Transfer learning of EGFR inhibitors from preclinical models of cancer to neurodegenerative disease using molecular biomarkers

The failure rate for new drugs entering clinics is in excess of 90%, with more than a quarter of drugs failing due to lack of efficacy. Earlier treatments for complex diseases like cancer considered a small number of patient factors and prescribed a fixed treatment regimen, with individual-specific, severe drug side effects and highly-varying outcomes. Recently, personalised treatments have become popular through the discovery and standardised use of genetic markers that differentiate patient response to therapies.

The deluge of data from screening thousands of drugs and profiling thousands of genes have made machine learning (ML) a useful tool for predicting genetic associations to drug response.
Several research teams, including ours, have recently tested a variety of ML approaches for predicting drug response in different cell lines using genomic data, however, prediction accuracy was low whenever predictions were attempted for new disease contexts due to the limited number of samples, such as rarer diseases or for animal models.

By learning from the entire dose-response function and modelling uncertainty using Gaussian Processes, we can extend our existing strategies for predicting drug responses and improve on the transfer learning of drug response across different biological contexts. Simply, transfer learning focuses on storing knowledge from solving one problem in order to help solve a similar but different problem. Through this project, we hope to provide proof of concept for transfer learning applied to drug response prediction from cancer to neurodegenerative disease.

The objectives are to perform a series of in silico experiments by applying machine learning to existing dose-response data to:

Aim 1. Identify drug responses in cancer cell lines that correlate better with MND (Motor Neuron Disease) as a function of genomic features.
Aim 2. Develop a machine learning model to be trained on response curves of cancer cell lines (identified in Aim 1) and few MND cell lines.
Ami 3. Prediction and validation using genomic features that correspond to cancer and MND cells.