Page 39 - AIH-2-2

Artificial Intelligence in Health New drug discovery in the AI era

Figure 3. Drug toxicity evaluation: wet-laboratory experiments and artificial intelligence/machine learning-based in silico prediction tools
Abbreviation: OECD: Organization for Economic Co-operation and Development.

Bridging the gap between static in vitro models and dynamic of drug discovery and development, from hit identification
in vivo homeostatic systems is difficult. Furthermore, these to candidate selection for clinical trials.
models need to undergo extensive validation to demonstrate During the lead discovery phase, AI models such as
limited interexperiment and inter-laboratory variability recurrent neural networks and generative adversarial
as well as reproducibility in in vitro in vivo correlation. networks generate NCEs, predict target binding affinities,
Therefore, these models are currently used for exploratory and expedite candidate selection. Molecular dynamic
toxicology and establishing PK/PD relationships. simulations and ML approaches enhance the efficiency and
For in silico predictions, knowledge-based standalone accuracy of de novo drug design.
quantitative structure-activity relationship tools have long Compounds predicted to have poor in vitro ADME
been available. Examples include Derek, Meteor, StAR, and (solubility, permeability, chemical, and metabolic stability),
TopKat for drug toxicity, Ecosar for ecotoxicity, and Biowin suboptimal PK properties (low oral bioavailability), drug–
for biodegradability prediction. However, traditional drug interaction (DDI) potential (CYP inhibition or
approaches based on structure-activity relationships induction), and toxicity (mechanism/off-target) that could
and physicochemical attributes did not account for drug affect the clinical safety and efficacy of NCEs are effectively
interactions with the human-specific liver proteome, identified and circumvented by AI/ML-powered virtual
resulting in inaccurate predictions of DILI. Over the strategic planning platforms. 19,20 Consequently, the holistic
18
past decade, it has become clear that integrating various AI/ML-driven evaluation across all drug discovery verticals
drug discovery verticals into a unified computational tool facilitates early no-go decisions. 21,22 Figure 4 outlines the
is essential to lowering attrition rates and shortening the brief history of AI-driven drug discovery beginning - 2017
drug discovery timeline. AI/ML models that combine – 2018.
physicochemical attributes, anticipated on-target biological
interactions, and predicted off-target toxicity in humans 3.1. AI
can help address gaps in predicting DILI.
AI, a branch of computational science, focuses on creating
Recently, there has been a parallel emphasis on AI/ systems that perform tasks typically requiring human
ML-based approaches to accelerate the drug discovery intelligence. In drug discovery, AI has been successfully
process and reduce NCE failures during clinical trials and applied to target protein structure identification, 23,24 de
phase IV withdrawals of marketed drugs. novo drug design, 25,26 compound docking studies, virtual
27
3. New drug discovery in the AL/ML era screening, 28,29 retrosynthesis reaction prediction, 30,31
bioactivity and toxicity prediction,
and in silico clinical
32,33
Over recent years, numerous AI/ML approaches have been trials. Figure 5 illustrates AI, its subsets, and respective
34
developed and successfully implemented at various stages tools.
Volume 2 Issue 2 (2025) 33 doi: 10.36922/aih.4423