Prediction and Prioritisation of Novel Anthelmintic Candidates from Public Databases Using Deep Learning and Available Bioactivity Data Sets.

The control of socioeconomically important parasitic roundworms (nematodes) of animals has become challenging or ineffective due to problems associated with widespread resistance in these worms to most classes of chemotherapeutic drugs (anthelmintics) currently available. Thus, there is an urgent need to discover and develop novel compounds with unique mechanisms of action to underpin effective parasite control programmes. Here, we evaluated an in silico (computational) approach to accelerate the discovery of new anthelmintics against the parasitic nematode Haemonchus contortus (barber's pole worm) as a model system. Using a supervised machine learning workflow, we trained and assessed a multi-layer perceptron classifier on a labelled dataset of 15,000 small-molecule compounds, for which extensive bioactivity data were previously obtained for H. contortus via high-throughput screening, as well as evidence-based datasets from the peer-reviewed literature. This model achieved 83% precision and 81% recall on the class of 'active' compounds during testing, despite a high imbalance in the training data, with only 1% of compounds carrying this label. The trained model was then used to infer nematocidal candidates by in silico screening of 14.2 million compounds from the ZINC15 database. An experimental assessment of 10 of these candidates showed significant inhibitory effects on the motility and development of H. contortus larvae and adults in vitro, with two compounds exhibiting high potency for further exploration as lead candidates. These findings indicate that the present machine learning-based approach could accelerate the in silico prediction and prioritisation of anthelmintic small molecules for subsequent in vitro and in vivo validations.