; pure GNN and pure FPN models comprised approximately 4.3%.The authors interpreted the ratios as showing relatively balanced contributions from the graph and fingerprint modules.Source: The Ablation Experiment of FP-GNN · p. 19
  • On LIT-PCBA, FP-GNN outperformed FPN and GNN on 10 of 15 targets, and was below GNN but above FPN on most of the other five.For these targets, the reported FP-GNN performance was medium, slightly lower than GNN.Source: The Ablation Experiment of FP-GNN · p. 19
  • The authors suggested that setting the GNN ratio to 0.5 may have left less information from an unfavorable module on the five LIT-PCBA targets where FP-GNN lagged GNN.The ratio was used as the default when building the FP-GNN model.Source: The Ablation Experiment of FP-GNN · pp. 19–20

The influence of different types of fingerprints

  • Mixed-fingerprint FP-GNN outperformed Morgan-based FP-GNN on both classification and regression datasets.The comparison is shown in Figure 5.Source: The influence of different types of… · p. 20
  • Mixed fingerprints encode atomic and bond properties, chemical structures and substructures, and pharmacophore features that may be absent from molecular graph features.The paper attributes their stronger complementarity to graph features, and notes that Morgan records local atom environments that may resemble graph features.Source: The influence of different types of… · p. 21

The Anti-noise Ability of FP-GNN

FP-GNN was evaluated on HIV data with artificially corrupted training and validation labels and unchanged test labels. It achieved state-of-the-art performance in anti-noise tests, suggesting potential for handling poor-quality data in real drug discovery.

The Anti-noise Ability of FP-GNN

  • Drug-discovery data are often scarce and mediocre in quality, and noise can affect training and reduce a model’s real-world practicality.Source: The Anti-noise Ability of FP-GNN · p. 22
  • The HIV dataset of 41,127 compounds was split 8:1:1, with test labels unchanged and training and validation labels altered at predetermined noise rates.Source: The Anti-noise Ability of FP-GNN · p. 22
  • The comparison used the same data, split, ROC-AUC metric, and noise rates as Wu et al. to ensure fairness.The comparator methods were Attentive FP and HRGCN+ (deep learning) and XGBoost (conventional machine learning).Source: The Anti-noise Ability of FP-GNN · p. 22
  • FP-GNN achieved state-of-the-art performance in anti-noise tests, supporting its potential to handle poor-data conditions in real drug discovery.The authors compared it with Attentive FP, HRGCN+, and XGBoost from Wu et al. [54].Source: The Anti-noise Ability of FP-GNN · p. 22

The Interpretation of FP-GNN

FP-GNN’s graph-attention and fingerprint modules offered complementary interpretations of molecular property predictions. The examples linked attention to BBB permeability, while the fingerprint analysis identified bits relevant to hydration free energy.

The Interpretation of FP-GNN

  • FP-GNN computes attention between adjacent atoms and maps it to their connecting bonds; darker molecular portions indicate greater contribution to BBB prediction.The attention coefficients quantify how much chemical fragments contribute to molecular-property predictions.Source: The Interpretation of FP-GNN · p. 23
  • ClogP 2.142 in the red benzene ring versus 1.389 in the grey fragment; FP-GNN emphasized the low-polarity ring in its active-molecule BBB prediction.The ring was described as having the least polarity and the greatest contribution to BBB penetration.Source: The Interpretation of FP-GNN · pp. 23–24
  • ClogP -0.905 in the red fragment versus 0.934 in the grey; high attention to the hydrophilic exposed amino group aligned with the inactive molecule’s BBB prediction.The passage attributes the red fragment’s lower ClogP to its greater hydrophilicity and difficulty crossing the BBB.Source: The Interpretation of FP-GNN · p. 24
  • The 4th, 5th, 7th, and 10th fingerprint bits represented substructures with strong polarity and high water solubility, relevant to free solvation.The analysis perturbed each of 1,489 mixed fingerprint bits in turn; greater prediction deviation indicated a more critical bit.Source: The Interpretation of FP-GNN · pp. 24–25
  • Pearson correlations for the 3rd, 6th, and 10th bits each exceeded 0.7 with hydration-free energy, indicating that the model captured significant fingerprint information.The fingerprint interpretation used the FreeSolv dataset of small-molecule hydration free energies.Source: The Interpretation of FP-GNN · pp. 24–25
  • Of the ten crucial bits, four came from MACCS FP, three from Pharmacophore ErG FP, and three from PubChem FP; together, the three fingerprints played an important role in FP-GNN.Source: The Interpretation of FP-GNN · p. 25

Conclusions

FP-GNN couples a graph-attention network over molecular graphs with an artificial neural network over mixed molecular fingerprints to generate more comprehensive molecular representations.

Conclusions

  • FP-GNN performed outstandingly compared with four recently published graph-based deep-learning algorithms and XGBoost on 13 classical public datasets.The compared algorithms were MoleculeNet, Chemprop, Attentive FP, and HRGCN+.Source: Conclusions · p. 25
  • FP-GNN was highly competitive on the unbiased LIT-PCBA dataset and 14 phenotypic drug-screening datasets related to breast cancer cell lines.Source: Conclusions · p. 26
  • Ablations found that molecular graphs and mixed fingerprints improve prediction, and that their complementary representations play a key role in model performance.Source: Conclusions · p. 26
  • Fingerprint choice affects predictive performance; the current mixture of three fingerprints spanning atomic and bond properties, substructures, and pharmacophores complements the graph module optimally.Source: Conclusions · p. 26
  • FP-GNN showed anti-noise ability for noisy or poor drug-discovery data and can identify important chemical fragments to support molecule design and optimization.The conclusions also describe the model as intuitively interpretable.Source: Conclusions · p. 26
  • Future work may explore pretraining given biological datasets’ insufficient quantity and poor quality, and incorporating protein-target information into a modified FP-GNN.The proposed protein-target approach would combine molecule and target features for prediction.Source: Conclusions · pp. 26–27

Key points · Figure legends · Table legends

FP-GNN combines molecular graphs and mixed fingerprints to predict molecular properties. The paper reports competitive performance, complementary graph and fingerprint information, and chemically interpretable predictions.

Key points

  • FP-GNN couples graph attention networks on molecular graphs with neural networks on mixed fingerprints to generate comprehensive molecular representations.Figure 1 depicts graph attention over each node’s neighbors and the combined FP-GNN model.Source: Key points · p. 27, Figure 1 · p. 34
  • Ablation experiments indicate that molecular graphs and fingerprints provide complementary information that improves predictive power.Figure 4 shows the ablation study on LIT-PCBA.Source: Key points · p. 27, Figure 4 · p. 38
  • FP-GNN’s interpretable predictions identify chemically important fragments that may assist molecule design or optimization.For BBBP examples, highlighted regions align with permeability predictions and differ in ClogP-related lipophilicity.Source: Key points · p. 27, Figure 7 · p. 41

Table legends

  • On LIT-PCBA, Figure 2 compares FP-GNN with fingerprint-based NB, SVM, RF, XGBoost, and DNN models and graph-based GCN and GAT models.Fingerprint baselines use Morgan or mixed MACCS, PubChem, and Pharmacophore ErG fingerprints; graph-based models and mixed-fingerprint models use the same benchmark.Source: Figure 2 · p. 35, Table legends · p. 36

Figure legends

  • Figure 5 compares Morgan-based and mixed-fingerprint FP-GNN models on three classification and four regression datasets, averaging results across 10 random seeds.The models were evaluated after hyperparameter optimization.Source: Figure 5 · p. 39

Table legends

  • Figure 6 reports anti-noise performance for Attentive FP, HRGCN+, XGBoost, and FP-GNN at different noise rates on HIV.The Attentive FP, HRGCN+, and XGBoost results were collected from Wu et al.[54].Source: Figure 6 · p. 40

Key points

  • FP-GNN was highly competitive against classic machine-learning and state-of-the-art deep-learning methods across the reported experiments.The performance tables cover 13 public datasets and 14 breast cell-line datasets, with comparisons against graph-based and other baseline models.Source: Key points · p. 27, Table 1 · p. 42, Table 2 · p. 43

Table legends

  • The performance tables compare FP-GNN with prior models using corresponding published data splits for fair evaluation.Table 1 covers public-dataset comparisons, while Table 2 covers breast cell-line datasets and identifies the models outperforming all others in bold.Source: Table 1 · p. 42, Table legends · pp. 42–43, Table 2 · p. 43

Related papers

Loading 2205.03834v1…