The current challenges of AI in pharma
Recently, through my work on applying the latest AI models in drug discovery and discussions with friends using AI in the pharmaceutical space, I’ve made several observations I’d like to share:
✅ Exciting Breakthroughs in Academia: AI has led to significant breakthroughs, particularly in academia. For example, AlphaFold has revolutionized protein structure prediction, and ESM-3 has enabled the design of novel fluorescent proteins. These groundbreaking achievements from top labs have made the pharmaceutical and biotech industries eager to apply AI to the costly, hit-or-miss, and time-consuming drug discovery process, which often exceeds $1 billion. The hope is that AI can improve the odds of success in this complex field.
✅ The Need for Experimental Validation: AI’s potential cannot be fully realized without experimental validation. While it’s possible to run computer simulations in drug discovery, the real impact comes when an AI-developed drug is experimentally proven to work and generate value. Anyone with basic AI knowledge can build models using existing datasets, but creating a streamlined AI model that effectively reduces experimental cycles and labor requires careful planning and a deep understanding of the problem at hand. AI can inspire ambitious visions, but without a realistic grasp of the model’s capabilities, dataset limitations, and the actual problem, these visions may fall short.
✅ High-Quality Data and Well-Defined Problems: AI trained on high-quality datasets for well-defined problems can make relevant predictions. However, questions often arise about the sufficiency of data points for an accurate machine learning model. Key factors influencing AI success include the cost of generating high-quality data, selecting the appropriate AI models to answer the right questions, and knowing when to trust a model’s predictions. In the early stages of drug discovery, where AI is used to reduce costs and increase the chances of success, deprioritizing promising candidates can be more detrimental than prioritizing ineffective ones.
✅ The AI Dilemma in Pharma: AI in pharma faces a unique dilemma. When processes are unknown or expensive, there’s hope that AI can provide relevant answers. However, AI, particularly generative AI, often produces more results than can be feasibly analyzed. Without a efficient method to filter good candidates for further development, valuable opportunities may be missed. Some have tried to extrapolate known drug targets to unexplored territories, hoping to identify unknown targets through patterns that AI can model. However, this approach often leads to two problems: (1) the AI consistently identifies known targets, which doesn’t contribute to novel drug development, and (2) it produces irrelevant results because many factors influencing target identification are specific to the landscape (disease/application). In other words, ensuring AI consistently outperforms empirical methods in project-specific tasks is challenging, even though successful results are often highlighted in literature.
✅ The Ultimate Goal - Helping Patients: Ultimately, drug development is about helping patients and improving their quality of life. The further a drug candidate progresses towards clinical trials, the more successful it is considered. While clinical success can be statistically calculated, AI’s value lies in providing additional evidence for known targets, improving patient recruitment strategies, or identifying the right patient types for trials. There are methods that integrate data from various sources (omics, genetics, literature), but the explanatory power of AI is still evolving. Clinicians are more interested in understanding why a drug target can cure a disease, such as cancer, rather than just identifying potential targets for treatment—many of which are already well-documented in GWAS and gene expression analyses.
The above might change with the rapid development of AI and the understanding gained in the space. Special thanks to friends within Novartis and connections outside for the insight. If you have a different perspective and would like to share, please feel free to reach out to me.