A Pragmatic Perspective on the Hype and Hope of Large Language Models in Life Science
Angela Radcliffe
Bristol Myers Squibb

arge language models (LLMs) have been used in drug discovery for some time now, but many in the life sciences industry are just becoming aware of Generative AI: a buzzword that has become increasingly pervasive in society. Frequently and inaccurately associated with OpenAI, the company responsible for creating ChatGPT among numerous tools in the expansive AI domain, this technology garners a wide array of responses: lauded as a transformative innovation at one end of the spectrum and decried as a potential societal threat at the other. This article explores the transformative potential of LLMs underpinning Generative AI, their possible impacts on drug development, the regulatory hurdles they face, and the pressing need for a balanced approach to their utilization. There is a serious need for a judicious blend of innovation, regulatory compliance, and education to reap the maximum benefits of these tools while managing potential risks such as a proliferation of racial bias in healthcare, safety concerns that might arise from the use of AI-enabled decision support, or the undermining of public trust in our healthcare institutions if data is improperly consented for use in training data sets.

Testing Innovation

Stakeholders in the drug development ecosystem must adopt a “test and learn” strategy when innovating with LLMs. As the democratization of this technology is currently in progress, it provides a unique opportunity to become “citizen computer scientists.” We must grasp this chance to generate hypotheses, test them using foundational models, and learn in an iterative process. Training a model using a data set from closed studies in a particular disease to discern how changes in protocol design may impact a specific patient cohort, allowing us to rapidly amend protocols for future trials, is a good sample use case for a test-and-learn methodology. Another example might be training a model to scan “failed” new molecular entities (NMEs) to explore untapped potential for use in a new therapeutic target, renewing our portfolios in unexpected ways. This approach will equip us to identify new applications of LLMs for our most pressing treatment challenges and to actively advocate for appropriate regulation where LLMs are most likely to impact the drug development lifecycle.


LLMs in the life sciences, and more specifically in the drug discovery and development continuum, face a nascent regulatory landscape. While US regulators (FDA) have issued guidance on the use of AI in healthcare, particularly for Software as a Medical Device (SaMD), and the European Union has proposed AI regulations, specific or meaningful guidance for LLMs has not yet been developed.

Regulatory implementation of LLMs faces several challenges, including difficulties in traceability, explainability, data privacy, and consent for data use in training models. To fully realize the potential benefits of LLMs, stakeholders across the healthcare ecosystem must collaborate with regulatory bodies to establish regulations that balance risks without stifling innovation. Until clear regulations emerge, we need to be smart about setting our own policies that hold us accountable to the ethical principles and values of our organizations. We need to set policies and standards that respect privacy and keep our data safe such as revisiting our data governance policies with respect to the types of data we will train our AI models with and setting standards that ensure we hold ourselves accountable to ethical methods including the use of adversarial debiasing to reduce model bias. This may mean restricting the use of open-source tools within our organizations as we find our footing. It may also require enabling new technical environments and governance guardrails to mitigate risk and maximize value. While these policies will be individualized to each organization, the promotion of trust and transparency must be common core.


We must not overlook the critical role of education in the successful adoption of LLMs. As part of this endeavor, we need to update our data and digital literacy playbooks to include new terminologies like vectors and to develop new skills like prompt engineering and FAIR (Findable, Accessible, Interoperable, and Reusable) data management practices. The skill gaps are different for different roles, but closing them through a deliberate upskilling effort can help us unlock the full potential of LLMs and shift our workforce from routine tasks to more valuable knowledge work aimed at solving the greatest drug development challenges of today.

Conclusion and Future Trends

The integration of LLMs and AI into modern clinical trials holds immense promise, poised to usher in transformative changes across the industry. These cutting-edge tools are forecasted to significantly alleviate the administrative strain on clinical research investigators and their teams, enhance patient recruitment for the most crucial trials, pinpoint and craft new drug targets, and fine-tune patient visit schedules. However, this technological metamorphosis is not without its inherent risks, demanding proactive and meticulous management. It is essential to consider potential biases and the dangers of perpetuating or exacerbating misinformation. The successful navigation of these complex challenges calls for concerted collaboration across the entire drug development ecosystem. This includes key stakeholders such as biopharmaceutical sponsors, research investigators, their affiliated institutions, Institutional Review Boards (IRBs), and regulatory agencies. To ensure the ethical, responsible, and ultimately beneficial implementation of these transformative tools, a judicious mixture of innovation, regulation, and education is vital. This multifaceted approach will serve as a guiding compass, paving the way for the utilization of these technologies in a manner that benefits humanity at large.