The fusion of artificial intelligence (AI) and machine learning (ML) has reshaped drug discovery, expediting the development of innovative treatments. Initially, AI and ML models pinpoint potential drug targets by analyzing biological data like genomics, proteomics, and metabolomics, accurately predicting protein structures and interactions. These technologies refine lead compounds by forecasting pharmacokinetics and pharmacodynamics, hastening virtual screening and novel drug design for safer candidates. AI platforms optimize preclinical and clinical trials by predicting toxicity, patient categorization, and treatment outcomes, enhancing trial efficiency and cost-effectiveness through data integration. Despite hurdles like data quality and ethical concerns, AI and ML synergies hold immense promise in revolutionizing drug discovery and improving patient care.
TopIntroduction
The drug research and development process involves several stages, including recognition of drug target, target drug authentication, lead refinement and determination of preclinical molecule, preclinical assessment, and clinical testing. Despite significant financial investments, the expected clinical approval rate for new small agents in drug discovery is a mere 13%, with a substantial risk of non-success. The advent of computer-assisted drug design technology is seen as a hopeful method to transform this difficult situation, depending on strategic guidance throughout the development process. Computational methods are crucial for systematically assessing molecular characteristics (including selectivity, physicochemical properties, biological activity, pharmacokinetic parameters, and adverse effects) on a theoretical level. These approaches create optimized molecules with favorable attributes in silico, and multi-objective refinement through computational methods helps minimize the failure rate of preclinical hit molecules (Hassan et al., 2016). Artificial intelligence (AI) is employed via computer software applications to learn, analyze, and reveal vast pharmaceutical-related datasets, incorporating advancements in machine learning (ML) in a seamlessly integrated and automated fashion in drug design. The inception of machine learning models distinguishing between drugs and nondrugs goes back to 1998, with BASF and Vertex scientists independently proposing models for estimating drug-likeness. These models demonstrate the possibility of training machine learning models using chemical characteristics to distinguish between drugs (particularly, compounds intended for biological testing) and non-drugs (compounds without pharmaceutical applications). The difficulty encountered by machine learning models in assessing drug-likeness reflects the larger challenge in drug discovery, where the quality of a drug is not inherent to chemicals, and regulatory approval criteria may evolve over time (Hasselgren & Oprea, 2024).
Computational tools have been utilized to assess the potential carcinogenicity of impurities associated with drugs, resulting in the incorporation of structure-activity predictions in submissions to regulatory authorities. European Medicines Agency (EMA) in 2018 reflected on in silico tools use for evaluating nonmutagenic impurities-related risks in the absence of experimental data. Regulatory authorities have recently offered more extensive viewpoints on the application of AI/ML in the development and production of drugs, actively soliciting input from stakeholders. Incorporating AI into regulatory practices requires transparent models that health authorities can evaluate for reliability and usability (Kovarich & Cappelli, 2022). Industries across the globe are currently engaged in efforts to establish standards for the utilization of in silico tools, including AI/ML models. These endeavors aim to ensure transparency regarding the origins and quality of datasets, as well as the algorithms used, to achieve regulatory acceptance.
In modeling and simulation realm, which has become a standard method for demonstrating effects over physiology or safety across various indications or clinical populations, endeavors are underway to establish standards for assessing models. The SafetAI initiative, initiated by the National Center for Toxicological Research in collaboration with the Center for Drug Evaluation and Research, is dedicated to developing artificial intelligence models for toxicological endpoints important in evaluating safety of drug and potentially influencing the Investigational New Drug (IND) review process (Soni & Hasija, 2022). The first AI-developed chemical moiety, DSP-1181, a potent and long-lasting agonist of the 5-HT1A receptor, was synthesized within 12 months using Exscientia’s MPO approach, Centaur Chemist. It entered Phase I clinical trials in January 2020 for obsessive-compulsive disorder, and its chemical structure has not been disclosed. In-silico Medicine's ISM01-055 has emerged as a promising candidate, being the first AI-designed compound to advance into Phase II clinical trials, scheduled for June 2023. Developed using Chemistry platforms and the PandaOmics, ISM018-055, potentially safeguarded by a patent, is intended for addressing idiopathic pulmonary fibrosis by focusing on NCK-interacting protein kinase (Hasselgren & Oprea, 2024).
Essential AI methodologies are