OBJECTIVES: In the light of significant risk of failures and ambition to access the market as quickly as possible, predicting which molecules will finally access the market is a fundamental need for pharmaceutical industry and physicians. Nowadays, artificial intelligence algorithms expand and could provide solution to healthcare questions which involve high-dimensional data analysis. With this in mind, we investigated how a deep learning model can be used to predict efficacy in phase I oncology trials from unstructured preclinical literature.

METHODS: All available phase I clinical trials and their relative preclinical abstracts evaluating oncology agents for adult patients were extracted from PubMed without date limitation. For each drug, anti-tumor clinical activity was retrieved. We developed a Hierarchical Attention Network model that was trained to predict clinical efficacy of drugs. Therefore, a transfer learning method was used to provide a slight improvement of the model’s performance and allow us to reduce overfitting.

RESULTS: In total, 367 drugs were identified representing 6,879 preclinical abstracts. Our deep learning model allowed to predict clinical activity of drugs with an accuracy of 0.70 on the main task. The model enables to predict clinical efficacy of drugs with a specificity of 0.67 and a sensitivity of 0.64.

CONCLUSIONS: In this work, we proposed a state-of-the-art deep learning model applied on drug development leveraging unstructured scientific knowledge. Although no gold standard exists to compare with our model’s performance, our deep learning approach seems to predict phase I efficacy results with a promising accuracy. This tool aims to guide oncologist on the most valuable phase I and/or drugs.