Can graph similarity metrics be helpful for analogue identification as part of a read-across approach?

Abstract

Read-across is a technique used to fill data gaps for substances lacking specific hazard data. The technique relies on identifying source analogues with relevant data that are ‘similar’ to the substance of interest (target). Typically, source analogues are identified on the basis of structural similarity but the evaluation of their suitability for read-across depends on other contexts of similarity. This manuscript aimed to review the ways in which source analogues are identified for read-across using chemical fingerprint/scaffold approaches before describing graph-based approaches including; graph kernel, graph embedding, and deep learning. To demonstrate how these could be practically used for analogue identification, five different toxicity datasets of varying size and diversity were selected that had been the subject of previous read-across or QSAR analyses. One dataset was an analogue set whereas the other four datasets comprised substances evaluated for their skin sensitisation, skin irritation, fathead minnow aquatic toxicity and genotoxicity potential. The analogues and their associated similarities using the different graph based approaches were compared with the outcomes from two chemical fingerprint approaches (ToxPrints and Morgan). The results for each dataset are briefly described. Based on the examples evaluated, graph kernel approaches were found to have some promise, in contrast unsupervised whole graph embedding approaches were ineffective for all the datasets evaluated. Graph convolutional networks produced meaningful embeddings for the genotoxicity dataset evaluated. Depending on use case, availability and size of training data, graph similarity approaches have the potential to play a larger role in analogue identification and evaluation for read-across.