A Graph Convolutional Network–Based Method for Chemical-Protein Interaction Extraction: Algorithm Development (Preprint)

Abstract

BACKGROUND Extracting the interactions between chemicals and proteins from the biomedical literature is important for many biomedical tasks such as drug discovery, medicine precision, and knowledge graph construction. Several computational methods have been proposed for automatic chemical-protein interaction (CPI) extraction. However, the majority of these proposed models cannot effectively learn semantic and syntactic information from complex sentences in biomedical texts. OBJECTIVE To relieve this problem, we propose a method to effectively encode syntactic information from long text for CPI extraction. METHODS Since syntactic information can be captured from dependency graphs, graph convolutional networks (GCNs) have recently drawn increasing attention in natural language processing. To investigate the performance of a GCN on CPI extraction, this paper proposes a novel GCN-based model. The model can effectively capture sequential information and long-range syntactic relations between words by using the dependency structure of input sentences. RESULTS We evaluated our model on the ChemProt corpus released by BioCreative VI; it achieved an F-score of 65.17%, which is 1.07% higher than that of the state-of-the-art system proposed by Peng et al. As indicated by the significance test (<i>P</i><.001), the improvement is significant. It indicates that our model is effective in extracting CPIs. The GCN-based model can better capture the semantic and syntactic information of the sentence compared to other models, therefore alleviating the problems associated with the complexity of biomedical literature. CONCLUSIONS Our model can obtain more information from the dependency graph than previously proposed models. Experimental results suggest that it is competitive to state-of-the-art methods and significantly outperforms other methods on the ChemProt corpus, which is the benchmark data set for CPI extraction.

A Message Passing Perspective on Learning Dynamics of Contrastive Learning

“…For a feature matrix H (t) ∈ R n×m at the t-th layer, the canonical graph convolution scheme in GCN (Kipf & Welling, 2017) gives H (t+1) = ĀH (t) . Later, DGC (Wang et al, 2021b) generalizes it to H (t+1) = (1 − ∆t)I + ∆t ĀH (t) with a flexible step size ∆t for finite difference, which reduces the canonical form when ∆t = 1. Comparing it to the alignment update rule in Eq.…”

“…Message Passing Graph Neural Networks (MP-GNNs) are the prevailing designs in modern Graph Neural Networks (GNNs), including numerous variants like GCN (Kipf & Welling, 2017), GAT (Veličković et al, 2018), and even the Transformers (Vaswani et al, 2017). There is a vast literature studying its diffusion dynamics and representation power Oono & Suzuki, 2020;Wang et al, 2021b;Li et al, 2022;Dong et al, 2021;Xu et al, 2019;Chen et al, 2022). Therefore, establishing a connection between contrastive learning (CL) and MP-GNNs will hopefully bring new theoretical and empirical insights for understanding and designing contrastive learning methods.…”