Clinical Trial Outcome Prediction using Deep Learning

BMI707 Final Project (Spring 2022) by Aishwarya Chander, Benedikt Geiger, Kezia Irene, Man Qing Liang, Thomas Smits at the Department of Biomedical Informatics, Harvard Medical School.

Description:

Clinical trials for therapeutics are a time and resource intensive process influenced by several trial, disease and drug properties. We aim to predict clinical trial success rates to accelerate healthcare research and enable the use of new therapeutics that will benefit patients. To this effect, we present a multilayer perception (MLP) model that uses a combination of drugs, drug targets, diseases and trial features to predict the success or failure of a phase III clinical trial. Our MLP model achieved an ROC-AUC of 0.65 and outperformed baseline logistic regression and random forest. Incorporating eligibility criteria, and target data improves model performance, while excluding drug structure embeddings also improved our model's performance, suggesting feature redundancy. Future work should develop better strategies to address data extraction and embedding challenges, analyze feature importance, and shift efforts to predicting different outcomes.

Poster:

GitHub Structure

1. Data Processing: Files used to generate Data Embeddings as described in the final report.
2. MLP Model: Deep learning multilayer perceptron model.
3. Baseline Models: Logistic Regression and Random Forest Model for Baseline comparison.
4. Reports: Project checkins and reports.
5. Images: Image files associated with Repository.

Additional Data Files:

1. Embedding's Data on Google Drive
2. MLP Clinical Trial Model Weights on Google Drive
3. Model Weights Plots on Google Drive