Raphaël Rubrice

Helmholtz Munich · Institute of Computational Biology

Raphaël Rubrice

Hi, I'm Raphaël.
I work on
AI for biology.

More precisely, I'm an aspiring computational-biology researcher. Right now I work on AI-assisted experimental design, helping decide which experiments to run next, and I keep an eye on one stubborn problem: how to help our models not fall apart under biological contexts they haven't seen. I'm a research assistant at Helmholtz Munich in the Fabian Theis Lab, and in January I start my PhD, beginning in Montréal at Mila.

Research interests

I'm interested in problems working towards virtual experimental biology.

The dream is models good enough that a lot of experimental work could start in silico: faster wet-lab campaigns, AI in the loop when deciding which experiments to run, and maybe a bit more scientific discovery along the way. The virtual cell is one avenue toward that.

I suspect AI-assisted experimental design is where progress is likeliest in the next few years, before the harder generalization questions are solved, given the amount of effort being put into it.

What also bugs me is how badly our models do out of distribution. Put them in a context or condition they weren't trained on and they often fail, in ways I'd argue are worse than how we humans cope. Faced with something new, we lean on what we already know, reason about how it's similar to or different from the situation in front of us, and we're often roughly right. We're far from perfect, sometimes we grab the wrong prior and do worse than chance, but the gap between that and how a model behaves in an unfamiliar biological context feels like something worth trying to close. So I'm trying to keep one foot in each.

AI-assisted experimental designhelping navigate the design space
The virtual cell & out-of-distributionmodels that hold up in unfamiliar biological contexts

Work

A few things I've worked on, different angles trying to get structure out of biological or clinical data, and now, to decide which experiments to run next.

ELLIS UnConference · NGDL Workshop2026 · Master's thesis · Theis Lab, Helmholtz Munich★ Accepted · poster, Palais des Congrès de Paris, 8 Dec

Xiaotong Fu, Alessandro Palma, Lorenzo Consoli, Fabian Theis, Guy Wolf

BANmyBED: Bayesian experimental design for lab-in-the-loop CRISPR screens
Deciding which genes to test next in a CRISPR screen, and letting the scientist steer.

CRISPR screens are costly, so choosing which genes to test next, round after round, is an experimental-design problem. Policies trained on past screens help, but they transfer imperfectly to a new screen and can't easily be steered when the scientist's goal changes. BANmyBED is a batched, amortized design policy that can be steered at inference time, between more hits and more diverse ones, together with a Bayesian correction that adapts a frozen policy to the screen at hand as outcomes arrive. It's the core of my Master's thesis; a public version will follow at the NGDL workshop.

AI in the Life Sciences · Elsevier2025 · Master 1 project · UMR MIA Paris-Saclay & Micalis

Rubrice R., Guéneau V., Briandet R., Cornuéjols A., Guigue V. (2025). A machine-learning framework for the prediction and analysis of bacterial antagonism in biofilms using morphological descriptors. AI in the Life Sciences, Elsevier.

Predicting bacterial antagonism from biofilm morphology


Can you tell whether a beneficial Bacillus strain will out-compete a pathogen from the morphology of single-species biofilms alone, without exhaustive co-culture screening? Supervised by Antoine Cornuéjols and Vincent Guigue (AgroParisTech, UMR MIA Paris-Saclay), and working with Romain Briandet and Virgile Guéneau from the biofilm team at INRAE's Micalis Institute, we built a first framework for it. My part was a homology-controlled cross-validation scheme so the model couldn't cheat on closely related strains; together we used SHAP and permutation importance to see which morphological features actually drive exclusion.

Dassault Systèmes20256-month internship · Clinical Decision Team

Aziliz Cottin, Marine Zulian

Explainable clustering for clinical data
Helping clinicians read what a clustering result actually means.

A six-month research internship on making clustering results interpretable: when patients fall into subgroups, finding ways to explain what actually separates them, with a focus on small cohorts where standard methods struggle. This one is proprietary and under NDA, so can't say more sorry :(

Owkin & Servier2025★ AI Methodology Award

Jhonathan Felix-Ramos, Adrien Bouchet, Timothée Sanchez, Jean Radig

Extending THREADS for glioblastoma
A multimodal idea that taught us how hard the problem is.

We proposed extending THREADS, a multimodal foundation-model approach, to bring together the many modalities in the MOSAIC glioblastoma dataset. In the short time we then had after the event to actually finish the project, our results were honestly quite poor, and that was the instructive part: integrating that many modalities across different biological scales, in a very low-N setting with noisy signals, is genuinely hard. The heads converge at different rates, the modalities carry different noise levels and structures, and naive integration just doesn't hold up.

AI4AS Conference · ETH Zurich2025 · IODAAOral presentation

Tran G., Genin R., Lauront A., Petitet M., Rubrice R., Heuzé V., Guigue V., Cornuéjols A.

Predicting forage nutritional value with AI
Estimating energy and protein from chemical composition and textual data.

Can AI accurately predict the energy and protein content of forages from chemical descriptors and natural-language metadata, without expensive in vivo trials? Working with AgroParisTech (Antoine Cornuéjols, Vincent Guigue) and animal nutrition specialists (Véronique Heuzé, Gilles Tran), we built models combining chemical composition data with textual descriptions of forage samples. The project was selected for an oral presentation at the 1st EAAP Conference on Artificial Intelligence in Animal Science (AI4AS), ETH Zurich, June 2025.

IRCAN, Nice20243-month internship · Bachelor project · Cristofari Lab

Sophie Lanciano, Gaël Cristofari

TELLAM: reading the repeats genomes usually discard
An RNA-seq pipeline for retrotransposon expression in senescent cells.

A bioinformatics pipeline to quantify retrotransposon expression in senescent cells: the repetitive-element signal that standard RNA-seq workflows often discard, but that may matter for how cells age.

Distinctions

Unaite Fellowship for AI in Biology, inaugural cohort, Paris
2026-27
Impact+ Doctoral Scholarship, Canada Impact+ Research Training Awards
2026
PR[AI]RIE-PSAI Institute Scholar, PR[AI]RIE Paris School of AI scholarship
2025
AI Methodology Award, Owkin & Servier glioblastoma hackathon
2025
Regional Excellence Scholarship, Guadeloupe Region
2019

Volunteering

French (native) · English, C2 proficient

Trajectory

From wet to dry lab.

Each step pulled me toward more ML without ever leaving the biology behind. Last year focused on the maths side, and I think that mix is worth it.

01
2020-2022

Prépa BCPST · Lycée Masséna, Nice

Two years of France's competitive science track (mathematics, physics, chemistry, biology) before the grandes écoles entrance exams.

02
2022-2026

Engineering degree, life sciences · AgroParisTech

The national institute for life sciences and industries. I specialized in health engineering, and met biostatistics and machine learning for the first time.

03
2024-2025

MSc Computational Biology · Université Paris-Saclay

Highest Honors. Where bioinformatics merged with modeling: learning theory, graph theory, deep learning, gene-expression analysis.

04
2025-2026

Master MVA · ENS Paris-Saclay & Université Paris-Cité

Highest Honors. Applied Maths and Computer Science, one of France's most demanding programs. Health track.

05
2026

Master's thesis · Helmholtz Munich

In the Theis Lab at the Institute of Computational Biology, supervised by Xiaotong Fu, Alessandro Palma, Lorenzo Consoli, Fabian Theis and Guy Wolf. The thesis is BANmyBED, Bayesian experimental design for lab-in-the-loop CRISPR screens. I stay on as a research assistant at Helmholtz until the PhD starts.

06
Jan 2027 →

Incoming PhD, applied mathematics · Université de Montréal / Mila & TUM / Helmholtz Munich

Joint PhD between Université de Montréal / Mila, supervised by Guy Wolf, and TUM / Helmholtz Munich, supervised by Fabian Theis. It starts in January 2027, with the first year in Montréal at Mila.

Now

More science ahead.

Research assistant at Helmholtz Munich, presenting BANmyBED at the ELLIS UnConference in Paris on December 8, then starting my PhD in Montréal in January, with Munich as the second home. Always glad to talk experimental design, single-cell, generative models, or AI for biology.