Introduction

What I will describe below is a rough first approximation of what it is like to work in the field of machine-learning-assisted small-molecule design.

Imagine that you are tasked with solving the following machine-learning problem:

There are 116 billion balls of varying colors, textures, shapes, and sizes in front of you. Your job is to predict which balls will stick to a velcro strip. To help start you off, you’re given a training set of 10 million balls that have already been tested; which ones stuck and which ones didn’t. Your job is to predict the rest. You give it your best shot, train a very large transformer on 80% of the (X, Y) labels, and discover that you’ve achieved an AUC of .76 on a held out 20% set of validation balls. Not too shabby, especially given that you only had access to .008% of the total space of all balls. But, since you’re a good hypothetical scientist, you look more into what balls you did well on, and which balls you did not do well on. You do not immediately find any surprises; there is mostly uniform error across color, textures, shapes, and sizes, which are [...]