Daniele Pessina

Modeling & Optimization Scientist · Deep Origin

Scientific machine learning · mechanistic modelling · uncertainty

I build scientific machine learning tools for complex biological systems. At Deep Origin, I connect mechanistic simulations with trainable models to study patient-level toxicity. My PhD at Imperial focuses on uncertainty-aware modelling and experimental design.

Current work

Virtual patients & in-silico clinical trials

Cell Sims team · ARPA-H CATALYST program · PREDICTS consortium · Deep Origin

I’m building tools that connect different kinds of biological simulators with machine learning. The aim is to train patient-level toxicity models, examine how uncertain their predictions are, and make use of clinical evidence even when only aggregate trial results are available.

Trainable simulations: Made mechanistic models work together with ML predictors, including solvers that do not provide ordinary gradients.
Evidence from limited data: Developed virtual-cohort and uncertainty methods for modelling from aggregate trial summaries.
Understanding model behaviour: Used bifurcation analysis to find stability boundaries and changes in biological response.

Experience

2026 – now
Modeling & Optimization Scientist · Deep Origin
Cell Sims team · ARPA-H CATALYST · PREDICTS consortium
- Built a framework combining mechanistic simulations and machine learning, enabling patient-level toxicity models to train across multiple simulation engines.
- Made previously non-differentiable scientific models trainable and accelerated model training by approximately 15× through gradient methods, precomputation and surrogates.
- Developed virtual-cohort and uncertainty methods for patient-level modelling from aggregate clinical-trial summaries when individual patient records are unavailable.
- Applied bifurcation analysis to large biological models to identify stability boundaries and treatment-response transitions.
2022 – now
PhD Researcher, Process Systems Engineering · Imperial College London
Sargent Centre · CASE studentship with AstraZeneca · applied to pharmacokinetics and protein crystallisation
- Developed mechanistic ODE/PDE simulators, hybrid mechanistic-ML and pure ML models of dynamic systems from sparse, noisy time series.
- Uncertainty quantification and identifiability methods for pharmacokinetic and process models: profile likelihood and Bayesian inference (MCMC, variational inference, ABC).
- Methods and open-source implementations for fast global sensitivity analysis and model-based design of experiments.
- Closed-loop in-vitro / in-silico work: I was in the lab running the experiments, built the models on that data, and fed the analysis back into experimental and process design.
2026
Data Scientist Intern · Quaisr
London
- Agentic LLM workflows for regulated pharma process development.
2025
Consultant Data Scientist · Haleon
- ML soft-sensors reading product quality in real time to within 10% of lab reference, enough to retire up to half of offline QA.

Education

2022 – now
PhD, Chemical Engineering · Imperial College London
2018 – 2022
MEng, Chemical Engineering, 1st Class Honours · Imperial College London

Thesis: a hybrid in-silico / in-vitro approach to the design and optimisation of large-molecule crystallisation, supervised by Maria Papathanasiou and Jerry Heng at the Sargent Centre.

Open source

jaxgsa · JAX library for global sensitivity analysis of large dynamic models (Sobol, RS-HDMR, multi-output), hundreds of times faster than SALib on multi-output workloads. Try selected jaxgsa methods directly in your browser with the WebAssembly demo.

jaxhybridmodels · JAX/Equinox library for combining user-written ODE dynamics with trainable predictors, supporting bounded physical quantities and regular or irregular time-series experiments. Documentation

CriSTool.jl · Julia package for simulating batch crystallisation, fitting kinetic parameters, and propagating parameter uncertainty through population-balance models.