Artificial intelligence (AI) is big business in the pharmaceutical industry. According to Deep Pharma Intelligence, cumulative investments in AI-related drug development between 2014 and 2023 topped $60 billion. With the recent $1 billion launch of Xaira Therapeutics, and seemingly endless new deals between various pharmaceutical companies and NVIDIA, $60 billion seems to be a low projection for 2024 spending.

Niven R. Narain, PhD

Alongside the PR hype, however, there are real conversations taking place among industry leaders questioning whether the money flowing into AI will meaningfully improve productivity and output in the industry. While biotechnology companies have been touting a decade-old narrative about AI drug discovery being exponentially faster and cheaper than conventional drug discovery, so far these companies have put only a few drugs into clinical trials, and none have made it through Phase III and FDA approval yet.

Early high-profile clinical failures last year, including Exscientia’s Phase I/II study of its cancer drug EXS-21546 and BenevolentAI’s Phase II study of its dermatitis drug BEN-2293, suggest the return on investment for AI in drug development may be further delayed.

The conversation around AI in the pharmaceutical industry seemingly portrays AI tools as a shortcut to clinic-ready compounds. This exaggerates the power of AI to master something as diverse and complex as human biology.

AI tools alone aren’t a shortcut to a drug in the way that ChatGPT might be a shortcut to developing a term paper comparing the existentialist philosophies of Sartre and Camus. For us, the existential question is not, “Will AI revolutionize drug discovery?” AI undoubtedly will once it is applied properly. Instead, we need to ask, “How do we eliminate the translation problems that have plagued some of the early AI-derived compound trials?” If we could answer that question, we would be in a better position to validate AI-derived findings.

Why AI hasn’t lived up to the hype

Much of the excitement around AI in the pharmaceutical industry centers around the use of AI tools to identify new drugs by using sophisticated computer algorithms to crunch through publicly available datasets. Conventional wisdom suggests that the computer can’t be wrong. But if that were the case, why are we seeing challenges in moving from the dry lab to the clinic?

The answer likely lies in the data itself and in what companies do with their AI-derived findings. AI models are excellent at identifying correlations, but as we all know, correlations do not explain causation. The success of generative AI methods using data from public datasets is dependent upon the accuracy and completeness of the datasets used to train the AI models.

There is no evidence suggesting that we have yet fully digitized human biology. Even the most accurate and complete datasets can at best get researchers to identify proper correlations. Simply put, more data is needed. Companies developing drugs need to understand causation, which requires going back to the wet lab to validate AI-derived findings.

Real biological samples and longitudinal studies

In the present land grab for access to AI, there are few AI platforms focused on the use of real biological samples to feed their AI models, and even fewer using real biological samples in longitudinal studies to produce data. Few companies and groups are pioneering Bayesian AI versus more traditional machine learning models to derive insights from their samples.

The value of initiating research with samples from pre- and post-longitudinal disease samples and using a Bayesian approach is that it offers hypothesis-free discovery and holds the potential to redefine the conceptualization, discovery, and development of drugs. Neural AI can be used in concert as a next step to decode the intricate relationships between genetic factors and common diseases, aiding crucial decision-making about drug
development pathways.

With the use of real biological samples taken from the same patients at different times, AI can help researchers go beyond preset hypotheses and the traditional try-and-fail approach, and truly understand the causation of diseases and guide us to new discoveries. By validating pharmacological approaches in real biological samples either prior to preclinical testing or as part of their efforts to understand the results of clinical trials, AI helps us not only make novel discoveries faster, but provide imperative insights to ensure clinical trial success.

Leveraging these approaches, we’ve identified trial populations for therapeutic assets that have shown early promise in clinical-stage studies for difficult-to-treat cancers including glioblastoma multiforme and pancreatic cancer.

What success with AI tools really looks like

Despite the early failures that have raised skepticism about the value of AI to pharmaceutical development, I remain bullish on the opportunity ahead. Rich, comprehensive, and free of human bias, AI tools can, with the support of wet lab validation and preclinical translational models, bring us closer to precision medicine and provide value across the value chain of pharmaceutical development.

As widely predicted, AI tools can help us identify compounds for clinical development, but we must leverage real biological inputs—that is, inputs other than those from public datasets—and rigorously validate findings derived from AI models to ensure that we understand the underlying mechanisms of action in our therapeutic candidates, and what types of patients will be most likely to benefit. From there, we can design clinical trials to include only those patients likely to benefit.

Once we run a clinical trial, we can collect and analyze clinical samples using AI modeling of the patient’s biology before and after treatment. Insights derived from this modeling can help us better understand the biological effects of our therapeutic candidate and further refine our understanding of which types of patients respond to treatment, and which do not.

Finally, AI-derived insights can help direct a path toward label expansion or drug repurposing once a therapeutic is approved, by leveraging understandings of the mechanism of action and responding population characteristics and identifying other patient populations with similar biological characteristics.

Niven R. Narain, PhD, serves as the CEO of BPGbio

Biology-First Artificial Intelligence

BPGbio uses its NAi Interrogative Biology AI platform to operate a clinically annotated, longitudinal, 100,000-plus patient/sample biobank, and the company’s researchers have taken tissue, blood, and urine samples and subjected them to metabolomic, lipidomic, proteomic analyses. BPGbio officials say that analyzing integrated multiomics data with domain-specific AI models enables the company to better understand the underlying biology of the diseases for which it is designing therapies, and the biological changes that occur when its therapeutic candidates are administered.

Previous articleFinding “Impossible” Antibodies in the Human Body
Next articleUpstream Bioprocessing: More Intense and More Stable