LINT: LLM Interaction Network for Clinical Trial Outcome Prediction

Chufan Gao, Tianfan Fu, Jimeng Sun

Keywords: Clinical Trial Outcome Prediction, Healthcare, Clinical Trial, Drug Discovery

Abstract:

Clinical trial outcome prediction aims to predict the success probability of a clinical trial that reaches its desirable endpoint. Most of
the effort focuses on developing machine learning models for making accurate predictions with diverse data sources, including clinical
trial descriptions, drug molecules, and target disease conditions. Accurate trial outcome prediction helps trial planning and asset portfolio prioritization. Previous works have focused on small-molecule drugs; however, biologics are a quickly growing intervention type
that lacks information that is traditionally known for drugs, like molecular properties. Additionally, traditional methods like graph
neural networks are much more difficult to apply to biologics data which are a fast-growing type of drug. To address these points, we
propose a Language Interaction Network (LINT), a novel method for trial outcome prediction using only free-text descriptions. We validate the effectiveness of LINT with thorough experiments across three trial phases. Specifically, LINT obtains 0.770, 0.740, and 0.748 ROC-AUC scores on phase I, II, and III, respectively, for clinical trials with biologic interventions.

Submission Number: 3