Introduction
In a groundbreaking study published in the July 2026 issue of Science (Volume 393, Issue 6806), researchers have introduced TranscriptFormer, a generative artificial intelligence model capable of constructing a comprehensive cell atlas that spans 1.5 billion years of evolution. This innovative approach leverages deep learning to map cellular states across diverse species, offering unprecedented insights into the evolutionary origins of cell types and gene regulatory networks.
What Is TranscriptFormer?
TranscriptFormer is a generative AI model designed to predict and reconstruct gene expression profiles across a wide range of organisms. By training on vast datasets of single-cell transcriptomics, the model learns the underlying patterns of gene regulation that are conserved or divergent across evolutionary time. The result is a dynamic cell atlas that not only catalogues known cell types but also predicts ancestral and intermediate cellular states.
The model's architecture is based on transformer networks, similar to those used in large language models like GPT, but adapted for biological sequence data. It processes gene expression data as a language, where each gene's expression level is akin to a word in a sentence, and the cellular context provides the grammar. This allows TranscriptFormer to generate plausible expression profiles for cells that have never been experimentally observed, effectively filling gaps in our evolutionary understanding.
Key Findings
The study demonstrates that TranscriptFormer can accurately reconstruct cell types from organisms separated by up to 1.5 billion years of evolution, including animals, plants, fungi, and protists. The model identifies core gene regulatory programs that are universally conserved, as well as lineage-specific innovations that gave rise to complex tissues and organs.
One of the most striking findings is the prediction of a common ancestral cell type that likely existed in the last eukaryotic common ancestor (LECA). This hypothetical cell type exhibits a combination of features seen in modern stem cells and immune cells, suggesting that the earliest eukaryotes possessed a primitive form of cellular differentiation.
Additionally, TranscriptFormer reveals how certain gene regulatory networks have been repurposed across evolution. For example, genes involved in neural development in animals show similarities to genes controlling stress responses in plants, indicating deep evolutionary connections between seemingly unrelated biological processes.
Methodology
The researchers compiled a massive dataset of single-cell RNA sequencing data from over 100 species, representing major branches of the eukaryotic tree of life. This data was used to train the TranscriptFormer model in a self-supervised manner, where the model learned to predict masked gene expression values based on the surrounding context.
To validate the model's predictions, the team compared them to experimentally determined cell atlases from species not included in the training set. TranscriptFormer achieved high accuracy in reconstructing known cell types and also identified novel cell states that were later confirmed through targeted experiments.
The model's generative capabilities were further tested by simulating evolutionary scenarios, such as the transition from unicellular to multicellular life. TranscriptFormer successfully predicted intermediate cell types that bridge the gap between single-celled ancestors and complex multicellular organisms, providing a computational framework for studying evolutionary transitions.
Implications for Biology and Medicine
TranscriptFormer has far-reaching implications for both basic biology and applied medicine. By providing a comprehensive view of cellular diversity across evolution, the model can help identify conserved genes and pathways that are critical for cell function. This knowledge can inform the development of new therapies for diseases that involve cellular dysfunction, such as cancer and degenerative disorders.
Moreover, the generative nature of TranscriptFormer allows researchers to explore 'what-if' scenarios, such as how a cell might respond to genetic perturbations or environmental changes. This could accelerate drug discovery by predicting off-target effects or identifying novel drug targets.
The study also opens up new avenues for evolutionary developmental biology (evo-devo), enabling scientists to test hypotheses about the origin of cell types and the genetic changes that drove major evolutionary innovations.
Limitations and Future Directions
While TranscriptFormer represents a major advance, the authors acknowledge several limitations. The model's predictions are only as good as the training data, and biases in species representation or experimental conditions could affect accuracy. Additionally, the model does not account for epigenetic modifications or post-transcriptional regulation, which play important roles in cell identity.
Future work will focus on integrating multi-omics data, including chromatin accessibility and protein levels, to create a more holistic view of cellular states. The researchers also plan to expand the atlas to include more species, particularly those from understudied branches of the tree of life.
Conclusion
TranscriptFormer marks a new era in computational biology, where generative AI can reconstruct the evolutionary history of cells with remarkable fidelity. By spanning 1.5 billion years of evolution, this cell atlas provides a unifying framework for understanding cellular diversity and the principles of gene regulation. As the model continues to evolve, it promises to transform our understanding of life's cellular foundations.
This article is based on reporting by Science (AAAS). Read the original article.
Originally published on science.org
