Integrating Artificial Intelligence into Pharma R&D
Michael N. Liebman, Managing Director and Co-Founder of IPQ Analytics, LLC
Artificial intelligence (AI) is transforming the pharmaceutical industry by enhancing drug discovery and optimizing clinical trial processes. With the potential to improve efficiency and accuracy in research and development, AI-driven technologies present both significant opportunities and challenges.
1. Which challenges in Pharma R&D can AI most effectively address today, and how do you foresee its role evolving to tackle emerging complexities, such as regulatory demands or personalized medicine, in drug development?
This seemingly simple question actually opens to addressing the underlying complexities in both pharmaceutical development and the use of AI methods.
AI represents a broad umbrella under which reside many technologies including natural language processing and text mining (NLP), machine learning (ML), knowledge graphs (KG) and generative AI (genAI), among others.
NLP can be used to process clinical notes, published papers, patient-physician interactions with increasing fidelity but can still be limited until it can also “process the context”.
ML is striving to move towards causality but remains a powerful tool for correlative analysis and identifying complex patterns in large data sets, i.e. advanced statistics
KG present a potentially dynamic representation of the complex interactions between/among concepts and relationships but is limited in their potential to include the “unknown unknowns”.
Analogously, the primary “silos” of Pharma R&D, namely discovery, toxicology, pre-clinical, clinial trials, regulatory and commercialization, all have individualized needs and expectations as to what “AI” can address for them…
The key to AI addressing existing and evolving complexities will always require a human intervention to accurately assess and define what the critical question is needing an answer as a first step towards selecting the appropriate AI methods, if available, and achieving the desired goals.
2. From your experience, what are the primary regulatory and operational obstacles to integrating AI technologies in the pharmaceutical industry, and how can these challenges be overcome?
As noted, the regulatory challenges to integration of AI methods will vary across the R&D process. There are few obstacles to using AI to predict the 3-dimensional structure of a protein or the small molecule that might be used as an inhibitor. There can be challenges, however, when AI is used to select patient cohorts for clinical trials or even clinical trial sites because of the potential for inclusion of biases in development of the big data used for the analysis. The focus of regulatory agencies is primarily on patient safety, first, and efficacy, second, and such potential biases may result in impacting those priorities. The closer you come to actual patient care, the more stringent the evaluations need to be…and it also presents a challenge to regulatory agencies in training/hiring staff with the critical skills to enable identifying potential issues.
In terms of operationalizing AI methods, the first step is aligning the appropriate methods with the question that needs to be addressed and recognizing that it may be more complex than conventionally stated, i.e. need to look at root cause issues. AI methods are becoming more accessible and able to be implemented even without specialists and a significant challenge is that data scientists and engineers are being well trained in implementing the technology but not necessarily in the critical thinking that is needed to bridge between the real world problem and the accessible technology. We need to develop better “translators” who can bridge this gap.
3. How can the industry maintain transparency in AI-driven decision-making, ensuring that predictive algorithms are both interpretable and actionable by all stakeholders, including regulatory bodies, clinicians, and patients?
There are at least two key stages of transparency that need to be addressed: 1) transparency in the method and potential biases/limitations; 2) transparency in the data and potential biases/limitations. Correlative analysis of “big data” will always be limited to finding novel patterns in existing data that may include integration of non-equal data, i.e. data with the same label but different contextual meaning because of different lab tests, thresholds, even diagnostic criteria, etc. Most AI methods focus on correlative analysis of such big data and need to explore more causal relationships to provide the optimal basis for critical clinical (and pharmaceutical) decision making.
While these are the low-hanging fruit in this issue, there is a significant need to identify, evaluate and improve the understanding of disease, as a process, and the current limitations due to the complexities of the patient, of the disease and of clinical practice. Failure to adequately address these poorly aligned aspects of real-world medicine will remain a challenge that AI is not likely to be able to overcome.
4. What are the most pressing ethical considerations that arise with the use of AI in Pharma R&D, from data privacy to algorithmic bias, patient safety, and patient outcomes, and what proactive measures should the industry take to mitigate these concerns?
"Do no harm" is a core ethical principle which essentially means avoiding causing harm to others; it is a fundamental concept in medical ethics, particularly associated with the Hippocratic Oath and naturally extends to Pharma R&D. The challenge that arises when applying AI in R&D is the potential of unintended harm because of limited control over the underlying algorithms, e.g, deep learning, and the need to access “big data” without full understanding of potential biases and limitations in its generation.
Algorithmic approaches that identify correlative relationships in data sets frequently operate without adequate human clinical oversight until review of final results. The identification of relationships without the potential for evaluating the significance both of individual critical factors and of their interactions can yield high statistical significance but fail in clinical relevance. These may misdirect efforts aimed at identifying key clinical processes and their potential targets for drug development and this has been observed across multiple data sources, e.g. from imaging data to EHR’s.
These issues are bi-directional, transcending the gap between Pharma R&D and clinical medicine. The challenges, that impact the accuracy of diagnosis and increasingly include the use of AI methods for data analysis and clinical decision support, have the potential to limit the potential effectiveness of drug discovery and development. Similarly, the application of AI approaches into cohort and site recruitment as well as earlier in discovery as to target selection for drug development, are susceptible to the limitations resulting from data-driven modeling (DDM) vs model-driven analytics (MDA). The emphasis on aggregation of big data sets frequently overlooks key aspects, i.e. all data are not created equal even if they are labelled the same way! And this goes beyond the obvious risk that aggregating large data sets frequently results in the potential leak of personal identification data.
It is essential that a collaborative effort across industry, clinic, academia and regulatory groups focus on how to approach the testing and validation of new algorithms as well as how to improve the contextual annotation of critical laboratory data and clinical notes. One consideration could be the development of a testbed data set that has “known” defects, limitations and biases that could be used to assess new algorithmic methods and provide a basis for comparative analytics.
5. How can AI and pharmaceutical experts collaborate more effectively to bridge knowledge gaps, ensuring that AI technologies are deployed in a way that complements human expertise while driving innovation?
The issue remains as to knowing what the right question is that needs to be answered, not what technology do I have to apply without understanding the question. It is of great value to use AI to summarize existing knowledge, e.g. publications, blogs, etc, but the high value opportunity is for AI to focus on identifying gaps in understanding the true complexity of the disease, patient and clinical practice that will impact successful drug development. Significant emphasis on the silos presents challenges to accomplishing this…for example, AI is being used (AlphaFold) to generate 3-dimensional structures of proteins, i.e. targets. Then AI is being used to design, refine, and screen potential molecules, e.g. inhibitors/modulators, for these targets. But the true complexity of a successful drug requires an understanding if this is the correct biological target, i.e. more than validation in laboratory experiments on pathways, to true biological impact. This requires collaborative interactions with clinicians, etc to understand what may be challenges in their diagnosis, disease subtyping, patient stratification, etc so further support selection of the right target, not simply the most convenient, “structure-predicted” target.
6. AI is often lauded for its ability to process vast datasets at unprecedented speeds. However, are there any limitations or risks associated with the over-reliance on AI in early-stage research, such as drug candidate screening or biomarker discovery?
AI as it is being implemented to read/digest the literature can be very helpful in early stages of research but there needs to be greater transparency as to what it can and cannot do so that it does not become the driver of that research. Lack of knowledge of the training base, i.e. journal articles, pre-prints, etc, leads to unknown potential biases in the results of such studies. And the AI approach cannot address the issue of “unknowns”…not to solve them but to be able to simply identify what and where they are so that inadequate assumptions do not become the base upon which significant research resources are dedicated. These AI tools are powerful but it need to be very clearly transmitted as to what they can and cannot do and what they are and are not representing.
7. With AI increasingly being integrated into every stage of Pharma R&D, what strategies would you recommend for organizations to scale AI applications while maintaining compliance with both regional and global regulatory standards?
The standards are influx and not well synchronized or harmonized across regional and global regulatory groups, so this will require some degree of flexibility and even parallel develop/implementation especially for global pharma…simple example is the nuances of difference between GDPR and HIPAA.
8. As AI continues to evolve and integrate more deeply into Pharma R&D, what is your vision for the future of the pharmaceutical industry in the next 10-15 years? What role do you see AI playing in shaping the future of drug discovery, clinical trials, and patient care?
I have both great hope and trepidation about the future integration of AI into Pharma R&D. It is potentially a very powerful tool but only one tool that needs to be incorporated into the complex process of drug development. The ultimate need will remain the understanding and differentiating between what is known vs what is unknown, and in particular how to effectively deal with our truly limited understanding of disease as a process and not a state. The complexity that exists in the patient, in the disease and in clinical practice will always present challenges that need to be considered and incorporated in any comprehensive decision support environment, especially one that involves AI where there is an increasing tendency to accept its results without adequate evaluation and validation. The challenge will always be to first truly understand what the real critical question is that needs to be answered and whether AI is the right tool to provide the solution.
As AI continues to evolve and integrate more deeply into Pharma R&D, what is your vision for the future of the pharmaceutical industry in the next 10-15 years? What role do you see AI playing in shaping the future of drug discovery, clinical trials, and patient care?
Author Bio
Michael N. Liebman is currently working as a Managing Director, IPQ Analytics, LLC, has experienced both an academic and pharma/diagnostic career at Mt. Sinai, UPenn, Vysis, Wyeth, Roche where he has led programs and teams in Bioinformatics, Pharmacogenomics, Computational Biology, Cancer Biology. He has had senior advisory roles in PhARMA, HIMSS, IUPAC and leads IPQ in its international advanced analytics as a service (AAS) business in EU, China, Africa and Australia with an associated non-profit that is focused on critical issues in women’s health: infant/maternal morbidity and mortality and personalizing the perimenopause-.menopause transition and its impact on post-menopausal disease risk
