Unlocking AI in drug discovery: A roadmap to scalable solutions

Dr Mutlu Dogruel, Vice President of AI Solutions at Cresset, examines AI’s transformative potential, highlights the current challenges and introduces a series of potential solutions to implement AI in the drug discovery landscape successfully. The use of artificial intelligence (AI) is creating a paradigm shift in drug discovery and optimisation processes, allowing for the analysis of large-scale datasets, the design of new molecules, and the prediction of drug properties more efficiently. This can unlock new opportunities for pharmaceutical and agricultural companies to develop innovative solutions for unmet clinical and agronomic needs. AI-driven drug discovery is increasingly commonplace, with companies actively exploring hundreds of AI-driven candidates in the discovery and early clinical stages1. However, no AI-discovered or designed drugs have yet reached the market. To harness AI’s full potential in drug discovery and deliver life-saving medicines to patients, the biopharmaceutical industry must overcome the challenges creating ‘AI bottlenecks’ in the discovery process. The transformative power of AI While Foundation Models (FMs), including Large Language Models (LLMs) and image-based models, may not exhibit true intelligence, their potential to revolutionise fields like drug discovery and materials science is undeniable. These models can process vast datasets, recognise patterns and generate valuable insights – capabilities that are essential in many scientific domains. Gartner predicts that by 2025, more than 30% of drugs and materials will be systematically discovered using Generative AI (GenAI) techniques. Considering that no AI-discovered or designed drugs are currently on the market, this forecast represents a significant shift within a short timeframe. The biopharmaceutical industry is already making considerable progress toward this goal, with several biotech companies advancing AI-guided drug candidates to clinical trials2. Not all language models need to be large. Small Language Models (SLMs) are gaining traction, particularly in the life sciences. They excel in tasks such as product feature orchestration and function calling, where the demand for detailed responses is lower. SLMs offer advantages like reduced computational requirements and faster processing times, making them ideal for efficiency-focused applications. Moreover, LLMs are not limited to human languages – they can also be trained to understand and generate biological and chemical languages, such as amino acids, genes, ligands and proteins. This capability enables their use in various bioinformatics and cheminformatics applications, including predicting DNA regulatory elements and optimising drug properties. As these models continue to prove their worth, AI's integration into drug discovery workflows will only deepen, offering researchers numerous benefits, including: Integrated generative AI chatbots and copilots: AI copilots can significantly boost productivity by enabling users to interact with chatbots to search for features, engage with product functionalities, generate programming code and ask questions about the product’s usage. Over time, copilots may evolve to execute product functions on behalf of the user, breaking down complex problems into manageable parts, each handled by a specialised AI agent. For instance, generative AI can be incorporated into established drug discovery tools to improve protein-ligand docking, quantitative structure-activity relationship (QSAR) modelling, and molecular simulations, thus enhancing predictive analytics and user experience. Expanding accessible chemical space: Generative models have a remarkable ability to design novel drug-like molecules based on specific chemical or pharmacological properties, or even detailed molecular information about drug targets. This expands the range of explorable chemical space, opening up new avenues for discovering molecules that could potentially address difficult-to-treat diseases. Even when designing entirely new molecules isn't the goal, these models excel at suggesting innovative substructures, thereby enhancing efficacy and optimising properties. Enhancing biological modelling and molecule design: Machine learning (ML) can significantly improve the analysis of protein-drug interactions and help identify potential adverse effects, such as toxicity. By processing large datasets and accurately identifying patterns, biological modelling can streamline drug discovery, providing researchers with the precision needed to make informed decisions. This improved modelling can guide medicinal chemists in designing better-performing molecules by leveraging expanded chemical space. Rapid virtual screening of ultra-large libraries: AI-powered virtual screening allows for the rapid processing of billions of compounds, significantly accelerating compound analysis. By filtering out candidates that lack desired properties, researchers can focus on the most promising compounds, examining far more than would be possible through traditional wet lab methods. Integrating these in silico experiments is crucial for reducing reliance on expensive, time-consuming wet lab experiments, thereby enhancing efficiency and streamlining the drug discovery pipeline. However, despite claims of improved speed and accuracy, AI-based methods may not always outperform traditional, well-established screening and docking techniques. In these cases, AI can still play a crucial role by orchestrating the execution of the right methods in the correct order. Unlocking the power of quantum mechanics at classical speeds with AI: In traditional computational chemistry, quantum mechanics (QM), also known as ‘ab initio’, is the gold standard for atomic-level simulations, providing highly accurate system analysis based on fundamental physical laws. However, ab initio simulations are notoriously computationally expensive. To address this challenge, deep learning models are currently being developed using datasets of hundreds of thousands of molecules with pre-calculated properties. These models can generate QM-quality atomic charges for a molecule in just a fraction of a second. Technological breakthroughs like this enable precise yet rapid calculations in classical (non-quantum) atomic simulations, which are essential for tasks such as conformer generation, molecular dynamics, and free energy perturbation. Producing treatments for unmet medical needs: AI offers the tools to examine complex diseases comprehensively. Unlike traditional approaches that often focus on one specific area (such as a key protein), AI has the capacity to consider the broader disease pathway and leverage knowledge from a vast array of scientific literature. This capability allows AI to identify potential treatment strategies that consider multiple factors simultaneously, leading to the discovery of drug candidates that might not have been previously considered. By providing researchers with valuable insights, AI can help develop drugs targeting complex and hard-to-treat diseases. Leveraging AI’s capabilities allows drug developers to enhance efficiency, accuracy, and patient-centricity in drug discovery, ultimately leading to improved treatments and healthcare outcomes. The primary strategic benefit of AI lies in its ability to look at the big picture and focus on multiple aspects simultaneously. In the context of drug discovery, this means analysing various biological pathways simultaneously and integrating results with supporting datasets in a meta-analysis. However, fully integrating AI into drug discovery is a complex undertaking with many challenges. Navigating the complexities of AI integration Integrating AI into drug discovery efforts can significantly advance research and develop innovative solutions for better patient outcomes. However, introducing or optimising AI systems within existing software tools and workflows can be complex and resource-intensive. Several factors must be carefully considered to navigate AI integration effectively: Interoperability with existing systems: Ensuring that AI solutions seamlessly integrate with current software tools and platforms is crucial. This requires a comprehensive understanding of both legacy systems and the new AI technologies being introduced. Creating APIs and middleware that facilitate smooth communication between systems can significantly ease the integration process. Scalability and performance: AI models often require substantial computational resources. It is essential to plan for scalable infrastructure that can handle increased demand, whether through cloud computing solutions or on-premises hardware. If using a cloud vendor, a secondary geographical region with the same AI models must be set up to ensure high availability by redirecting applications to it in case of outages in the original region. Additionally, performance optimisation techniques should be employed to ensure that AI models run efficiently without causing bottlenecks. User training, adoption and change management: The successful deployment of AI solutions depends heavily on user adoption. Providing extensive training and support to end-users ensures they are comfortable and proficient in utilising new AI-driven tools. User-friendly interfaces and intuitive design can further enhance the adoption process. Effective change management strategies are essential to address potential user resistance or concerns, equipping organisations to ensure a smooth transition and maximise the benefits of AI integration. Dedicated training systems, support materials, and staff development programmes are crucial for facilitating user understanding and adoption. These efforts not only ensure effective AI integration into drug discovery workflows but also help alleviate any anxiety about AI within the organisation. By proactively fostering AI literacy and upskilling the workforce, organisations can cultivate a more receptive and capable team, ready to leverage the full potential of AI technologies. Ethical and responsible AI practices: Implementing AI responsibly is non-negotiable. This involves ensuring transparency in AI decision-making processes, maintaining fairness to avoid biases, and adhering to ethical guidelines. Regular audits and updates of AI systems are necessary to uphold these standards, especially to spot and fix any issues that may arise when new versions of your LLM become available. Additionally, responsible AI filters might be updated behind the scenes by cloud vendors, which further underscores the importance of conducting these audits regularly. Continuous monitoring and maintenance: AI systems require ongoing monitoring to ensure they remain accurate and effective over time. Regular updates and maintenance are essential to adapt to new data, refine algorithms and address any emerging issues. Establishing a robust monitoring framework can help in swiftly identifying and rectifying any anomalies. Collaboration and cross-disciplinary teams: Successful AI integration often necessitates collaboration across various departments, including IT, data science and domain experts from drug discovery and development. Building cross-disciplinary teams can facilitate a more holistic approach to AI integration, ensuring that all perspectives and expertise are considered. Tackling AI hallucinations: AI ‘hallucinations’ – the generation of unexpected or nonsensical outputs by AI models – present both challenges and opportunities in drug discovery. These hallucinations, when treated as novel, unseen drug candidates, can open up new avenues for therapeutic innovation. However, these AI-generated molecular candidates must undergo rigorous validation and optimisation to ensure their chemical and physical properties meet the desired criteria. To mitigate the risks associated with AI hallucinations, several techniques can be employed. Retrieval-Augmented Generation (RAG) architectures are one such approach. RAG enhances the reliability of AI-generated content by integrating an orchestrator layer that retrieves relevant documents from a repository. This layer uses the retrieved information to refine the AI model's prompt, thereby ensuring that the generated responses are both relevant and grounded in accurate data. In a similar vein, knowledge graphs provide a structured and contextual framework by linking entities and concepts in a manner that reflects real-world relationships. This structured information helps guide AI models, reducing the likelihood of generating irrelevant or incorrect outputs. Building on these principles, GraphRAG3, developed by Microsoft Research, combines the strengths of both RAG and knowledge graphs. By integrating knowledge graphs into the retrieval-augmented generation process, hybrid solutions like GraphRAG enhance the AI model's ability to generate responses that are not only contextually accurate but also semantically rich. Employing multiple LLMs with a majority voting approach can enhance output reliability by comparing responses across different models and selecting the most consistent or frequent answers. However, this method has its caveats: to avoid consensus bias, it is essential to use truly independent LLMs. If the models are too similar (e.g., trained on the same datasets, architectures, or even using the same underlying framework), they may exhibit similar biases or errors. Additionally, majority voting is best suited for short, factual answers rather than complex reasoning tasks. While this technique helps filter out unreliable outputs and mitigates the impact of hallucinations, it should be supplemented with other validation methods to avoid developing a false sense of confidence in the results. Agentic frameworks can further improve the handling of AI hallucinations and complex problem-solving. By breaking down tasks and distributing them among multiple autonomous agents, these frameworks can enhance the accuracy and relevance of generated responses. Each agent can focus on specific aspects of a problem or validation process, ensuring a more thorough and adaptive approach. Cross-verification among agents helps to identify and correct hallucinations, while contextual anchoring maintains the coherence and accuracy of the outputs. Additionally, adjusting parameters such as temperature and top-p sampling of each LLM further refines the generation process. Lowering the temperature reduces randomness, making responses more focused and coherent, while top-p sampling ensures that only the most probable and contextually relevant information is considered. Together with agentic frameworks and majority voting, these strategies enable LLMs to deliver more accurate and meaningful responses, improving their effectiveness in complex applications and reducing the likelihood of erroneous outputs. Finally, implementing post-processing algorithms to assess and filter AI outputs can further mitigate the effects of hallucinations. These algorithms can evaluate the coherence and relevance of generated content before it is used in practical applications. Above all, implementing a scalable roadmap allows companies to successfully navigate the complexities of AI integration and fully benefit from the transformative power of these technologies. Now, let’s consider the importance of having a robust AI strategy in place and the benefits it can bring to drug discovery and wider organisational goals. A roadmap to building a robust AI strategy Successfully implementing AI requires a well-thought-out strategy that takes into account drug discovery objectives and wider organisational goals. Researchers can take several steps to prepare for potential challenges and solve complex drug discovery problems with a robust AI strategy. Integrate AI across the entire drug discovery lifecycle To maximise benefits, AI integration must span the entire drug discovery life cycle. AI can be leveraged to identify new drug targets, design and optimise drug molecules using a streamlined ‘design, make, test, analyze’ (DMTA) cycle and predict drug-target interactions. By adopting AI across all drug discovery stages, productivity is enhanced and researchers gain actionable DMTA insights quickly. This can help reduce the time and cost of drug discovery and development, improving the success rate of clinical trials. Develop a systematic AI solution A structured and methodical approach, as part of a broader organisational strategy, is crucial to ensuring seamless and effective AI integration in drug discovery. This involves meticulous planning, execution, and management of AI solutions to align with the organisation’s overarching objectives. A systematic approach enables thorough implementation across the entire organisation and ensures consistency in applying AI, maximising its benefits in driving drug discovery forward. To achieve this, ML and LLM operations must align with the overall digital transformation goals and the organisation's broader commercial objectives to create efficient and scalable AI solutions. Prioritise ethical AI Prioritising the development of responsible AI models that align with ethical principles is crucial to ensuring compliance and delivering ethical drug treatments. To meet responsible AI demands, transparency and explainability in AI models must be demonstrated with clear and accessible information about the training data, algorithms and decision-making processes4. Additionally, it’s important to build fair and unbiased AI models. To minimise biases, proactive measures should be taken to address potential issues originating from the training data or algorithms. Doing so can help ensure the responsible use of AI technology. Leverage AI across the organization, not just for R&D users Beyond revolutionising drug discovery, AI can also optimise internal processes. Integrating AI internally streamlines operations, enhances efficiency, and empowers the workforce. Automating cumbersome or repetitive tasks with AI frees up time for employees to focus on high-value activities requiring creativity, expertise, and collaboration. Improvements in internal operational efficiency can fuel innovation in drug discovery, and integrating AI into every aspect of an organisation can drive future success. Looking ahead: Embracing a scalable AI solution Executing this AI roadmap can be challenging for some companies due to the complexity of integrating AI into existing workflows, the substantial resource and skill requirements, and the need for effective change management, but doing so can unlock AI's benefits and transform drug discovery and development. Incorporating AI throughout the drug discovery process may require a strategic and innovative partner who can leverage AI’s full potential to enhance outcomes. Developing the right strategic alliances can provide access to cutting-edge AI insights and help implement modular and scalable AI solutions. Given the rapid evolution of AI and the need for continuous improvement, opting for a modular approach is essential to simplify AI integration and future-proof drug discovery. Employing modular architectures combined with cloud-based solutions and integrating ML and LLM operations provides the flexibility necessary to consistently advance AI-driven systems, allowing them to be seamlessly adapted and continuously improved. Embracing scalable AI empowers researchers to unlock new possibilities in drug discovery and look forward to a future where medicines are delivered to patients at an accelerated pace. References AI in small molecule drug discovery, accessed June 28, 2024, from: https://www.elsevier.com/en-gb/industry/ai-in-small-molecule-drug-discovery These Six Biotechs are Winning the Race to Get AI-Designed Drugs to the Clinic, accessed July 15, 2024, from: https://www.insideprecisionmedicine.com/news-and-features/these-six-biotechs-are-winning-the-race-to-get-ai-designed-drugs-to-the-clinic/ GraphRAG: Unlocking LLM discovery on narrative private data, accessed August 2024: GraphRAG: Unlocking LLM discovery on narrative private data - Microsoft Research Ethics of Artificial Intelligence, accessed June 28, 2024, from: https://www.unesco.org/en/artificial-intelligence/recommendation-ethics About the author Dr Mutlu Dogruel is Vice President for AI Solutions at drug discovery specialists Cresset. He has previously worked at Microsoft, holding various positions at the company including Pharma GenAI Lead Architect and Senior Cloud Solution Architect in Data and AI, as well as similar roles with HP and software company Eagle Genomics. Dr Dogruel has a PhD in Bioinformatics from the University of Cambridge.