GraphDTA: Predicting drug-target binding affinity with graph neural networks

Thin Nguyen; Hang Le; Thomas P. Quinn; Tri Nguyen; Thuc Duy Le; Svetha Venkatesh

doi:10.1093/bioinformatics/btaa921

BioinformaticsOct 15, 2020BioinformaticsPeer-reviewed journal article

GraphDTA: Predicting drug-target binding affinity with graph neural networks

Thin Nguyen iD, Hang Le, Thomas P. Quinn iD, Tri Nguyen, Thuc Duy Le iD, Svetha Venkatesh iD

PDF DOI Preprint Code Data

In-depth full-text digest available · Jump to digest ↓

Abstract

The development of new drugs is costly, time consuming, and often accompanied with safety issues. Drug repurposing can avoid the expensive and lengthy process of drug development by finding new uses for already approved drugs. In order to repurpose drugs effectively, it is useful to know which proteins are targeted by which drugs. Computational models that estimate the interaction strength of new drug--target pairs have the potential to expedite drug repurposing. Several models have been proposed for this task. However, these models represent the drugs as strings, which is not a natural way to represent molecules. We propose a new model called GraphDTA that represents drugs as graphs and uses graph neural networks to predict drug--target affinity. We show that graph neural networks not only predict drug--target affinity better than non-deep learning models, but also outperform competing deep learning methods. Our results confirm that deep learning models are appropriate for drug--target binding affinity prediction, and that representing drugs as graphs can lead to further improvements. Availability of data and materialsThe proposed models are implemented in Python. Related data, pre-trained models, and source code are publicly available at https://github.com/thinng/GraphDTA. All scripts and data needed to reproduce the post-hoc statistical analysis are available from https://doi.org/10.5281/zenodo.3603523. ContactThin.Nguyen@deakin.edu.au

In-depth summary

Claude Sonnet 4.6 · xhigh effort · 8 Nov 2026 · from the full text of version 8 (2 Oct 2020) · CC BY 4.0

Representing drugs as molecular graphs, GraphDTA lowers drug-target affinity prediction error to MSE 0.229 on Davis and 0.139 on Kiba versus deep learning baselines.

Earlier deep models for drug-target affinity (DeepDTA, WideDTA) encode drugs as strings, which discards molecular structure. Graph neural networks had not been applied to continuous affinity regression, and this work tests them against those baselines.

Key results

ShownOn Davis, all four graph variants had lower MSE than every baseline. Concordance index (CI) improved for only two of the four variants (GAT and GIN).
Best MSE 0.229 vs best baseline 0.261 (14.0% lower). CI 0.892 (GAT) and 0.893 (GIN) vs best baseline CI 0.886. 72 drugs, 442 targets. · 3.1 Graphical models outperform the…; 2.5 Benchmark
ShownOn Kiba, three of four variants (GIN, GCN, GAT-GCN) had both the lowest MSE and the highest CI. GIN was the only variant best on both datasets and both metrics.
Best MSE 0.139 vs best baseline 0.179 (28.8% lower). 2,116 drugs, 229 targets. Baselines' numbers are those originally published, not re-run. · 3.1 Graphical models outperform the…; 2.5 Benchmark
SuggestedThe 128 GIN latent variables partly encode known chemistry. A minority of their variance is explained by standard descriptors, led by aliphatic OH group count.
Redundancy analysis with 38 JoeLib descriptors: 20.19% of latent space explained. Hidden nodes V58 and V14 activate strongly with more aliphatic OH groups. GIN on Kiba only. · 3.2 Graphical models discover known drug…; 2.6 Model interpretation
SuggestedPrediction error is concentrated in a few drugs and proteins. Hard-to-predict drugs sit near the origin of a latent-space PCA, while easy ones are more often PCA outliers.
CHEMBL1779202 (ALK inhibitor), CHEMBL1765740 (PDK1 inhibitor) and protein CSNK1E each had median absolute error above 2. No other pattern separated hard from easy drugs. · 3.3 A few drugs contribute disproportionately to…; 2.6 Model interpretation

Methods, data and reproducibility

Drugs are converted from SMILES to atom-feature graphs and encoded by GCN, GAT, GIN or GAT-GCN with global pooling. Proteins are integer-encoded, embedded and passed through 1D CNNs. The two vectors are concatenated and regressed to affinity by fully connected layers.

Davis benchmark, Kd binding affinities (range 5.0 to 10.8), same train/test examples as the earlier benchmark studies · n = 72 drugs, 442 targets
Kiba benchmark, KIBA scores (range 0.0 to 17.2), same split as the baseline studies · n = 2,116 drugs, 229 targets
Protein sequences retrieved from UniProt by gene name, truncated or padded to 1000 residues
JoeLib molecular descriptors from ChemMine Tools used for latent-space redundancy analysis · n = 38 descriptors

Methods & tools: GraphDTA (GCN, GAT, GIN, GAT-GCN variants), RDKit, DeepChem atom features, UniProt, ChemMine Tools (JoeLib descriptors), Redundancy analysis, Principal component analysis, DeepDTA, WideDTA
Code: https://github.com/thinng/GraphDTA
Data & accessions: https://doi.org/10.5281/zenodo.3603523 https://github.com/thinng/GraphDTA/blob/master/supplement.pdf
Software & environment: Implemented in Python; RDKit used for SMILES-to-graph conversion. No package versions or containers stated.
Reproducibility gaps: Hardware, runtime and software versions are not stated. Hyper-parameters are given in Table 3, which is not reproduced in the provided text.

Limitations (5)

The model ignores molecular stereochemistry, and proteins are encoded only as 1D sequences. Whether 3D drug or protein representations would improve accuracy was not tested. · stated by the authors
Most learned latent variables remain unexplained by the available descriptors, so much of what the model learned is uninterpreted. The authors note the unique-substructure pattern does not hold cleanly, since many easy drugs also lack unique structures. · stated by the authors
The authors could not find an obvious latent-space pattern separating hard-to-predict from easy-to-predict drugs; the only trend was that easy drugs more often appear as PCA outliers. · stated by the authors
Baseline scores were copied from earlier publications rather than re-run, and hyper-parameters were fixed a priori without tuning. No variance across seeds or splits is reported, so the MSE and CI gaps carry no uncertainty estimates. · noted in this summary
Latent-space analyses (descriptor regression and PCA) use only the GIN model on Kiba, so those findings may not transfer to other variants or to Davis. · noted in this summary

Read this if you want a baseline for graph-based drug-target affinity regression on Davis and Kiba, or an example of probing GNN latent spaces with descriptors.

Citations: 1,253
FWCI: 41.50
Field percentile: top 1%