Gradient Oriented Active Learning for Candidate Drug Design

Abstract

One of the primary challenges of drug design is that the complexity of Biology often comes to the fore only when proposed candidates are eventually tested in reality. This necessitates making the discovery process more efficient by making it actively seek what it wants to know of reality. We propose Gradient Oriented Active Learning (GOAL), a technique for optimizing sequence design through active exploration of sequence space that interleaves performing experiments and learning models that propose experiments for the next iteration through gradient based descent in the sequence space. We demonstrate the promise of this method using the challenge of mRNA design as our guiding example. Using computational methods as a surrogate for experimental data, we provide evidence that for certain objectives, if one were restricted by the bandwidth or the number of experiments they can perform in parallel, increasing the number of iterations can still facilitate optimization using very few experiments in total. We show that availability of high-throughput experiments can considerably bring down the number of iterations required. We further investigate the intricacies of performing multi-objective optimization using GOAL.