MT-RGCN-DTI: A Multi-Task Relational Graph Convolutional Network for Drug-Target Interaction Prediction
Authors: Dheeran Sankaran, R Yadunandan, Daniele Santoni, Nimisha Ghosh
Organizations: Department of Computer Science and Engineering, Shiv Nadar University Chennai, Tamil Nadu, India · Institute for System Analysis and Computer Science “Antonio Ruberti”, National Research Council of Italy, Rome, Italy
Prediction of drug-target interaction (DTI) has become a very crucial part of drug discovery. With the increase in the data volume, it has become very important to improve prediction accuracy by integrating multiple data sources. In this regard, utilising multiple resources in the form of heterogeneous information network to address DTI prediction can be rewarding. In this work, we propose a heterogeneous information network based multi-task DTI prediction model which leverages the advantages of relation graph convolutional network (RGCN). To achieve this, the model employs drug-target interaction network, drug-drug similarity network and protein-protein similarity network to construct a heterogeneous information network. Moreover, we use a knowledge graph embedding scoring functions to model interactions between drug and targets, specifically DistMult and ComplEx as decoders over the learned node representations. We performed extensive experiments on multiple benchmark datasets and compared with state-of-the-art methods. We observed that the proposed model shows competitive performance, thereby confirming the effectiveness of our approach. The codes and datasets are available at https://github.com/Dheeran-S/MT-RGCN-DTI/.
Figures & tables
Dataset
Task
Drug
Protein
Interactions
EN
Classification
445
663
2925
IC
Classification
210
204
1476
DrugBank
Classification
1555
1593
5837
Davis
Regression
68
379
25772
KIBA
Regression
2068
229
117657
TABLE I: Summary of Data
Fig. 1: Pipeline of the work
Dataset
Method
Acc
Prec
Sen
Spec
MCC
AUC
AUPR
Enzyme
HIN-MTDTI
0.916 ± 0.016
0.970 ± 0.007
0.859 ± 0.029
0.974 ± 0.006
0.838 ± 0.029
0.965 ± 0.007
0.973 ± 0.005
MT-RGCN-DTI-DM
0.914 ± 0.008
0.920 ± 0.012
0.906 ± 0.020
0.921 ± 0.014
0.827 ± 0.017
0.968 ± 0.006
0.972 ± 0.005
MT-RGCN-DTI-CX
0.919 ± 0.004
0.936 ± 0.008
0.899 ± 0.015
0.939 ± 0.010
0.839 ± 0.008
0.974 ± 0.005
0.978 ± 0.004
IC
HIN-MTDTI
0.989 ± 0.001
0.949 ± 0.016
0.782 ± 0.010
0.998 ± 0.000
0.834 ± 0.017
0.912 ± 0.010
0.807 ± 0.017
MT-RGCN-DTI-DM
0.989 ± 0.002
0.926 ± 0.016
0.746 ± 0.031
0.998 ± 0.001
0.826 ± 0.021
0.974 ± 0.005
0.863 ± 0.023
MT-RGCN-DTI-CX
0.990 ± 0.003
0.903 ± 0.024
0.785 ± 0.052
0.997 ± 0.001
0.836 ± 0.039
0.973 ± 0.008
0.877 ± 0.031
TABLE II: Performance comparison of MT-RGCN-DTI with existing state-of-the-art across Enzyme, IC and DrugBank datasets for classification tasks
Dataset
Method
MSE
CI
Davis
HIN-MTDTI
0.228 ± 0.270
0.623 ± 0.007
MT-RGCN-DTI-DM
0.226 ± 0.016
0.876 ± 0.005
MT-RGCN-DTI-CX
0.213 ± 0.014
0.882 ± 0.003
KIBA
HIN-MTDTI
0.317 ± 0.417
0.572 ± 0.011
MT-RGCN-DTI-DM
0.316 ± 0.010
0.796 ± 0.003
MT-RGCN-DTI-CX
0.379 ± 0.198
0.755 ± 0.112
TABLE III: Performance comparison of MT-RGCN-DTI with existing methods across Davis and KIBA datasets for regression tasks
Dataset
Model
MT-RGCN-DTI-DM
MT-RGCN-DTI-CX
Enzyme
DTI-only
0.962 ± 0.009
0.967 ± 0.005
DTI + DDI
0.964 ± 0.009
0.970 ± 0.007
DTI + PPI
0.967 ± 0.006
0.971 ± 0.007
DTI + DDI ( λ =0.5) + PPI ( λ =0.5)
0.966 ± 0.006
0.972 ± 0.005
DTI + DDI ( λ =1.0) + PPI ( λ =1.0)
0.972 ± 0.005
0.971 ± 0.005
Learned- λ
0.969 ± 0.007
0.973 ± 0.005
TABLE IV: Ablation Study of MT-RGCN-DTI in terms of AUC across Enzyme, IC and DrugBank datasets for classification tasks
Fig. 2: Ablation Study of MT-RGCN-DTI for Regression Tasks
Dataset
Held-out pairs
Positive /
Linear-probe
PC 1 / PC 2 var.
Negative
AUROC
(%)
Enzyme
1,170
585/585
0.981
14.5/9.9
IC
8,568
294/8,274
0.980
10.2/7.3
DrugBank
2,334
1,167/1,167
0.909
17.4/11.2
TABLE V: Held-out fold-5 evaluation underlying Fig. 3 . AUROC is from a class-balanced logistic-regression linear probe on the 128 -d ComplEx interaction embedding ϕ (trained on training folds, evaluated on fold 5). PC 1/PC 2 variance refers to the two PCA components shown in the scatter plots.
Fig. 3: PCA projection of the held-out fold-5 ComplEx interaction embeddings for RGCN-CX on the (a) Enzyme, (b) IC, and (c) DrugBank datasets. Each point is one held-out drug–target pair: orange = true DTI, blue = non-interaction. Shaded contours are the KDE of the positive class; the top and right strips are class-conditional marginal KDEs of PC 1 and PC 2.