Daniele Pessina
I build modelling software for virtual patients and toxicology prediction. Mechanistic simulators, machine learning and biological data, wired into one differentiable system that trains end to end and runs in production.
Currently
Virtual patients & in-silico clinical trials
I build the software behind virtual patients: multiscale models of human physiology coupled to ML layers, calibrated against heterogeneous data and exposed as workflows scientists can query.
Much of it is making legacy scientific code behave like modern ML code. I wrap C++ ODE engines, SBML models, optimisation solvers and agent-based simulations as differentiable PyTorch components, then build the gradient backends, surrogates and analysis tools around them. Once a model trains end to end, it can be stress-tested under uncertainty and reduced to the mechanisms that drive the output, which makes it usable for experiment design and in-silico decisions.
What I build
At Deep Origin I build the differentiable pipelines behind virtual patients, and the software around them: surrogates, gradient backends, analysis and deployment code. In my PhD I work on neural ODEs for sparse crystallisation time-series, hybrid pharmacokinetic models that fuse clinical data with neural components, and probabilistic inference in JAX and Julia. At Haleon I shipped ML soft-sensors that read product quality to within 10% of lab reference, enough to retire up to half of offline QA; at Quaisr, agentic LLM workflows for regulated pharma.
Across all of it I lean on model interpretation and design: sensitivity analysis, SHAP and PDP/ICE, uncertainty quantification, model reduction, design of experiments. That work is packaged in gsax, my open-source JAX library for global sensitivity analysis of large dynamic models.
Doctoral research
Alongside Deep Origin I'm finishing my PhD at the Sargent Centre for Process Systems Engineering, Imperial College London, supervised by Maria Papathanasiou and Jerry Heng on a CASE studentship with AstraZeneca. The thesis builds differentiable, uncertainty-aware models of biomolecule crystallisation that combine mechanistic simulation, machine learning and probabilistic inference to guide experiments and speed up process development.