Synthetic accessibility assessment using auxiliary responses

Abstract

Despite the recent advances in computational approaches to discovering new chemical compounds, accessibility assessment of designed compounds has still been a difficult task to automate because it is a heavily knowledge intensive task. A promising solution to such “AI-hard” tasks is collective intelligence approaches that aggregate opinions of a group of human non-experts or semi-experts. However, the existing aggregation methods rely only on synthetic accessibility evaluation scores given by humans, and they do not exploit auxiliary information obtained as byproducts of human evaluations such as that related to chemical structures. In this paper, we propose to exploit such auxiliary responses to obtain better aggregations. We introduce a new two-stage aggregation method of semi-expert judgments consisting of synthetic accessibility evaluation scores along with auxiliary responses that select substructures of targets obstructive to their synthesis. The first stage divides both semi-experts and substructures into clusters using stochastic block models to identify similar skills or properties. The second stage aggregates judgments while considering groups of semi-experts and substructures, and predicts synthetic accessibility. Our experiments show that the use of auxiliary responses improves the prediction performance and gives insight into evaluators and the structure of evaluated compounds.