Conformal Prediction-based Machine Learning in Cheminformatics: Current Applications and New Challenges

Conformal Prediction (CP) is a distribution-free Machine Learning (ML) framework that has been developed in the last ~25 years to provide well calibrated prediction subsets/intervals that include the true label with a user pre-defined probability, only requiring data exchangeability. It is based on the concept of nonconformity (or dissimilarity) of the new prediction compared to previous data and their predictions, so that the prediction subset/interval size is larger for new “unusual” instances and smaller for “typical” instances. Given its simplicity and ease of applicability, since 2012 it has been widely adopted in Cheminformatics, especially in the Quantitative Structure-Activity Relationship (QSAR) modeling and Molecular Screening areas. This rapid popularization of CP in Cheminformatics can be explained on the grounds that: a) it can handle the applicability domain (AD) issue of ML models, of large importance in Cheminformatics due to the immense size of the chemical space; b) it deals with classification of heavily imbalanced datasets typical in Molecular Screening; and c) it quantifies compound-specific prediction uncertainties, especially useful as it allows to implement gain-cost strategies to accelerate drug discovery by reducing compounds to test. This comprehensive review introduces the method, provides a full appraisal of the work done in the field of Cheminformatics (with special emphasis in the QSAR and Molecular Screening arenas), and discusses its pros and cons and new challenges, especially for Deep Learning applications and nonexchangeable datasets, a very frequent situation in Cheminformatics.