TARJINDER SAHOTA 1
1 Glaxosmithkline (Stevenage , United Kingdom)
Objectives
Hybrid AI/ML-pharmacometrics approaches are increasingly discussed for drug development [1], but productive collaboration is often slowed by terminology gaps between pharmacometricians and AI/ML experts. Recent literature in the space has shifted from proof of concept toward operationalisation of workflows and reporting to enhance rigor and transparency [2]; complementary attention is needed to reduce communication-based friction during cross-functional execution. Unlike long-established working patterns with statisticians and other quantitative disciplines, many pharmacometricians have limited exposure to AI/ML language, conventions, and buzzwords (e.g., digital twins, virtual populations). Because AI/ML terminology is more widely used across industry, it is pragmatic for pharmacometricians to develop AI/ML literacy rather than expect AI/ML colleagues to learn pharmacometric terminology. This literacy also helps pharmacometricians contextualise emerging AI/ML advances, separating genuinely new and useful capabilities from renamed familiar ideas or methodological innovations that do not materially advance the clinical decision. While AI/ML and pharmacometrics differ methodologically, many core concepts share structural parallels; making these explicit reduces cognitive load and accelerates interdisciplinary reasoning. The objective of this work is to provide (i) a concise, pharmacometrics-oriented glossary that maps common AI/ML terms to familiar NONMEM concepts where appropriate, and (ii) an opinionated communication proposal describing how pharmacometricians can act as effective translational interfaces between clinical development teams and AI/ML experts to accelerate deployment of hybrid methodologies on projects.
Methods
An AI/ML-pharmacometrics glossary was iteratively developed within GSK based on direct experience of cross-functional interactions in emerging hybrid initiatives, focusing specifically on recurrent communication failure modes observed when engaging with AI/ML specialists (e.g., different meanings attached to the same word; “false friends”; ambiguous buzzwords). Candidate terms were selected to cover the typical lifecycle of AI/ML work (data representation, optimisation/training, evaluation/generalisation, and deployment/inference) and then translated into pharmacometrics-accessible language using: (1) nearest conceptual analogues in NONMEM and NLME workflows where they exist, (2) explicit “non-equivalence” notes where analogies are tempting but misleading, and (3) plain-language definitions intended for rapid uptake by pharmacometricians. The glossary and proposed team-working rubric were refined through internal presentation and feedback from pharmacometricians and AI/ML subject matter experts.
Results
The primary output is a working glossary of >35 terms, designed for pharmacometricians collaborating with AI/ML subject matter experts. The glossary includes direct mappings that enable rapid anchoring of AI/ML concepts to established pharmacometrics intuition, including loss function ↔ objective function value (OFV); epoch ↔ iteration during parameter estimation; backward propagation ↔ parameter gradient calculation during estimation; and out-of-distribution generalisation ↔ extrapolation to a new dosing regimen, population, or clinical setting. It also explicitly addresses high-risk “false friends” and buzzwords that can derail early discussions. For example, digital twins is clarified in pharmacometrics terms as individualised prediction conditioned on posthoc ETA information ($EST MAXEVAL=0), while virtual populations is distinguished as simulation-based generation of plausible individuals ($SIM).
Beyond the glossary itself, this work proposes a pragmatic operating model for hybrid teams: pharmacometricians can serve as a translation layer between the clinical teams they sit on and AI/ML implementation details, provided they build working literacy in AI/ML terminology and workflow constructs (training/validation/test splits; hyperparameter optimisation; generalisation and out-of-distribution behaviour). Early internal usage at GSK suggests the glossary functions as a useful shared reference to align expectations and reduce cross-disciplinary miscommunication during project framing.
Conclusions
Terminology mismatch with AI/ML experts is a practical barrier that can prevent productive conversations, slow alignment, and reduce stakeholder engagement when deploying hybrid AI/ML methodologies in drug development. A curated glossary that (i) maps AI/ML terms to pharmacometrics analogues when justified, (ii) flags non-equivalences and false friends, and (iii) demystifies buzzwords, provides a low-cost, high-leverage first step toward efficient cross-functional deployment. Given pharmacometricians’ familiarity with clinical development decision-making and their established role in cross-functional quantitative teams, developing AI/ML language fluency positions pharmacometricians to act as effective translational interfaces, improving collaboration with AI/ML specialists while keeping modelling work anchored to clinical questions and project priorities.
References:
[1] Terranova N, Renard D, Shahin MH, Menon S, Cao Y, Hop CECA, Hayes S, Madrasi K, Stodtmann S, Tensfeldt T, Vaddady P, Ellinwood N, Lu J. Clinical Pharmacology & Therapeutics. 2024;115(4). doi:10.1002/cpt.3053
[2] Fochesato A, Brooks L, Bazgir O, Pierrillas PB, Jamois C, Lu J, Mercier F. CPT: Pharmacometrics & Systems Pharmacology. 2026;15:e70113. doi:10.1002/psp4.70113
Reference: PAGE 34 (2026) Abstr 12152 [www.page-meeting.org/?abstract=12152]
Poster: Methodology – AI/Machine Learning