Finetuning Foundation Models for Temporal Clinical Transcriptomics Data

In plain English

The team adapted pretrained gene representations to follow how patients' gene-expression programs changed during treatment.

What they found

• The authors report that the models recovered known drug-response mechanisms and identified differences in activated and repressed pathways.
• In ulcerative colitis, the reported differences included B-cell activation and mitochondrial activity.

The catch

• The publisher abstract does not report cohort sizes or complete benchmark metrics.
• This is the accepted manuscript and may change during copyediting and typesetting.

Methods at a glance

Study type

Computational analysis of longitudinal clinical transcriptomics

Biological system

Treatment response in four inflammatory diseases

Organism

Human

Dataset

Longitudinal responder and non-responder transcriptomics for ulcerative colitis, Crohn's disease, alopecia areata, and psoriasis

Sample size

Not reported

Model architecture

Fine-tuned gene embeddings with temporal graph neural networks

Training objective

Model temporal gene-expression differences between treatment responders and non-responders

Code

Code and data are available from the Sanofi-Public/GNN-Timeseries repository cited by the publisher.

Data

The publisher links a public code and data repository.

Tasks and metrics

Tasks: Recover known response mechanisms; Identify pathway differences between responders and non-responders

Metrics: Complete quantitative metrics are not reported in the publisher abstract

Baselines: Traditional methods described in the accepted manuscript

Principal findings

• The authors report stronger detection of known response mechanisms than traditional methods across four disease datasets.

What remains unresolved

External validation: Not reported

• The abstract does not provide sample sizes.
• Complete quantitative comparisons require the accepted manuscript.