AI Model Predicts Single-Cell Protein Levels Directly from Gene Expression Data
A new large generative AI model infers single-cell proteomes from transcriptomes, unlocking richer disease and aging biology at lower cost.
Summary
Researchers at Tencent AI and City University of Hong Kong developed scTranslator, a large pre-trained generative AI model that predicts protein abundance in individual cells using only gene expression (RNA) data. Measuring proteins at the single-cell level is critical for understanding disease and aging, but current proteomic technologies are expensive, limited in coverage, and technically demanding. Inspired by both language translation and the central dogma of biology, scTranslator bridges this gap. The model was validated across multiple experimental platforms, diverse cell types, and tissues including blood, lung, and brain. It also demonstrated strength in downstream tasks such as identifying cancer cell origins, predicting responses to biological perturbations, and correcting batch effects. This tool could dramatically accelerate research into age-related diseases by making proteome-level insights accessible without requiring costly proteomic experiments.
Detailed Summary
Understanding how proteins behave at the level of individual cells is one of the most powerful windows into disease and aging. Proteins are the functional molecules that drive cellular processes, yet measuring them at single-cell resolution remains technically challenging, expensive, and prone to noise and batch variation. A new AI model called scTranslator aims to solve this by predicting single-cell protein abundances directly from gene expression (transcriptomic) data.
The model was developed by researchers at Tencent AI Lab and City University of Hong Kong. Drawing inspiration from both natural language processing and the biological central dogma — where DNA is transcribed to RNA and translated into protein — scTranslator functions as a generative translator between two molecular languages. It is pre-trained on large biological datasets and fine-tuned to generalize across experimental contexts.
The team validated scTranslator across multiple single-cell proteomic platforms including CITE-seq, spatial CITE-seq, REAP-seq, and NEAT-seq. Performance was benchmarked across biologically diverse cell types such as monocytes, macrophages, T cells, and B cells, and across tissues including blood, lung, and brain. Crucially, the model held up across disease contexts spanning infectious disease, metabolic conditions, and cancer.
Beyond raw prediction, scTranslator proved useful for a range of downstream applications: inferring gene-protein interaction networks, predicting cellular responses to perturbations, improving cell clustering, correcting batch effects, and identifying the tissue of origin for cancer cells in pan-cancer datasets. These capabilities are directly relevant to aging research, where multi-omic integration is essential.
For longevity science, the implications are significant. The ability to infer proteome-level information from cheaper, widely available transcriptomic data could democratize multi-omic aging research. Caveats include the fact that transcription and translation are not always tightly coupled, and the abstract alone was available for this review.
Key Findings
- scTranslator accurately infers single-cell protein abundance from RNA data alone across multiple platforms and tissues.
- The model generalizes across metabolic, infectious, and oncologic disease contexts without retraining from scratch.
- It improves downstream tasks including batch correction, cell clustering, and cancer cell origin identification.
- Perturbation prediction capability enables in silico testing of how cells respond to biological interventions.
- Reduces reliance on costly single-cell proteomic experiments by leveraging widely available transcriptomic data.
Methodology
scTranslator is a pre-trained large generative AI model benchmarked systematically on independent datasets from multiple single-cell multi-omic platforms (CITE-seq, REAP-seq, NEAT-seq, spatial CITE-seq). Validation spanned diverse cell types, tissues, and disease conditions. Downstream analytical tasks including perturbation prediction and pan-cancer cell origin recognition were used to assess practical utility.
Study Limitations
The summary is based on the abstract only, as the full paper is not open access. A key biological caveat is that mRNA levels do not always reliably predict protein abundance due to post-transcriptional regulation, protein degradation, and other factors, which may limit prediction accuracy in specific contexts. Competing interests exist, as several authors are Tencent employees and hold related patents, warranting independent replication.
Enjoyed this summary?
Get the latest longevity research delivered to your inbox every week.
Enter your email to subscribe: