Scoring gene importance by interpreting single-cell foundation models

2026-05-27 · Source: Machine learning : nature.com subject feeds · Field: Science & Research — Life Sciences & Biology, Health & Medical Research, Artificial Intelligence & Machine Learning · Depth: Expert, extended

Summary

SIGnature is a novel framework designed to score gene importance in single-cell RNA-sequencing (scRNA-seq) data, addressing the inherent unreliability of absolute gene expression levels. This framework adapts attribution methods from explainable AI (XAI) and applies them to single-cell foundation models (FMs), quantifying each gene's influence on a cell's position within the model's latent space. The associated SIGnature package facilitates rapid gene set searches across extensive scRNA-seq atlases, capable of processing 22 million cells in minutes. The method effectively reduces technical noise, prioritizes regulatory genes like transcription factors, and enables robust cross-dataset comparisons. Its practical utility was demonstrated by identifying novel associations between the MS1 monocyte signature, known for its activation in severe COVID-19 and sepsis, and previously unassociated hyperinflammatory conditions such as Kawasaki disease (KD) across 400 studies. Experimental validation further confirmed that serum from KD patients can induce the MS1 phenotype in vitro.

Key takeaway

For research scientists analyzing single-cell RNA-sequencing data, adopting the SIGnature framework can significantly improve the identification of functionally important genes and robustly compare gene signatures across diverse datasets. You should consider integrating attribution-based scoring to overcome technical noise and uncover shared disease mechanisms, as demonstrated by its success in linking the MS1 signature to Kawasaki disease and other hyperinflammatory conditions. This approach offers faster, more reliable insights than traditional expression-based methods.

Key insights

SIGnature uses XAI attributions on scRNA-seq foundation models to robustly score gene importance and enable cross-dataset biological discovery.

Principles

• Attribution scores reduce technical noise.
• Attributions highlight regulatory genes.
• FMs enable cross-dataset comparisons.

Method

SIGnature selects a scRNA-seq FM, computes attributions using gradient-based XAI (e.g., Integrated Gradients, Input x Gradient, DeepLIFT) with a summation layer for multidimensional embeddings, then aggregates scores for gene set analysis.

In practice

• Query gene signatures across 22 million cells.
• Identify shared disease mechanisms.
• Generate testable biological hypotheses.

Topics

• Single-cell RNA-sequencing
• Foundation Models
• Explainable AI
• Gene Importance Scoring
• Cross-disease Analysis
• Kawasaki Disease

Code references

• Genentech/SIGnature

Best for: AI Scientist, Research Scientist, Machine Learning Engineer

Related on AIssential

• Toward informed batch correction for single-cell transcriptome integration — Batch effects in single-cell data hinder comparisons; future methods need informed, interpretable modeling.
• Discovery of candidate therapeutic targets with Geneformer — Geneformer leverages a transformer model and transfer learning for gene network analysis and therapeutic target discovery with limited data.
• Probabilistic modelling of single-cell bisulfite sequencing data with MethylVI — MethylVI disentangles biological signals from technical noise in single-cell bisulfite sequencing data.
• A new tool is revealing the invisible networks inside cancer — RNACOREX reveals hidden gene networks in cancer, predicting survival with AI-level accuracy and interpretable molecular explanations.
• Machine Learning-Based Pre-Test Risk Stratification for PCR-Confirmed Chlamydia Using Patient-Reported Data and Urine Biomarkers — Urine biomarkers offer a reliable, consistent signal for Chlamydia pre-test risk stratification, complementing patient-reported data.
• Accelerating discovery of liver disease mechanisms — AI-driven literature analysis can uncover novel biological hypotheses and candidate therapies for complex diseases.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Machine learning : nature.com subject feeds.