Predictive AI model for drug development
Manu Kumar Shetty, MBBD, MD, Maulana Azad Medical College
Anubha Gupta, Department of ECE, Centre of Excellence in Healthcare, IIIT Delhi.
Incorporating AI predictive models into drug development demands thoughtful consideration. Challenges encompass navigating complex biology and acquiring ample human data. Computational methods, GAN models and transfer learning, address knowledge gaps. The use of XAI enhances decision transparency. Addressing challenges is pivotal for its adoption in predicting drug efficacy and safety.
The integration of artificial intelligence (AI) predictive models in drug discovery and development demands a thoughtful examination. AI models, particularly large Language Models, have undoubtedly proven their effectiveness in certain tasks. However, expecting the same levels of accuracy and output in drug discovery and development field is unrealistic. AI application in drug development requires careful consideration and a comprehensive understanding to avoid futile efforts. The underlying reason lies is the fact that LLM models were trained using extensive human-generated data, a luxury not easily replaceable in the complex biology field of drug development. To harness the power of AI in this sector, a substantial amount of real and intricate biological data must be generated.
Challenges
In this context, the initial step involves comprehensively identifying all factors influencing the outcome related to drug development. Even if we may not understand the either positive or negative influences or their direct or indirect effects on outcome, recognising these factors is crucial. Secondly, animal model generated data can not be useful in developing AI models and predicting the drug efficacy in humans. AI models developed using animal data would generate outcomes that will have the same level of attrition rate in drug success Therefore, it is essential to gather a sufficient amount of human data to develop accurate AI models . Third, the development of an explainable AI (XAI) model is necessary. This XAI model should provide clarity on how and why the AI model reaches its decisions. By doing so, we can analyse established features and their correlation with the XAI decision, gaining insights into the intricate nature of biological processes.
Solutions
Currently, tackling these challenges is important before applying AI. This is complicated by limited understanding of complex biology and limited resources. However, by employing computational methods and leveraging AI, we can address a significant portion of these issues with the right approach. The primary challenge involves comprehending all factors and relationships influencing outcomes in drug development. In these conditions, fully trained AI models using animal data directly cant be used for humans, but these models can be applied to generate the data of common biology concepts and relationships. After transfer learning is applied to develop generative AI models. Considering the complexities of biology, from DNA and RNA sequences to amino acid structures and the impact of external and internal factors on protein structures, computational methods such as sequence analysis models and generative adversarial network (GAN) models prove valuable. These models help fill gaps in our understanding. Once we understand that, it is possible to generate the extra data using generative AI models. Finally, by using transfer learning and XAI models, more reliable AI models can be developed.
Applications of AI
Effective decision-making in the implementation of AI involves the thoughtful selection of appropriate AI models at various stages of drug discovery and development. This encompasses the prediction of key drug characteristics. The integration of AI and data science has uncovered valuable applications in the execution of clinical trials, with numerous industries seamlessly incorporating these cutting-edge technologies.
Hospitals that embrace electronic health records (EHRs) contribute significantly to the pool of high-quality healthcare data. Such data not only supports the swift adoption of AI in clinical trials but also plays a pivotal role in advancing medical research. Notably, advancements in speech-to-text technology and other language models have empowered healthcare professionals by saving time in managing electronic health records. The transformation of doctor-patient conversations into electronic health records not only streamlines the process but also alleviates the stress burden on healthcare workers. This EHRs data proves instrumental in various aspects, including patient cohort selection, expediting subject recruitment, facilitating the coordination of multi-center trials, and predicting early outcomes. Despite these advancements, challenges persist in other phases of drug development, enhancing drug efficacy, optimising drug pharmacokinetics, and predicting offtarget side effects.
The pharmacokinetics of a drug, a critical determinant of both its efficacy and safety, is intricately influenced by factors such as lipophilicity, solubility, permeability, molecular structure, metabolising enzymes, blood pH, and in-vivo physiochemical reactions. Developing a Multi-level AI model that incorporates all these factors is essential to derive a holistic understanding of the four key elements of drug pharmacokinetics: absorption, distribution, metabolism, and excretion.
Currently, pharmacokinetic parameters are derived from in-vitro and animal studies. However, AI models built solely on these data may lack accuracy, given the higher complexity of human biology compared to in-vitro or animal conditions. Achieving a more precise AI model requires a significant infusion of human data that reflects the nuanced pharmacokinetics of drugs in humans. Furthermore, the early stages of drug design benefit from the adoption of AI for achieving desired pharmacokinetics. Utilising concepts such as structure-activity relationships and other in-vivo factors specific to human biology enhances the accuracy and effectiveness of AI models in predicting and optimising pharmacokinetics profiles during the drug development process.
Drug Efficacy: Despite a drug exhibiting effectiveness in animal models, the attrition rate during phases 2 and 3 trials remains high due to a lack of drug efficacy. Relying mainly on animal studies for efficacy results is not entirely trustworthy. Various factors influence drug efficacy, such as target and drug features—these include the drug's affinity to the target, whether the binding is reversible or irreversible, and the conditions of drug binding in-vivo. Implementing the aforementioned solutions is essential for the development of a reliable AI model to predict drug efficacy.
Ensuring Drug Safety and Pharmacovigilance: Anticipating offtarget side effects prior to the marketing of a drug requires a comprehensive understanding of its affinity to various targets. AI models play a crucial role in facilitating this prediction by taking into account factors beyond the mere protein and drug chemical structures. The intricate nature of protein structures, influenced by both physiological and pathological conditions, demands an advanced AI model capable of navigating the intricacies of the body's biology and physiochemical reactions. Post release of a drug in the market, monitoring side effects becomes a key aspect of pharmacovigilance, and AI proves instrumental in this sector as well. Unstructured data is efficiently transformed into structured data using AI methodologies. Leveraging EHR significantly expedites the pharmacovigilance process, enhancing its efficiency and effectiveness. The integration of AI not only streamlines the monitoring of side effects but also brings agility to the overall pharmacovigilance workflow.
The dynamic landscape of AI in drug development continues to evolve, promising innovative solutions while simultaneously necessitating thoughtful consideration of the associated challenges. Adopting AI in drug development involves recognising factors that influence the drug efficacy, safety, PK, And addressing challenges in understanding complex biology, and utilising computational methods. Overcoming the limitations of animal data necessitates gathering sufficient human data for accurate AI models. GAN models and transfer learning contribute to filling knowledge gaps between human and animal data, while XAI models enhance the reliability of AI models by providing transparency into decision-making processes. This approach enables a more informed and comprehensive application of AI in navigating the intricacies of drug development.
Author Bio
Dr. Manu Kumar Shetty, MBBS, MD, is an Associate Professor and data scientist at Maulana Azad Medical College, New Delhi, he excels in patient care, drug discovery, and teaching medical professionals. Proficient in AI/ML, he has contributed to notable projects, including ECG-based AI prediction model and risk stratification AI model.
Anubha Gupta received her PhD. from IIT Delhi (2006) in Electrical Engineering and has more than 30 years of work experience. Currently, she is a full Professor at IIIT- Delhi and had earlier served as the Dean (Academic Affairs), and a member, Board of Governors of IIIT Delhi.
