Source-linked AI summary
Chemi-net: a graph convolutional network for accurate drug property prediction
Ke Liu, Xiangyan Sun, Lei Jia, Jun Ma, Haoming Xing, Junqiu Wu, Hua Gao, Yax Sun, Florian Boulnois, Jie Fan
TL;DR
ADME prediction requires methods that can represent heterogeneous molecular graphs rather than relying solely on fixed-size descriptors or fingerprints. Chemi-Net combines molecular graph convolution with multi-task deep learning and was evaluated against Cubist. The study reports higher predictive accuracy for Chemi-Net across the tested datasets and presents the approach as a proof of concept.
Problem
ADME prediction is challenged by heterogeneous molecular graph structures that do not match the homogeneous inputs expected by many learning algorithms.
Method
Chemi-Net combines molecular graph convolutional networks with multi-task deep neural learning, dynamic batching, and fine-tuning for ADME property prediction.
Results
Chemi-Net improved predictive accuracy compared with Cubist, with higher R2 values and higher accuracy across all data sets.
Takeaways & Limitations
The proof-of-concept study reports the first use of a molecular graph model in this application and finds significant performance advantages over Cubist-based methods.
Takeaways & Limitations
The proof-of-concept study is limited by the fewer data points in an individual dataset.
Abstract
from arXiv · showhide
Absorption, distribution, metabolism, and excretion (ADME) studies are critical for drug discovery. Conventionally, these tasks, together with other chemical property predictions, rely on domain-specific feature descriptors, or fingerprints. Following the recent success of neural networks, we developed Chemi-Net, a completely data-driven, domain knowledge-free, deep learning method for ADME property prediction. To compare the relative performance of Chemi-Net with Cubist, one of the popular machine learning programs used by Amgen, a large-scale ADME property prediction study was performed on-site at Amgen. The results showed that our deep neural network method improved current methods by a large margin. We foresee that the significantly increased accuracy of ADME prediction seen with Chemi-Net over Cubist will greatly accelerate drug discovery.
Introduction
ADME prediction must handle heterogeneous molecular graph structures that do not fit the fixed-size inputs expected by many learning algorithms. Chemi-Net addresses this challenge by combining molecular graph convolution with multi-task deep learning and evaluates the approach against Cubist in a large-scale study.
- Introduction: Molecular conformations are represented as heterogeneous graph structures, creating a challenge for algorithms that expect homogeneous input features.Unlike image data, molecular graphs vary in structure among molecules.
- Introduction: Traditional approaches transform molecular graphs into fixed-size descriptors or fingerprints that existing machine-learning algorithms can process.Fingerprint representations encode substructures as fixed-length bit strings.
- Introduction: Fingerprint-based representations can be very large, sparse, and noisy, limiting their performance.These limitations motivate methods that learn smaller, denser molecular representations.
- Introduction: Multi-task deep neural networks can exploit information across activity datasets and outperform traditional single-task methods when datasets share connections and non-random activity patterns.Prior work linked this advantage to borrowing signal from structurally similar molecules across tasks.
- Introduction: Chemi-Net combines molecular graph convolutional networks with multi-task learning, dynamic batching, and fine-tuning to improve ADME prediction stability and accuracy.The study describes this combination as a new application and tests it across five ADME tasks with over 250,000 data points.
- Introduction: Chemi-Net showed significant performance advantages over existing Cubist-based methods in the large-scale ADME prediction test.The evaluation was conducted once on the testing dataset in a restricted environment.
Results and Discussion
Chemi-Net’s MT-DNN improves R2 prediction accuracy over Cubist across the evaluated ADME datasets, with gains influenced by dataset quality and training–test similarity. Performance also varies with descriptor choice, endpoint, and dataset size.
- MT-DNN improves predictive accuracy compared with Cubist across the evaluated ADME datasets.
- Model and dataset effects: Larger and less noisy datasets produced higher single-task improvements, and MT-DNN provided additional accuracy improvement.
- ADME prediction accuracy: ST-DNN underperformed Cubist on some lower-quality solubility subsets, whereas MT-DNN achieved higher accuracy versus Cubist for all datasets.
- Compound similarity: For all 13 datasets, average within-training similarity was 0.878, compared with 0.679 between training and test compounds.
- Compound similarity: R2 increased with training–test similarity for both Chemi-Net and Cubist, supporting the hypothesis that similarity affects prediction performance.
- Descriptor comparison: Chemi-Net descriptors performed better than Amgen descriptors on large, relatively high-quality datasets, while Amgen descriptors performed better on small and noisy datasets such as PXR and bioavailability.
Conclusion
Chemi-Net combines molecular graph convolution with multi-task deep learning to predict drug properties in industrial-grade datasets. Its feature extraction and multi-task design improve performance across assay types and support wider application in drug discovery.
- Conclusion: Chemi-Net applies a molecular graph convolutional network with MT-DNN to predict drug properties in industrial-grade datasets.The approach is reported as a first use of this combination in the study.
- Conclusion: Graph convolution extracts compound features from SMILES files instead of relying on preset descriptors.
- Conclusion: Multi-task DNN improves on individual models limited by fewer data points in single datasets.
- Conclusion: Performance improvements across all assay types motivate wider application of the approach in drug discovery tasks.
Methods
Chemi-Net represents molecules as atom-and-pair feature graphs, applies shared-weight graph convolutions and pooling, and predicts ADME properties with deep neural networks. The methods also include joint multi-task learning and fine-tuning ensembles.
- Molecular representation: Chemi-Net represents each molecule with distributed atom and atom-pair features, including atom types, radii, aromaticity, distances, and bond orders.
- Graph convolution: The assembled features form a molecule-shaped graph processed by graph convolution operators that transform neighbor information and reduce variable-sized inputs.
- Graph convolution: Shared weights across atoms provide position-invariant convolution, while dynamic filter connections adapt to molecular structure rather than remaining fixed.
- Prediction: Pooling converts variable numbers of atom vectors into a fixed-size molecule embedding, which fully connected layers use to predict an ADME property.
- Convolution filter: Neighbor features are transformed and aggregated with commutative reductions, combining sum, maximum, and average operations before producing atom representations.
- Multi-task learning and fine-tuning: The multi-task model shares ligand embeddings across tasks, sums task losses with optional non-uniform weights, and fine-tuning ensembles selected models to improve accuracy.
Competing interests
The authors report employment at Amgen Inc. and Accutar Bio Inc.
- L.J., H.G., Y.S., and F. B. are employed at Amgen Inc., while K.L., X.S., L.J., J.W., and J.F. are employed at Accutar Bio Inc.
Figure legends
The figure legends and tabulated results identify benchmark comparisons between MT-DNN, ST-DNN, Chemi-Net, and Cubist across ADME endpoints and dataset subsets.
- The reported comparisons include MT-DNN and Cubist for solubility, and MT-DNN and Cubist for PXR inhibition.
- The network figure uses red for the central atom, pink for neighboring atoms, and purple for information flow from the previous convolution layer.
- The tabulated rows report train size, test size, and multiple result values for the listed subsets.
Supplementary information
Chemi-Net requires dynamic batching because molecule-dependent graphs differ across inputs, unlike the static graphs used by conventional CNN implementations. Its algorithm analyzes and merges compatible operations while preserving computation correctness.
- Motivation: Conventional mini-batching assumes a computation graph that is static and shared across inputs, but molecule-dependent graphs are dynamic and heterogeneous.
- Dynamic batching: Chemi-Net addresses this issue with an efficient dynamic batching algorithm that analyzes each input batch and merges CNN operations without affecting correctness.
- Dynamic batching: The approach batches neural-network operations within and across instances, supporting batch normalization for heterogeneous computation graphs.
- Algorithm steps: The algorithm comprises computation-graph initialization, batching analysis, and gather/scatter generation, with compatible dependency-free operations batched together.
- Algorithm steps: A gather/scatter optimization can eliminate scatter-gather pairs between consecutive batched operations.