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

Molecular property prediction needs representations that capture relevant chemical information across diverse properties and datasets. FP-GNN jointly learns from molecular graphs and mixed fingerprints, and the reported evaluations found strong performance across public benchmarks, LIT-PCBA, and phenotypic screening datasets. The authors conclude that both graph and fingerprint information contribute to prediction performance.

  • Problem

    Selecting informative molecular features for accurate and interpretable prediction remains challenging, while the relative performance of graph-based deep learning and conventional models is unsettled.

  • Method

    FP-GNN couples molecular graphs with mixed molecular fingerprints to learn graph features and incorporate complementary fixed fingerprint information.

  • Results

    FP-GNN performed outstandingly compared with four recently published models, while ablations found that molecular graphs and mixed fingerprints contributed to improved prediction performance.

  • Takeaways & Limitations

    The results support using combined graph and fingerprint representations for molecular property prediction across varied benchmark and screening settings.

  • Takeaways & Limitations

    The authors identify insufficient data quantity and quality as a motivation for exploring pre-trained methods in future work.

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

Molecular property prediction requires informative, interpretable representations, but the relative advantages of graph-based deep learning and conventional models remain unsettled. FP-GNN combines molecular graphs with fingerprints and performs competitively across diverse evaluation settings.

  • Accurate prediction of physicochemical, bioactive, and ADME/T properties remains a fundamental challenge in molecular design.
  • Selecting property-relevant hand-crafted descriptors from many predefined molecular descriptors affects both prediction accuracy and model interpretability.
  • Whether graph-based deep learning is better than conventional descriptor- or fingerprint-based machine learning remains controversial.
  • Prior comparisons found conventional descriptor-based models, especially RF and XGBoost, outperformed graph-based deep learning in prediction accuracy and computational efficiency.
  • FP-GNN combines molecular graphs and fingerprints, using attention-based graph encoding alongside fixed complementary fingerprint information.
  • The authors report that fingerprints improved graph-based model generalization, while FP-GNN retained predictive power under noise and supported extraction of significant fragments and substructures.

Methods and materials

The graph pathway represents each molecule as an undirected atom-and-bond graph and updates node states by aggregating neighboring information. An attention mechanism weights neighbor contributions before producing a graph-level representation.

  • Each molecule is transformed into an undirected graph G(V, E), with atoms as nodes and chemical bonds as edges.
  • The spatial-GNN updates each node by aggregating information from itself and its neighbors.
  • Graph attention: Graph attention computes neighbor-specific weights for information passed from node j to node i.
  • Graph attention: The attentions of all neighbors are normalized, then used to update each node; repeated attention calculations are averaged.
  • Graph-level aggregation: After node updates, the model aggregates the complete molecular graph into an output representation.

Initial molecule featurization

FP-GNN featurizes each molecule through parallel fingerprint and graph pathways. It combines three complementary fingerprints with attention-based molecular-graph representations before producing the final prediction.

  • Parallel molecular representations: SMILES notation is input into two parallel FP-GNN pathways: one for fingerprints and one for molecular graphs.The graph pathway initializes node features from molecular properties, aggregates each node with its neighbors by attention, and averages node representations.
  • Fingerprint pathway: The fingerprint pathway concatenates MACCS, PubChem, and Pharmacophore ErG fingerprints into a mixed fingerprint vector.These fingerprints provide substructure-key, chemical-structure, and pharmacophore representations.
  • Fingerprint pathway: The fingerprint vector is passed through an artificial neural network to obtain its learned representation.This transformation is represented by Equation 9.
  • Representation fusion: The learned fingerprint and graph representations are fitted together and passed through fully connected layers to produce the final output.The architecture therefore combines complementary graph and fingerprint information before prediction.

Hyperparameter optimization and training protocol

FP-GNN hyperparameters were optimized with Bayesian optimization and evaluated across diverse molecular-property datasets using task-appropriate metrics and repeated-seed protocols. Across 16 tasks from 13 public benchmarks, FP-GNN was best on seven tasks, while scaffold splitting was more challenging than random splitting.

  • Optimization: Bayesian optimization with Hyperopt selected six hyperparameters, including GNN dropout, attention settings, fingerprint-network settings, and the GNN ratio.The selected parameters included the number and hidden size of multi-head attentions, fingerprint-network hidden size and dropout, and GNN dropout.
  • Datasets: The evaluation covered 13 public drug-discovery datasets spanning physicochemical, bioactivity and biophysics, and physiology and toxicity tasks.The benchmark collection included ESOL, FreeSolv, Lipophilicity, MUV, HIV, BACE, PDBbind, BBBP, Tox21, SIDER, and ClinTox datasets.
  • Evaluation: Classification tasks used ROC-AUC or PRC-AUC, while regression tasks used RMSE.PRC-AUC was used for the highly imbalanced MUV datasets because it could better reflect classification performance there.
  • Training protocol: Datasets were randomly split into training, validation, and test sets at an 8:1:1 ratio, with BACE, BBBP, and HIV also evaluated using scaffold-based splits.Hyperparameters were optimized, and final metrics were averaged across 10 random seeds.
  • Benchmark results: Scaffold-based classification was lower-performing than random-splitting, yet FP-GNN performed best on all three scaffold-based datasets and showed stable performance on new scaffolds.The scaffold-based results were reported for BACE, BBBP, and HIV.
  • Benchmark results: Seven of 16 benchmark tasks favored FP-GNN, compared with five tasks favoring Chemprop.FP-GNN was best on three of eight learning tasks in one benchmark grouping, while Chemprop was best on four; MoleculeNet was best on MUV.

LIT-PCBA dataset

LIT-PCBA provides an unbiased, realistic evaluation setting with 15 targets and strongly imbalanced confirmed active and inactive compounds. FP-GNN showed the strongest overall competitiveness against fingerprint-based and graph-based baselines across the reported targets.

  • Dataset: The dataset was designed to address chemical biases associated with commonly used artificially constructed benchmarks.Training and validation sets were constructed with the asymmetric validation embedding method at a 3:1 ratio.
  • Dataset: LIT-PCBA contains 15 targets, 7,844 confirmed active compounds, and 407,381 confirmed inactive compounds.The dataset was collected from the PubChem BioAssay dataset.
  • Results: FP-GNN achieved the best performance on six targets, compared with two targets each for NB, DNN, and GCN.FP-GNN led on ADRB2, ALDH1, ESR1_ago, MAPK1, PPARG, and TP53; the other targets were led by multiple baseline models.
  • Results: FP-GNN showed similar outstanding performance to models using mixed fingerprints and outperformed fingerprint-based models while matching or exceeding GCN and GAT.ROC-AUC was used for classification evaluation, and direct baseline comparisons were reported in Supplementary Table 5.
  • Implication: The results support FP-GNN as a competitive approach for predicting molecular biological activity in drug-discovery campaigns.This conclusion was reported specifically for the challenging LIT-PCBA dataset.

based models on cell-based phenotypic screening datasets

Across 14 breast-cell-line phenotypic screening datasets, FP-GNN was the strongest overall model and combined graph and fingerprint information. Its component analyses indicate that both pathways contributed substantially, although a fixed module ratio may limit performance on some targets.

  • Phenotypic screening results: FP-GNN achieved the best performance on eight of 14 breast-cell-line datasets.Attentive FP was best on three cell lines, XGBoost on two, and GCN on one.
  • Phenotypic screening results: FP-GNN was second-best on HS-578T, MDA-MB-231, and MDA-MB-468.These were among the cell lines where another model achieved the top performance.
  • Phenotypic screening results: 0.849 was FP-GNN’s highest average AUC across the 14 cell lines, the best overall result among the compared models.The comparison included Attentive FP, GCN, GAT, MPNN, and XGBoost.
  • Ablation and pathway contributions: In optimized models across the 13 public datasets, 54.3% of GNN ratios fell between 0.4 and 0.6, indicating relatively balanced pathway contributions.Pure GNN and pure FPN models accounted for approximately 4.3% of all models.
  • Ablation and pathway contributions: On LIT-PCBA, FP-GNN outperformed separate FPN and GNN models on 10 of 15 targets.It also exceeded FPN on most of the other five targets, while performing slightly below GNN there.
  • Ablation and pathway contributions: The combined model integrates local-neighbor and complete structural information from graphs with substructure and pharmacophore information from fingerprints.The authors link this complementary information to more accurate molecular-property predictions.
  • Ablation and pathway contributions: Using the default GNN ratio of 0.5 may have limited performance on five targets by retaining less information from the less favorable module.This explanation was presented as a possible reason for the observed target-specific pattern.

The influence of different types of fingerprints

FP-GNN performance varied with fingerprint choice: mixed fingerprints generally outperformed Morgan fingerprints across classification and regression datasets, while algorithm type affected which representation captured more information.

  • Mixed-fingerprint FP-GNN models performed better than Morgan-fingerprint FP-GNN models on both classification and regression datasets.This pattern was reported in Figure 5a for classification and Figure 5b for regression.
  • Mixed fingerprints did not show absolute superiority on LIT-PCBA, producing 42 best models versus 30 for Morgan fingerprints, with three results tied.
  • NB and SVM tended to extract more information from Morgan fingerprints, whereas RF, XGBoost, and DNN tended to capture more from mixed fingerprints.
  • The authors attribute mixed fingerprints’ complementarity to their coverage of atomic, bond, structural, substructural, and pharmacophore information.MACCS, PubChem, and Pharmacophore ErG fingerprints provide information that may not be included in molecular-graph features.
  • Morgan fingerprints record local atom-environment information, which may overlap with molecular-graph features, whereas mixed fingerprints can better complement those features.The Morgan fingerprint used in the architecture was ECFP-4 with 1024 bits.

The Anti-noise Ability of FP-GNN

FP-GNN was evaluated under artificially noisy labels and in interpretability analyses. It achieved state-of-the-art anti-noise performance, while attention patterns and fingerprint-bit analyses linked predictions to molecular fragments and fingerprint information.

  • Anti-noise evaluation: The anti-noise experiments artificially changed training and validation labels while leaving the test-set labels unchanged.The HIV dataset contained 41,127 compounds and was split 8:1:1 into training, validation, and test sets.
  • Anti-noise evaluation: FP-GNN achieved state-of-the-art performance in anti-noise tests against Attentive FP, HRGCN+, and XGBoost.The comparison used the same data, split, ROC-AUC metric, and noise rates as Wu et al. for fairness.
  • Molecular interpretation: FP-GNN attention coefficients quantify the contribution of chemical fragments by mapping adjacent-atom attentions to connected bonds.
  • Molecular interpretation: In an inactive molecule, FP-GNN emphasized a hydrophilic amino-group fragment associated with difficulty crossing the BBB.The red fragment had ClogP = -0.905, while the grey fragment had ClogP = 0.934.
  • Fingerprint interpretation: Fingerprint-bit analysis on FreeSolv found that the three fingerprint types jointly contributed important predictive information.Four of the ten most crucial bits came from MACCS FP, three from Pharmacophore ErG FP, and three from PubChem FP.

Conclusions

FP-GNN couples molecular graphs with mixed molecular fingerprints and showed strong performance across diverse molecular-property datasets. Its analyses also indicate complementary representations, robustness to noise, and interpretable chemical-fragment importance, while future work targets limited biological-data quantity and quality.

  • Architecture: FP-GNN couples graph-attention representations of molecular graphs with neural representations of mixed molecular fingerprints.The architecture combines information from both representation types to generate a more comprehensive molecular representation.
  • Benchmark performance: FP-GNN performed outstandingly against four recently published graph-based deep-learning algorithms and XGBoost across 13 public datasets.The evaluated graph-based baselines included MoleculeNet, Chemprop, Attentive FP, and HRGCN+.
  • Benchmark performance: FP-GNN also showed strong predictive performance on the unbiased LIT-PCBA dataset and 14 phenotypic drug-screening datasets related to breast-cancer cell lines.These evaluations extended beyond the classical public benchmarks.
  • Representation analysis: Ablation experiments found that molecular graphs and mixed fingerprints contributed to improved prediction performance, while different fingerprints affected predictive performance.The results support complementary information from the two representation types and a dependence on fingerprint choice.
  • Representation analysis: The three-fingerprint mixture of atomic and bond properties, substructures, and pharmacophores achieved optimal complementarity with the graph-based module.This mixture was identified as the current optimal fingerprint combination in the reported analyses.
  • Practical implications: FP-GNN showed anti-noise ability and intuitive interpretability by identifying important chemical fragments for molecular design and optimization.The authors connect these properties to handling noisy drug-discovery data and supporting molecules with desired properties or functions.
  • Future directions: Future optimization may use pre-training because biological datasets can have insufficient quantity and poor quality, and may incorporate protein-target information.The proposed extensions include pre-training on larger datasets and combining molecular and protein-target features.

Key points

The paper presents FP-GNN as a molecular-property predictor that combines graph and fingerprint information. Reported experiments characterize it as highly competitive, with complementary representations and interpretable fragment-level outputs.

  • Architecture: FP-GNN couples graph-attention networks based on molecular graphs with neural networks based on mixed molecular fingerprints.Together, the two paths generate a more comprehensive molecular representation.
  • Performance: Extensive experiments found FP-GNN highly competitive with classic machine-learning and state-of-the-art deep-learning methods.The comparison covered graph-based and fingerprint-based model families.
  • Representation analysis: Ablation experiments indicated that information from molecular graphs and molecular fingerprints is complementary for improving predictive power.The study evaluated graph models and models using Morgan or mixed fingerprints as molecular representations.
  • Interpretability: The model's intuitive interpretability can provide important chemical fragments to assist chemists and pharmacists in designing or optimizing molecules.In the BBBP examples, darker regions corresponded to more important structures, and captured portions were consistent with prediction results.
  • Interpretability: For BBBP examples, the highlighted molecular portions aligned with ClogP-related permeability interpretations and were consistent with the model's predictions.One molecule was described as permeable with higher-C logP-associated lipophilicity in darker regions, whereas another was impermeable with weaker lipophilicity.

Table legends

The legends cover FP-GNN’s graph-attention and fingerprint components, benchmark comparisons, ablations, robustness, interpretability, and predictive-performance tables.

  • Architecture: FP-GNN combines molecular graph information with fingerprints for molecular-property prediction.
  • Fingerprint comparisons: The benchmark legends distinguish Morgan-based models from models using mixed fingerprints, including MACCS FP, PubChem FP, and Pharmacophore ErG FP.
  • Benchmark evaluation: The figures evaluate FP-GNN on public datasets, including comparisons with graph-based and conventional machine-learning baselines.
  • Robustness and ablations: The legends include an ablation study, fingerprint-composition analyses, and anti-noise experiments using multiple models on the HIV dataset.
  • Interpretability: Interpretability analyses examine molecular-structure importance, with darker coloring indicating greater structural importance and examples drawn from the BBBP dataset.
  • Reported results: Table 1 reports predictive performance on 13 public datasets, while additional tables cover breast cell-line datasets and significant mixed-fingerprint bits.
Loading 2205.03834v1…