Actionable Decision-Making for Digital Twins in Health and Life Sciences

In this invited talk, Peter Coveney (University College London, Centre for Computational Science), introduced by Abani Patra (Tufts University), examines the theoretical and computational foundations of digital twins in health and life sciences. Coveney frames digital twins as a virtuous circle between models and observed reality, aiming ultimately at a personalized digital twin of the human. He stresses that actionable, trustworthy predictions require validation, verification, and uncertainty quantification (VVUQ), distinguishing parametric and aleatoric uncertainty and warning of the curse of dimensionality that afflicts high-parameter models. He describes open-source tooling including the EasyVVUQ toolkit and FabSim3 for job submission, and collaborations with DOE labs including the Oak Ridge Leadership Computing Facility, using exascale systems Frontier and Aurora. Turning to molecular dynamics, Coveney argues that single simulations are non-reproducible because the dynamics are chaotic and sensitive to random seeds and thousands of force-field parameters; ensembles of roughly 20 or more runs are needed to control uncertainty in binding free-energy calculations. Scalable UQ methods based on a nonlinear principal component analysis reveal that uncertainty is often concentrated in fewer than about ten parameters. He critiques AI methods such as AlphaFold 2, whose hundred-million connection weights lack physical insight and whose predictions are valuable hypotheses but not actionable. He presents a drug discovery workflow combining generative AI (developed with AstraZeneca), physics-based methods, and experiment, which yielded nanomolar-potent molecules against the WDR91 protein, and points toward in-silico clinical trials and cardiotoxicity testing on virtual patients.