Leveraging Machine Learning, Transformers, and Large-Scale Chemical Data to Predict TGF-β1 Inhibitors for Cardiac Fibrosis Treatment

Fibrotic diseases are dangerous conditions that cause the deaths of more than 800,000 people each year. One very prevalent type of fibrotic disease is cardiac fibrosis, which is also classified as a cardiovascular disease (CVD). Cardiac fibrosis is a dangerous disease caused by the overproduction of fibroblasts, crucial cells that play large roles in creating proteins and healing wounds. These cells produce extracellular matrices (ECMs), connective tissues whose over-accumulation can impair the function of the myocardium. One way to prevent the accumulation of ECM in the myocardium is by inhibiting the function of proteins such as transforming growth factorbeta receptor type 1 (TGF-β1), whose overexpression results in the overproduction of fibroblasts. Unfortunately, there are very few drugs currently in circulation that attempt to inhibit the function of TGF-β1. Past studies attempting to discover new inhibitory drugs, whether in vivo, in vitro, or in silico, have often revolved around modeling the heart and have encountered many limitations including a lack of complexity and applicability. Instead, this study focuses on using available chemical data to train traditional and transformer-based machine learning predictors. We tested the hypothesis that the efficacy of potential TGF-β1 inhibitors can be assessed through the use of traditional and natural language processing regressors solely using the SMILES notation and Lipinski descriptors of chemical compounds. These models achieved R2 scores up to 0.64 and showed immense promise as potential tools in the process of cardiovascular drug discovery.