Charlie Harris (Charles)

University of Cambridge PhD Student @ University of Cambridge

I’m finishing a PhD at Cambridge, where I develop generative and geometric deep-learning methods for structure-based drug discovery in the labs of Pietro Lio and Sir Tom Blundell.

Until recently, I served as Technical Lead in the UK Government’s Sovereign AI Unit. Working closely with the Prime Minister’s AI Adviser, Matt Clifford, I helped shape early thinking on AI for Science — particularly around the idea that targeted, high-quality datasets can be a practical lever for unlocking “AlphaFold-scale” advances.

In that role, I led the £8m seed investment to establish OpenBind, a consortium aimed at building large-scale protein–ligand datasets for AI-driven drug discovery, and supported a £5m DSIT expansion of the Encode AI for Science Fellowship. I also helped design the GPU compute allocation approach set out under Sovereign AI Compute, influencing how millions of GPU-hours were distributed, and contributed to the UK’s AI for Science Strategy.

Google Scholar GitHub Twitter CV
Education
  • University of Cambridge

    University of Cambridge

    PhD in Computer Science Oct. 2021 - Present

  • Imperial College London

    Imperial College London

    MSc in Bioinformatics and Theoretical Systems Biology Oct. 2020 - Sep. 2021

  • Imperial College London

    Imperial College London

    BSc in Biochemisty Oct. 2017 - Jun. 2020

Experience
  • UK Government

    UK Government

    Technical Lead (Opportunties), UK Sovereign AI Unit Mar. 2025 - Dec. 2025

  • IQ Capital

    IQ Capital

    Venture Fellow June 2024 - Oct. 2024

  • BenevolentAI

    BenevolentAI

    Machine Learning Research Intern July 2022 - Oct. 2022

Honors & Awards
  • UK-Italy Visiting Researcher Fellowship - Alan Turing Institute 2024
  • Polaris Fellowship - Entrepreneur First 2024
  • Cambridge Half Blue (for gliding) - Univesity of Cambridge 2023
  • CCAIM PhD Studentship - University of Cambridge 2021
  • Associateship - Royal College of Science 2020
  • Gold - UK Chemistry Olympiad 2017
Media
  • UK Foundation for Science and Technology Podcast 2024
  • YouTube Video on AlphaFold 2024
  • iGEM Synthetic Biology Podcast 2021
News
2025
Published the UK's AI for Science Strategy. Read more
Nov 28
We announced SovAI's £8M investment into OpenBind, the largest drug discovery dataset in history. Read more
Jun 09
SynFlowNet was presented at ICLR 2025
Apr 05
2024
DiffSBDD is published in Nature Computational Science!
Nov 01
Selected Publications (view all )
TABASCO: A Fast, Simplified Model for Molecular Generation with Improved Physical Quality
TABASCO: A Fast, Simplified Model for Molecular Generation with Improved Physical Quality

Carlos Vonessen*, Charles Harris*, Miruna Cretu*, Pietro Liò (* equal contribution)

Transactions on Machine Learning Research (TMLR)(Camera ready pending) 2026

We introduce TABASCO which re- laxes these assumptions: The model has a standard non-equivariant transformer architecture, treats atoms in a molecule as sequences and re- constructs bonds deterministically after genera- tion. The absence of equivariant layers and message passing allows us to significantly simplify the model architecture and scale data through- put. On the GEOM-Drugs benchmark TABASCO achieves state-of-the-art PoseBusters validity and delivers inference roughly 10× faster than the strongest baseline, while exhibiting emergent ro- tational equivariance despite symmetry not be- ing hard-coded.

[Paper] [Code]

TABASCO: A Fast, Simplified Model for Molecular Generation with Improved Physical Quality
TABASCO: A Fast, Simplified Model for Molecular Generation with Improved Physical Quality

Carlos Vonessen*, Charles Harris*, Miruna Cretu*, Pietro Liò (* equal contribution)

Transactions on Machine Learning Research (TMLR)(Camera ready pending) 2026

We introduce TABASCO which re- laxes these assumptions: The model has a standard non-equivariant transformer architecture, treats atoms in a molecule as sequences and re- constructs bonds deterministically after genera- tion. The absence of equivariant layers and message passing allows us to significantly simplify the model architecture and scale data through- put. On the GEOM-Drugs benchmark TABASCO achieves state-of-the-art PoseBusters validity and delivers inference roughly 10× faster than the strongest baseline, while exhibiting emergent ro- tational equivariance despite symmetry not be- ing hard-coded.

[Paper] [Code]

DiffSBDD: Structure-based Drug Design with Equivariant Diffusion Models
DiffSBDD: Structure-based Drug Design with Equivariant Diffusion Models

Arne Schneuing*, Charles Harris*, Yuanqi Du*, Arian Jamasb, Ilia Igashov, Weitao Du, Tom Blundell, Pietro Lió, Carla Gomes, Max Welling, Michael Bronstein, Bruno Correia (* equal contribution)

Nature Computational Science 2024

DiffSBDD was one of the first equivariant diffusion models for structure-based drug design.

[Paper] [Code]

DiffSBDD: Structure-based Drug Design with Equivariant Diffusion Models
DiffSBDD: Structure-based Drug Design with Equivariant Diffusion Models

Arne Schneuing*, Charles Harris*, Yuanqi Du*, Arian Jamasb, Ilia Igashov, Weitao Du, Tom Blundell, Pietro Lió, Carla Gomes, Max Welling, Michael Bronstein, Bruno Correia (* equal contribution)

Nature Computational Science 2024

DiffSBDD was one of the first equivariant diffusion models for structure-based drug design.

[Paper] [Code]

SynFlowNet: Design of Diverse and Novel Molecules with Synthesis Constraints
SynFlowNet: Design of Diverse and Novel Molecules with Synthesis Constraints

Miruna Cretu, Charles Harris, Ilia Igashov, Arne Schneuing, Marwin Segler, Bruno Correia, Julien Roy, Emmanuel Bengio, Pietro Liò

ICLR 2025 Spotlight

This work introduces SynFlowNet, a GFlowNet model whose action space uses chemically validated reactions and reactants to sequentially build new molecules. We evaluate our approach using synthetic accessibility scores and an independent retrosynthesis tool. SynFlowNet consistently samples synthetically feasible molecules, while still being able to find diverse and high-utility candidates.

[Paper] [Code]

SynFlowNet: Design of Diverse and Novel Molecules with Synthesis Constraints
SynFlowNet: Design of Diverse and Novel Molecules with Synthesis Constraints

Miruna Cretu, Charles Harris, Ilia Igashov, Arne Schneuing, Marwin Segler, Bruno Correia, Julien Roy, Emmanuel Bengio, Pietro Liò

ICLR 2025 Spotlight

This work introduces SynFlowNet, a GFlowNet model whose action space uses chemically validated reactions and reactants to sequentially build new molecules. We evaluate our approach using synthetic accessibility scores and an independent retrosynthesis tool. SynFlowNet consistently samples synthetically feasible molecules, while still being able to find diverse and high-utility candidates.

[Paper] [Code]

PoseCheck: Generative Models for 3D Structure-based Drug Design Produce Unrealistic Poses
PoseCheck: Generative Models for 3D Structure-based Drug Design Produce Unrealistic Poses

Charles Harris, Kieran Didi, Arian Jamasb, Chaitanya Joshi, Simon Mathis, Pietro Liò, Tom Blundell

MLSB Workshop @ NeurIPS 2023 Spotlight

This work introduced PoseCheck, an extensive analysis of multiple state-of-the-art methods and find that generated molecules have significantly more physical violations and fewer key interactions compared to baselines, calling into question the implicit assumption that providing rich 3D structure information improves molecule complementarity. We make recommendations for future research tackling identified failure modes and hope our benchmark will serve as a springboard for future SBDD generative modelling work to have a real-world impact.

[Paper] [Code]

PoseCheck: Generative Models for 3D Structure-based Drug Design Produce Unrealistic Poses
PoseCheck: Generative Models for 3D Structure-based Drug Design Produce Unrealistic Poses

Charles Harris, Kieran Didi, Arian Jamasb, Chaitanya Joshi, Simon Mathis, Pietro Liò, Tom Blundell

MLSB Workshop @ NeurIPS 2023 Spotlight

This work introduced PoseCheck, an extensive analysis of multiple state-of-the-art methods and find that generated molecules have significantly more physical violations and fewer key interactions compared to baselines, calling into question the implicit assumption that providing rich 3D structure information improves molecule complementarity. We make recommendations for future research tackling identified failure modes and hope our benchmark will serve as a springboard for future SBDD generative modelling work to have a real-world impact.

[Paper] [Code]

All publications