The future of drug discovery: Using AI to optimise the hit-to-lead process

Mirit Eldor, Managing Director, Life Sciences Solutions, Elsevier, looks at how artificial intelligence (AI) can improve the hit-to-lead process for small molecules. Early drug discovery takes an average of three-to-six years and accounts for 42% of total costs1 in the development of a new drug. It’s therefore no surprise that pharmaceutical companies are increasingly interested in using artificial intelligence (AI) to accelerate development and reduce costs. Research shows that companies could shorten this cycle by up to 30 months2 over the course of the small molecule drug discovery process by applying AI. In particular, interest is growing in how AI can be applied to the hit-to-lead (H2L) workflow – where the cost and difficulty of taking a small molecule hit to a potential lead compound is a longstanding industry challenge. Small molecule discovery has seen major developments throughout its history, with the introduction of computational methods alongside virtual screening leading to breakthroughs. Now, increasingly sophisticated AI provides a major opportunity to optimise the H2L process. However, as with any application of AI in the life sciences, research organisations must follow guidelines to ensure AI models deliver accurate and trustworthy results. How AI can drive improvements in H2L Designing a new molecule likely to interact with a target, synthesising, and then testing the molecule to identify the most promising lead candidates is time and resource intensive. There is great potential in the application of AI to H2L to accelerate the pivotal design-make-test-analyse cycle and reduce the number of iterations. AI can be used at each stage of this cycle. Here are some examples: Design: Conduct docking simulations, de-novo library design, high-throughput virtual screening, multi-parameter optimisation including synthetic accessibility and physicochemical property prediction. Make: Plan synthesis of new molecules and predict their yield and purity, identify problems with the synthesis process, identify more sustainable and green routes to synthesis. Test: Screen new molecules for potency and selectivity to interact with target proteins, predict efficacy and toxicity, identify most promising drug candidates. Analyse: Process large volumes of test data to identify correlations and trends, design further experiments to test most promising drug candidates. In the lead identification and optimisation phase of drug discovery, AI-based approaches offer significant advantages. For example, research3 has found AI can lead to “more than two times improvement over baseline on the key metric of “efficacy observed”. Additionally, AI models enable over a hundred times more in silico experiments compared to traditional screening methods. Lastly, AI can expedite the design of compounds to improve drug delivery efficacy in lead optimisation. AI guidelines companies must follow Pharmaceuticals is a highly regulated industry. Any AI application must be done safely, ethically, and with the necessary guardrails in place. To use AI responsibly, pharmaceutical companies must consider: 1. Sourcing comprehensive and high-quality data Several of the principal uses of AI in H2L phases are predictive; these include predicting structure-activity relationships, making accurate ADMET predictions, and predicting drug metabolism and pharmacokinetics (DMPK) endpoints. To ensure an AI model delivers an accurate prediction, high-quality and verifiable training data is essential. The data needed will depend on what the model is designed to achieve. Typically in H2L applications, training datasets include information on known compounds and their biological relationships, drawn from a range of sources such as proprietary data, scientific literature, regulatory data, clinical trials data and safety data. Assembling a finetuned dataset affords researchers a solid foundation from which to work to ensure the best possible outcomes and predictions. 2. Data structuring and enrichment Once researchers have the dataset, the data must be structured in a way that allows them to extract insights for H2L. Most datasets are not AI-ready – they are siloed and stored in myriad formats with insufficient metadata. So structuring datasets via ontologies (human-generated, machine-readable descriptions of categories) is a critical step. Ontologies standardise data against an agreed vocabulary, providing a shared language across an organisation. Next comes semantic enrichment. Through using text analytics to extract keywords, concepts and terms that can be 'fed' into AI models, researchers can transform text into clean, contextualised data, free from ambiguities. This harmonisation eliminates “noise” and is especially important in H2L when using data from multiple sources where technical terms and abbreviations are often used. 3. Domain expertise and human oversight Applying AI to H2L successfully necessitates a high level of domain expertise. Very few pharmaceutical companies have the right mix of skills needed for applying AI to H2L, such as building ontologies. Data scientists who can couple technology skills with scientific domain expertise are critical. These experts can leverage their contextual understanding of a process like H2L to parse questions and build models that will yield relevant results. The use of any AI model in pharmaceutical R&D must always be overseen – and validated – by a human. This human-in-the-loop approach ensures that researchers can combine their intuition, nuance and creativity with AI’s ability to carry out a high volume of work. Without a human-in-the-loop, 'black box' models can add risk with unvalidated findings during H2L and little clarity on the process for generating results. When human validation is lacking, a lead could be progressed erroneously to the optimisation phase, wasting considerable time and resources, and even posing a safety hazard. How to get the best results from AI AI has vast potential to reshape the drug discovery process – H2L is just one example. But using AI responsibly in the pharmaceutical industry also poses a significant challenge. It is only through the combination of trusted data, quality analytics and domain expertise that researchers can begin to unlock the opportunities offered by AI to improve their drug discovery processes. From DDW Volume 25 – Issue 3, Summer 2024 Read the digital issue References https://www.knowledgeportalia.org/costs-r-d https://www.sciencedirect.com/science/article/pii/S135964461830343X https://www.mckinsey.com/industries/life-sciences/our-insights/ai-in-biopharma-research-a-time-to-focus-and-scale About the author: Mirit Eldor is Executive Vice President for Strategy. Eldor is responsible for developing and enabling the execution of Elsevier’s strategy.