Clinical Trial Outcome Prediction Task Overview

Definition: Clinical trial outcome prediction is a machine learning task that aims to forecast the outcome of clinical trials, such as the approval rate of a drug or treatment. It utilizes various clinical trial features, including the drug's molecular structure, disease code representing the medical condition, and eligibility criteria that specify participant selection criteria. This task is formulated as a binary classification problem, where the machine learning model predicts whether a clinical trial will have a positive or negative outcome.

Impact: Clinical trial is the most time and cost-consuming step in the drug discovery process. Optimizing and designing trials with machine learning could drastically lead to the speedup of delivery of life-saving therapeutics to patients. Particularly, they can effectively alert potential fallouts of trials to practitioners by pointing out potential risks, optimizing safety monitoring protocols and ensuring participant well-being. They can also assist in identifying suitable patient populations, optimizing sample sizes, refining inclusion and exclusion criteria, and selecting appropriate endpoints and outcome measures.

Generalization: Machine learning models for clinical trial outcome prediction are expected to demonstrate robust generalization to novel drug molecular structures and rare diseases. This capability enhances the versatility and applicability of machine learning in clinical research, supporting advancements in personalized medicine and treatment discovery. The ability to generalize well to diverse and evolving conditions is crucial for the models to be adaptable and effectively contribute to the field of clinical trials.

Product: All pipelines require clinical trials.

Pipeline: Clinical trial.

Trial Outcome Prediction (TOP)

Dataset Description: TOP [1]consists of 17,538 clinical trials with 13,880 small-molecule drugs and 5,335 diseases. Out of these trials, 9,999 (57.0%) succeeded (i.e., meeting primary endpoints) and 7,539 (43.0%) failed. For each clinical trial, we produce the following four data items: (1) drug molecule information including Simplified Molecular Input Line Entry System (SMILES) strings and molecular graphs for the drug candidates used in the trials; (2) disease information including ICD-10 codes (disease code), disease description, and disease hierarchy in terms of CCS codes (; (3) trial eligibility criteria are in unstructured natural language and contain inclusion and exclusion criteria; and (4) trial outcome information includes a binary indicator of trial success (1) or failure (0), trial phase, start and end date, sponsor, and trial size (i.e., number of participants).

Task Description: Binary Classification. Given clinical trial features including drug molecular structure, disease code, and eligibility criteria (inclusion and exclusion criteria), predict if it would be approved.

Dataset Statistics: Phase I: 2,402 trials / Phase II: 7,790 trials / Phase III: 5,741 trials.

Dataset Split: Temporal split

from tdc.multi_pred import TrialOutcome
data = TrialOutcome(name = 'phase1') # 'phase2' / 'phase3'
split = data.get_split()


[1] Fu, Tianfan, et al. “Hint: Hierarchical interaction network for clinical-trial-outcome predictions.” Patterns 3.4 (2022).

Dataset License: This dataset and code (including data collection and preprocessing, model construction, learning process, evaluation), referred as the Works, are publicly available for Non-Commercial Use only at Non-Commercial Use is defined as for academic research or other non-profit educational use which is: (1) not-for-profit; (2) not conducted or funded (unless such funding confers no commercial rights to the funding entity) by an entity engaged in the commercial use, application or exploitation of works similar to the Works; and (3) not intended to produce works for commercial use..