Extrapolation is Not the Same as Interpolation

Abstract

AbstractWe propose a new machine learning formulation designed specifically for extrapolation. The textbook way to apply machine learning to drug design is to learn a univariate function that when a drug (structure) is input, the function outputs a real number (the activity): F(drug) → activity. The PubMed server lists around twenty thousand papers doing this. However, experience in real-world drug design suggests that this formulation of the drug design problem is not quite correct. Specifically, what one is really interested in is extrapolation: predicting the activity of new drugs with higher activity than any existing ones. Our new formulation for extrapolation is based around learning a bivariate function that predicts the difference in activities of two drugs: F(drug1, drug2) → signed difference in activity. This formulation is general and potentially suitable for problems to find samples with target values beyond the target value range of the training set. We applied the formulation to work with support vector machines (SVMs), random forests (RFs), and Gradient Boosting Machines (XGBs). We compared the formulation with standard regression on thousands of drug design datasets, and hundreds of gene expression datasets. The test set extrapolation metrics use the concept of classification metrics to count the identification of extraordinary examples (with greater values than the training set), and top-performing examples (within the top 10% of the whole dataset). On these metrics our pairwise formulation vastly outperformed standard regression for SVMs, RFs, and XGBs. We expect this success to extrapolate to other extrapolation problems.