Senior Data Scientist · PhD · Computational Biology

Amy
Francis

Data science, ML & drug discovery - and making these tools accessible to everyone.

I'm a Senior Data Scientist bridging computational biology and experimental research, with expertise in machine learning, bioinformatics, and tool development across cancer genomics, immunology, and drug discovery. I think a lot about how AI is changing science - and about making sure those changes reach every researcher, not just those with a computer science or machine learning background. My work tries to do that: building practical tools, leading interdisciplinary teams, and closing the gap between what these models can do and what most researchers can actually access.

Cancer Genomics Drug Discovery Protein ML RNA-seq Foundation Models AI in Science 2× Hackathon Winner

View projects Get in touch

About me

I try to read where science is heading,
and build the skills to meet it.

My background is wet-lab first. I built a CRISPR activation screening platform at the Crick Institute, working alongside bioinformaticians and attending seminars from some of the best scientists in the world. That exposure is what made the move into computation feel less like a leap and more like an obvious next step.

I'm a Senior Data Scientist at Nexus BioQuest, a contract research organisation in Portishead, Bristol, where I work across data science, machine learning, and analytical tool development in support of research programmes spanning pharmaceutical and biotech clients.

My PhD at the University of Bristol, funded by a competitive Cancer Research UK studentship, focused on predicting the functional impact of genetic variants in cancer genomes. I've since worked at Roche, exploring protein language models and antibody optimisation - and published peer-reviewed papers in variant effect prediction.

I've also led interdisciplinary teams to back-to-back hackathon victories at Cambridge and the Wellcome Collection, and co-organised Bristol's first AI in Health meeting, which secured two interdisciplinary research grants. The best science I've been part of has come from people with genuinely different mental models trying to solve the same problem. It generates questions that nobody in a single-discipline team would ever think to ask, and occasionally, those questions lead somewhere important.

Based in

Bristol, UK

Experience

5+ Years

Current role

Senior Data Scientist, Nexus BioQuest

PhD

Bioinformatics & ML, Bristol

PhD Funded by

Cancer Research UK

Interests

AI in science, open collaboration

A perspective on AI in science

AI is going to be part of everyday science.
I think the hard work now is getting these tools to the scientists who need them, including those who don't yet know they do, and that's where the next medicines will come from.

The tools exist, and the clinical evidence is starting to follow. The real challenge now is getting them out of research papers and into the hands of the scientists who could use them to find new medicines, and opening up possibilities for researchers who don't even know what they're missing yet.

"Some of the most powerful tools in the history of biology are sitting in research papers and GitHub repos that most bench scientists have never heard of. Building the bridges to get them there, through upskilling and through teams where computational and experimental scientists actually learn from each other, is how I think we find the next generation of medicines."

AlphaFold transformed protein structure prediction in 2021 - making accurate structural models accessible for hundreds of millions of proteins where before it had required years of experimental work. AI-designed drugs are now reaching clinical trials. The industry is adapting, with pharma companies partnering with specialist AI firms and embedding AI infrastructure directly into their R&D pipelines. The pace of change is accelerating.

But using these tools well still requires an unusual combination of skills - enough ML to run and adapt the models, enough compute to work with them, and enough domain knowledge to ask the right questions. Most biologists have one of those things, maybe two. Closing that gap - whether through upskilling or building teams that complement each other's strengths - is something I think the field needs to take seriously. I came into data science from the biology side, so I know what it feels like to have a question you can't answer because the tools are out of reach. That's what drives most of what I build and write about.

The deeper analysis - the partnerships, what the clinical evidence actually shows, how infrastructure investments could lead to more breakthroughs, and the honest open questions about whether any of this produces better drugs - is in the blog.

Foundation models reshaping biology

AlphaFold 2 & 3

DeepMind / Google

Predicted structures for 200M+ proteins by 2022. AlphaFold 3 (2024) extends to DNA, RNA, and small molecules - relevant to structure-based drug design.

ESM-2 & ESMFold

Meta AI

Protein language models trained on 250M sequences. Enable zero-shot fitness and function prediction - models I've used directly in my own research at Roche.

Evo 1 & 2

Arc Institute

Genomic foundation models trained across the tree of life on DNA sequences. Enable generative design of genes, regulatory elements, and CRISPR guides.

RFdiffusion

Baker Lab, UW

Diffusion model for de novo protein design - binders, enzymes, vaccine candidates - designed from scratch and experimentally validated.

BioNeMo

NVIDIA

An open platform wrapping biological foundation models - ESMFold, DiffDock, MolMIM - behind accessible APIs. Used by Amgen, Genentech, AstraZeneca, GSK, and Novo Nordisk. Over 100 firms on the platform as of 2024.

DiffDock & MolMIM

NVIDIA / MIT

DiffDock predicts protein-ligand binding poses using diffusion - faster than traditional docking. MolMIM generates novel small molecules optimised for target properties.

Geneformer

Broad Institute

Transformer pre-trained on 30M single-cell transcriptomes. Supports in silico gene perturbation and network inference - useful for target identification without wet-lab experiments.

scGPT

University of Toronto

Single-cell foundation model pre-trained on 33M cells. Enables cell type annotation, perturbation prediction, and gene regulatory network inference.

TxGNN

Harvard / Broad

Graph neural network trained on biomedical knowledge graphs for drug indication and contraindication prediction - a practical tool for repurposing existing compounds.

Pharma.AI / INS018_055

Insilico Medicine

End-to-end AI drug discovery platform. Used to design INS018_055, the first fully AI-generated drug to show efficacy in a Phase IIa trial (Nature Biotechnology, 2024). Target to Phase I in under 30 months.

Projects & Hackathons

Things I've built and worked on.

A mix of published tools, hackathon projects, and pipelines built for real research problems.

🏆 1st Place · Cambridge

Mapping novel compounds to biological pathways

Led the winning team at GetSeen Ventures' AI × Cancer Bio Hackathon. Used transformer encoders on SMILES strings and high-content image embeddings from the RxRx3-core dataset to predict molecular pathways. Ongoing collaboration likely to result in publication.

TransformersSMILESImage EmbeddingsDrug Discovery

🏆 1st Place · Wellcome Collection

Deep learning for protein fitness prediction

Led the winning team at the Roche & HDR Hackathon. Encoded protein sequences with pre-trained language models (ESM, AntiBERT) and explored CNNs to model sequence-function relationships using DMS data from Protein Gym. Secured a Roche AI internship as a direct result.

ESMAntiBERTCNNsDMSPyTorch

Read opinion piece

Alan Turing Institute

Toxicity prediction for drug discovery: Data Study Group with Ignota Labs

Participated in the Alan Turing Institute's Data Study Group in collaboration with Ignota Labs, working on machine learning approaches to predict drug toxicity. Contributed to the team summary and published technical report.

PythonMLDrug DiscoveryToxicity Prediction

Team summary Full report

Flow Cytometry

Flow cytometry analysis pipeline

Built a post-acquisition flow cytometry analysis pipeline with an intuitive Streamlit interface for processing and visualising high-dimensional cytometry data, enabling efficient downstream reporting across client projects.

PythonStreamlitScikit-learn

Cancer Genomics

DrivR-Base - variant annotation toolkit

Published a data mining toolkit integrating molecular annotations for SNVs, creating a centralised resource that reduces redundancy and accelerates machine learning model development for variant effect prediction.

PythonRDockerCOSMICGnomAD

Read paper

Antibody Optimisation

Predicting mutation impact on antibody-antigen binding

At Roche pRED, used TensorFlow models grounded in global epistasis and pre-trained protein language models to predict binding affinity from deep mutational scanning data.

TensorFlowESMHPCDMS

Global Epistasis repo

Mentoring & Leadership

Automating ELISA analysis at Nexus BioQuest

Currently developing a pipeline to automate ELISA data processing and reporting, reducing manual effort across client projects. Currently mentoring a team member through the project as part of their coding development - turning it into a hands-on introduction to Python and scientific data pipelines.

PythonPandasStreamlitMentoringIn Progress

Community

Bristol AI in Health Meeting

Co-organised Bristol's first interdisciplinary AI in Health Meeting in collaboration with the Elizabeth Blackwell Institute. Facilitated cross-disciplinary collaboration that resulted in two interdisciplinary grants for applied AI projects.

LeadershipEvent OrganisationGrant Facilitation

Skills & Tools

The tools I work with.

Picked up across academia, industry, and a few hackathons.

Languages & Systems

Python
R
SQL
Linux / HPC
Cloud platforms
Docker

Machine Learning

Scikit-learn
XGBoost / SVMs
Neural Networks
Foundation & Language Models
TensorFlow / PyTorch
MLflow

RNA-seq & Transcriptomics

Count matrices: Cell Ranger
Quantification: featureCounts, DESeq2
Differential expression: edgeR, limma
Unsupervised ML for cell population definition
Immunology cell type characterisation
UMAP / t-SNE dimensionality reduction
Pathway & gene set enrichment analysis

Bioinformatics & Data

Flow Cytometry
Deep Mutational Scanning
CRISPR / Image Analysis
COSMIC, GnomAD, TCGA
Protein / DNA Sequences
Proteomics (Olink)

Visualisation & Comms

Streamlit
Matplotlib / Seaborn
Scientific writing
Client consultation
Conference presenting

Leadership & Collaboration

Interdisciplinary team leadership
Hackathon team lead (2× winner)
Cross-functional collaboration
User-centred design
Grant facilitation
CRO client management

Publications

Published research.

Three peer-reviewed papers and one technical report, spanning cancer genomics, variant effect prediction, and drug discovery.

Journal Article · 2026

CanDrivR-CS: A Cancer-Specific Machine Learning Framework for Distinguishing Recurrent and Rare Variants

Bioinformatics Advances - Accepted & Published

doi.org/10.1093/bioadv/vbag008

Application Note · 2024

DrivR-Base: A Feature Extraction Toolkit For Variant Effect Prediction Model Construction

Bioinformatics - Accepted & Published

doi.org/10.1093/bioinformatics/btae197

Review Article · 2023

Predicting Pathogenicity from Non-Coding Mutations

Nature Biomedical Engineering - Accepted & Published

doi.org/10.1038/s41551-022-00996-x

Technical Report · 2024

Toxicity Prediction for Drug Discovery

The Alan Turing Institute, Data Study Group · Not peer-reviewed

doi.org/10.5281/zenodo.13882192

AmyFrancis

I try to read where science is heading,and build the skills to meet it.

AI is going to be part of everyday science.I think the hard work now is getting these tools to the scientists who need them, including those who don't yet know they do, and that's where the next medicines will come from.