It’s intuitive for many of us to think of genomes or protein sequences as "languages." This analogy is easy to grasp, as both DNA and protein sequences can be represented as long strings of characters—much like sentences in a book. These sequences follow patterns, convey information, and, just like language, can be decoded to reveal their meaning. This "language" perspective has fueled the development of large-scale AI models that can process and interpret these biological sequences, much like how natural language models (such as GPT) analyze text.

But what about molecular data that isn’t string-like, such as tabular data? For example, single-cell transcriptomics data, which captures the gene expression levels of thousands of genes across millions of individual cells, comes in the form of large, complex tables. At first glance, this type of data doesn't appear to fit neatly into the "language" analogy. However, foundational AI models are rapidly emerging in this space, with single-cell RNA sequencing (scRNA-seq) data leading the charge.

Models like **scGPT**, **scTab**, and **scFoundation** are pioneering efforts to convert these massive expression tables into a form that can be treated as "language." These models, inspired by language models like GPT, are pre-trained on millions of single-cell data points and then fine-tuned for specific downstream tasks such as cell type annotation and perturbation prediction. By doing so, they unlock new possibilities for interpreting the complexity of single-cell data. Recently highlighted in prestigious journals like Nature, these models are demonstrating how AI can extract meaningful insights from highly intricate biological datasets.

In conclusion, while it may seem unconventional to treat tabular biological data as a "language," transformers— the technology behind many AI models—offer unparalleled advantages in handling vast datasets with complex interdependencies. The rise of scRNA-seq foundational models, like scGPT, showcases how AI is revolutionizing multi-omics data interpretation. As these models continue to evolve, they promise to drive significant insights across diverse types of biological data, transforming the way we understand the molecular world.

Read this post on LinkedIn >