Heterogeneous Entity Representation for Medicinal Synergy Prediction

TL;DR

A novel deep hypergraph learning method named Heterogeneous Entity Representation for MEdicinal Synergy (HERMES) prediction to predict the synergistic effects of anti-cancer drugs demonstrates state-of-the-art performance on two benchmark datasets.

Abstract

Medicinal synergy prediction is a powerful tool in drug discovery and development that harnesses the principles of combination therapy to enhance therapeutic outcomes by improving efficacy, reducing toxicity, and preventing drug resistance. While a myriad of computational methods has emerged for predicting synergistic drug combinations, a large portion of them may overlook the intricate, yet critical relationships between various entities in drug interaction networks, such as drugs, cell lines, and diseases. These relationships are complex and multidimensional, requiring sophisticated modeling to capture nuanced interplay that can significantly influence therapeutic efficacy. We introduce a salient deep hypergraph learning method, namely, Heterogeneous Entity Representation for MEdicinal Synergy prediction (HERMES), to predict anti-cancer drug synergy. HERMES integrates heterogeneous data sources, encompassing drug, cell line, and disease information, to provide a comprehensive understanding of the interactions involved. By leveraging advanced hypergraph neural networks with gated residual mechanisms, HERMES can effectively learn complex relationships/interactions within the data. Our results show HERMES demonstrates state-of-the-art performance, particularly in forecasting new drug combinations, significantly surpassing previous methods. This advancement underscores the potential of HERMES to facilitate more effective and precise drug combination predictions, thereby enhancing the development of novel therapeutic strategies.

Figures (3)

• Figure 1: Overview of HERMES framework. HERMES has three key phases: (1) initialization: acquiring and transforming initial features of drugs, cell lines, and disease indications to a uniform dimensionality; (2) refinement: enhancing feature representations through the construction of a dual-relationship hypergraph and the application of hypergraph neural networks with gated residual connections; (3) consolidation: integrating refined features using a binary classification model to predict drug synergies with high accuracy. Each phase is integral to the framework’s ability to process diverse data types and generate precise drug synergy predictions.
• Figure 2: Model performance comparison for (A) ALMANAC Dataset and (B) ONeil Dataset. The left panels show the AUROC (%), the middle panels present the AUPRC (%), and the right panels display the F1-score (%) for different models across three validation modes (Random, CLine, DrugComb). Red asterisks indicate statistical significance between HERMES and HypergraphSynergy (*** $p$-value $<0.001$; ** $p$-value $<0.01$; * $p$-value $<0.05$; two-sample $t$-test).
• Figure 3: Performance comparison of HERMES and HypergraphSynergy in the DrugSingle and DrugDouble modes using the ALMANAC dataset. Red asterisks indicate statistical significance between HERMES and HypergraphSynergy(*** $p$-value $<0.001$; ** $p$-value $<0.01$; two-sample $t$-test).