OBJECTIVES: Artificial intelligence (AI) and experimental technologies are increasingly combined in drug discovery and preclinical research. The objective is to provide an overview of opportunities and challenges of currently used computed technologies.

METHODS: A targeted literature review was conducted on OVID. The selection focused on the two last years, reflecting the tremendous development of AI in non-clinical research.

RESULTS: 613 studies were identified, mostly from US and China, and were predominantly related to drug discovery. Tool bases were high-throughput docking, quantitative and visualized structure-activity relationship (QSAR/ VISAR), drug-target interaction and absorption, distribution, metabolism, excretion, and toxicity (ADME/T) processes.

Machine learning and neural networks, through mathematical modelling and feature extraction, have improved, sped up and drastically rearranged non-clinical research processes resulting in candidates with optimal efficacy and tolerability by predicting molecules properties in-silico from their directed structure and studying disease models. The performance of AI models is affected by the scarcity and high cost of reliable, high-quality databases. These latter will define the quality of descriptors required for reaching a performant and reproducible digital molecular triage.

Current AI systems are vulnerable to several issues, including non-reproducibility of results and the risk of over/under-fitting, which impact the accuracy of results.

CONCLUSIONS: Emerging AI techniques have renewed the drug discovery journey and decreased the costs by reducing the risk of a drug candidate failing in preclinical research. Nevertheless, original AI approaches must be optimized to address unmet need, such as application to complex living systems. Further AI implementations are required beyond drug discovery while complying to dogmas and ethics of regulatory and potentially Health Technology Assessment (HTA) agencies.