Research Engineer F/M – Aligning medical foundation models across histology, spatial transcriptomics and MRI

Context
The position is in the ARAMIS project-team (Inria, CNRS, Inserm, Sorbonne Université), at the Paris Brain Institute (ICM) on the Pitié-Salpêtrière hospital campus. ARAMIS develops computational and statistical methods for multimodal patient data in brain disorders. The engineer will be supervised by Reuben Dorent (Inria research scientist), whose work focuses on multimodal representation learning and statistical inference from incomplete medical data.
Assignment
Assignments :
With the help of Reuben Dorent, the recruited person will be taken to:
Build a reproducible pipeline that extracts and stores representations from several foundation models per modality (histology, spatial transcriptomics, MRI) on public multimodal datasets;
Set up an evaluation protocol for alignment from few pairs, using datasets in which the true correspondences are known but only a few hundred of them are revealed during training;
Extend the team's two-modality alignment method [1] to a first setting with three modalities, and compare it with existing baselines;
Support a PhD student of the team working on histology-spatial transcriptomics integration, by providing shared tools for data preprocessing, feature extraction and evaluation.
For a better knowledge of the proposed research subject :
[1] J. Kordes, R. Dorent. Representation Alignment from a Few Paired Data via Distilled Optimal Transport. arXiv:2610.12781, 2026.
[2] S. Ruyter, R. Dorent et al. Program-space Diffusion for Morphology-to-Transcriptomics Prediction. COMPAYL workshop, MICCAI 2026.
[3] J. Ostermaier, S. Ruyter, R. Dorent et al. Disentangled shared representations improve morpho-transcriptomic integration. COMPAYL workshop, MICCAI 2026.
Collaboration :
The recruited person will be working closely with a PhD student of the ARAMIS team.
Responsibilities :
Write data loaders and preprocessing for public datasets such as HEST-1k (paired histology and spatial transcriptomics), TCGA (MRI, histology and clinical data), Ivy GAP and the Allen Human Brain Atlas.
Extract embeddings from several pretrained foundation models per modality (e.g. pathology, single-cell / spatial transcriptomics and 3D MRI models) and organise them in a common, documented format.
Implement the evaluation protocol: train/test splits that reveal only a few hundred pairs, retrieval and matching metrics, and robustness to the choice of paired samples.
Implement and run baselines and the extension of the team's alignment method to three modalities; analyse the results.
Package the code (tests, documentation, example notebooks) so that it can be reused by the PhD students of the team and released as open-source software.
Take part in team meetings, present results and contribute to the writing of technical reports and, where results allow, scientific publications.
Main activities
Analyse and implement related work.
Design novel innovative solutions.
Write progress reports and papers.
Communicate advancement to the team.
Skills
Technical skills and level required :
Solid Python programming and experience with PyTorch.
Practical experience in deep learning; knowledge of representation learning or self-supervised learning.Good software practices: git, code review, testing, documentation, use of GPU clusters (SLURM).
Languages :English
Relational skills : Good communication skills
Other valued appreciated :