UNIVERSAL SINGLE-CELL TRANSCRIPTOMIC AGING CLOCK POWERED BY LLMS REVEALS TARGETS TO SLOW CELLULAR AGING

Sep 29, 2026

Upload by Cart

Large language models such as GPT are adapted to biology by converting single-cell transcriptomes into ordered gene-name sequences, treated as “cell sentences.” Pretrained language models are fine-tuned on diverse single-cell data to learn a universal latent representation of aging for biological age prediction. Trained across tissues and life stages, the framework captures cell-type-specific trajectories and conserved molecular signatures, generalizing across tissues and species. The clock achieves state-of-the-art single-cell accuracy and highlights aging-associated processes, while nominating gene targets whose modulation lowers predicted transcriptional age for therapeutic prioritization.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages1

Reading Time3 min

AccessFree to download

Abstract

Method: LLM adaptation for single-cell sequencing

Model capabilities: universal representation and generalization

Outputs: biological age prediction, embeddings, and target genes

Innovation in Aging, 2025, Vol. 9, No. S2 1291 Abstract citation ID: igaf122.4305 UNIVERSAL SINGLE-CELL TRANSCRIPTOMIC AGING CLOCK POWERED BY LLMS REVEALS TARGETS TO SLOW CELLULAR AGING Raghav Sehgal, Yale University, New Haven, Connecticut, United States Large language models such as GPT have shown impressive performance in capturing structure and meaning from natural language. We adapt this capability to biology by representing single-cell transcriptomes as ordered sequences of gene names, analogous to sentences in text. Each cell’s expression profile is converted into a “cell sentence,” enabling pretrained language models to be fine-tuned directly on single-cell data for age prediction. Trained on millions of cells spanning diverse tissues and life stages, the model learns a universal latent representation of aging. It captures both cell-type–specific trajectories and conserved molecular signatures, generalizing across tissues and species. Unlike other foundational models that must be trained from scratch on omics data, our framework leverages already well-validated LLMs, taking advantage of their robust priors while avoiding the cost and instability of building new architectures. The resulting clock achieves state-of-the-art accuracy in predicting biological age at single-cell resolution. Embeddings highlight aging-associated processes such as immune activation, proteostasis decline, and metabolic shifts, aligning with known hallmarks of aging. Importantly, the model can also identify genes whose modulation decreases predicted transcriptional age, nominating candidate targets for therapeutic intervention. These features make the approach valuable both for measuring aging and for prioritizing interventions that may slow or reverse it. This work establishes the first universal, multi-tissue aging clock at single-cell resolution, demonstrating that pretrained language models can be directly adapted to biological data to map, measure, and potentially modify cellular aging.

FAQs

How are single-cell transcriptomes represented for use with language models?

Single-cell transcriptomes are represented as ordered sequences of gene names, analogous to sentences, forming a “cell sentence” for each cell.

What enables the model to predict aging across tissues and species?

Training on millions of cells spanning diverse tissues and life stages allows the model to learn a universal latent representation that captures both cell-type trajectories and conserved molecular signatures, supporting generalization.

How does the framework identify potential therapeutic targets to slow cellular aging?

The model can identify genes whose modulation decreases predicted transcriptional age, nominating candidates for therapeutic intervention.

Sep 29, 2026

Upload by Cart