Deep Graph Learning for Analyzing Interactions in Chemical and Biological Systems

Abstract

Chemical and biological systems feature complex interactions. In chemistry, the concept of nucleophile-electrophile interactions is fundamental to understanding reaction mechanisms. Similarly, in multicellular organisms, cell-cell interactions, often mediated by ligand-receptor interactions, play a critical role in coordinating various biological processes. Exploring these interactions is crucial for understanding molecular synthesis and cellular communication, key aspects of drug discovery. However, the complexity of molecular and biological structures poses significant challenges. During my Ph.D. studies, I employed innovative deep graph learning methods to tackle these challenges, focusing on nucleophile-electrophile, cell-cell, and ligand-receptor interactions.

Nucleophilicity and electrophilicity are two key physicochemical properties that assist chemists in estimating the reaction rates between nucleophiles and electrophiles. However, determining the nucleophilicity parameter traditionally requires labor-intensive experiments. Recent advancements leveraged traditional machine learning models for predicting nucleophilicity, which typically depend on manually crafted structural descriptors. We propose a graph neural network (GNN) model that integrates molecular structural and electronic features, calculated using Density Functional Theory, to achieve state-of-the-art performance in predicting nucleophilicity. The model effectively identifies different nucleophilic atoms within a nucleophile by analyzing their local chemical environments. It maintains high accuracy with new types of nucleophiles and shows stable performance across various datasets, indicating robustness to coarse molecular conformations.

Recent advances in spatial transcriptomics (ST) combine gene expression data with cell location information. This integration allows us to transform tissue samples into cell graphs that represent cell-cell interactions (CCIs), facilitating the use of deep graph learning to derive low-dimensional embeddings for downstream analyses. However, existing methods rely solely on spatial proximity to define graph structures, which may not fully capture the diversity of CCIs. We propose a novel framework based on graph structure learning, STAGUE, which addresses this limitation by learning both cell graph structures and embeddings directly from ST data. Cell embeddings enable various downstream analyses like spatial clustering and trajectory inference, and novel CCIs can be inferred from the learned cell graph structure. Experimental results demonstrate that STAGUE outperforms existing methods in clustering performance across 86 real and simulated ST datasets. It also provides insights into human breast cancer tissue heterogeneity. Furthermore, STAGUE identifies CCIs that align more closely with established biological knowledge and pinpoints key regulatory genes, offering deeper understanding of the biological processes involved.

Ligand-receptor interactions (LRIs) are pivotal in facilitating cell-cell communications (CCCs), a specific type of CCIs used to relay signals. Existing approaches for deciphering CCCs depend heavily on established LRI databases, which are frequently incomplete, thereby limiting the precision and sensitivity of CCC inference. We introduce AlphaTalk, a new method that enhances CCC analysis by incorporating quasi-LRIs identified using advanced techniques including deep graph learning, network-based algorithms, and random walk. AlphaTalk identifies quasi-LRIs with a high likelihood of interaction, incorporating data from protein structures and LRI network topologies. These identifications are further refined to confirm spatial proximity and their potential to initiate intracellular signaling pathways. Applied to six spatial transcriptomics datasets, AlphaTalk significantly enhances the reconstruction accuracy of spatial structures at the single-cell level and identifies spatially varying correlations that influence immune responses in the tumor environment. AlphaTalk also reveals unique communication patterns linked to specific signaling pathways in human breast cancer, providing insights into cellular dynamics during disease progression.

Date of Award	28 Feb 2025
Original language	English
Awarding Institution	City University of Hong Kong
Supervisor	Yao Fu (External Supervisor) & Shuaicheng LI (Supervisor)