Table 1.

Examples of drug discovery applications of various machine learning to targets and diseases from AI companies.

Area of research / Disease	Target/property	Outcome	Company	References
Canavan disease	aspartate N-acetyltransferase	AtomNet deep neural network for structure-based drug discovery uses a model trained on bioactivity data and protein structures. They scored 10M molecules and 60 were tested in vitro with 5 compounds having low or sub μM activity.	Atomwise	[62]
Infectious disease	COVID-19	Workflow used knowledge graph information from recent literature using machine learning (ML) based extraction to identify baricitinib. This molecule progressed from a clinical trial to emergency FDA approval.	BenevolentAI	[63]
Various	Various drug rediscovery examples	de novo generative design benchmarking study used rediscovery of various drugs with different algorithms.	BenevolentAI	[64]
Rare disease	Fragile X	Disease-Gene Expression Matching approach to repurposing identified sulindac which rescued the phenotype in the Fmr1 KO mouse.	Healx	[65]
Fibrosis	DDR1 kinase	Generative machine learning to discover novel compounds validated in vivo	In silico Medicine	[52]
Infectious disease	Antibacterials against E. coli	Machine learning, virtual screeing and in vitro testing	In silico Medicine	[66]
Various	Various	Different generative approaches were used and evaluated including entangled conditional adversarial autoencoder, reinforced adversarial neural computer, and Adversarial threshold neural computer. They either purchased compounds similar to those proposed and then tested them in vitro against various kinases or alternatively they synthesized proposed compounds and tested them	In silico Medicine	[67] [68] [69]
Various	sEH, ERa and c-KIT	Applied machine learning algorithms (random forest or graph convolutional neural network (GCNN)) to DNA encoded libraries then validated the predictions in vitro. GCNN models had higher hit rates and potencies.	X-Chem	[56]
Various	IMPDH, JNK3 etc.	Graph based deep generative model to create linkers for combining two fragments for scaffold hopping and PROTACS using a gated graph neural network incorporating 3D information. Molecules were assessed with a range of 2D and 3D metrics and outperformed a baseline.	ExScientia Ltd	[70]
Various	Various	Multiple machine learning approaches applied to searching commercial and proprietary libraries, lead optimization and repurposing.	Collaborations Pharmaceuticals, Inc.	See Table S2