Grant: Discovery Grants Program - Individual

De novo drug discovery is expensive, time--consuming, and includes high risk investments. The repurposing of approved drugs potentially provide more cost--effective treatments in less time than it would take to develop new drugs. The explosive and large--scale growth of molecular, omics, semantic, bio-molecular network, and expression data, and phenotypic data of pharmacological compounds has enabled the development of computational drug repurposing (CDR) approaches that use machine learning (ML) to discover new indications for given approved drugs [drug--centric CDR], or identify effective drugs as a potential treatments for given diseases [disease--centric CDR]. ML approaches to CDR from such heterogeneous data has many limitations due to lack of (i) appropriate integration of homogeneous and heterogeneous data, (ii) of balanced data, or (iii) of model explainability/intelligibility. The performances of current network--based CDR (NB-CDR) methods proposed to integrate heterogeneous data are hindered by the high level of noise, sparseness, and incompleteness of current bio-molecular networks. Graph neural networks (GNNs) are recently proposed to deal with graph--structured data, but they naively conflate the known drug-disease, drug--drug, and disease--disease networks into a single heterogeneous graph and cause a substantial loss of domain--specific information. My long-term goal is to devise deep graph representation learning (GRL) approaches for learning compact low-dimensional representations of biomarkers, drugs, and diseases from multiple-types and multiple-sources of omics data while applying them to various (i) molecular property prediction (MPP) problems and (ii) to bio-molecular association prediction problems in general, including computational drug repurposing (CDR), adverse drug side-effect (ADE) prediction, drug-target interaction (DTI), and poly-pharmacy tasks (eg, drug-drug interaction [DDI] prediction, and effective drug combination [EDC] prediction). We explore the use of large language model (LLM) based embeddings (ie, representations) or predictions such as BERT, GPT, and LLaMA to enhance or improve association prediction tasks. My short-term goals are as follows. O1: Devising GRL methods for learning and integrating low-dimensional representations of biomarkers, drugs, and diseases by integrating heterogeneous data into one or more heterogeneous multi-relational knowledge graphs that incorporate domain-specific information of biomarkers, drugs and diseases, and capture complex topological patterns across multiple network types from multiple sources. O2: Devising LLM-based embedding and prediction techniques in combination with network-based embedding methods (graph neural network [GNN] methods, in particular).