GET is a new model for understanding human biology; China is advancing in AI despite US restrictions; AI users on social media is not a totally dumb idea; Google wants to simulate the physical world

AI startups drive VC resurgence; five real-world early uses of AI agents; how AI regulation will shake out in 2025; Sam Altman's big Bloomberg interview; AI is bringing the end of search as we know it

Jan 10, 2025

∙ Paid

DNA and RNA are two essential molecules that store and process genetic information in all living cells. They work together to create the essence of who we are: while DNA carries the genetic information we need to survive (essentially acting as an instruction manual for our body), RNA copies that information so it can be used to make proteins in our body.

One reason why scientists have long sought to decode these two molecules is because most cancers are caused by glitches in our DNA that cause problems in our proteins later down the line. Luckily, RNA can give us clues about what might be wrong in the DNA, and therefore help us understand how the interplay of genes, proteins, and regulatory sequences gives rise to the diversity of human cells and their functions.

Every cell in your body contains the same DNA, yet a neuron behaves nothing like a liver or muscle cell. This specificity is orchestrated by transcriptional regulation—where proteins called transcription factors interact with DNA sequences to control gene expression and therefore turn genes on and off in different patterns in the messenger RNA.

Now, an ambitious new AI model called GET (General Expression Transformer) promises to rewrite the rulebook for transcriptional regulation, a cornerstone of biological processes.

Unveiled by researchers from MBZUAI, Columbia University and Carnegie Mellon, GET is not just another computational biology tool—it’s a leap forward in the form of a foundational model designed to predict gene expression across 213 human cell types.

Foundational models like OpenAI’s GPT have been pretty versatile, finding applications in fields such as natural language processing, image generation or coding but they fall short in more specialized domains such as computational biology. On the other hand, existing computational models like Enformer have made strides in predicting gene expression, but their limitations are stark: they often fail when tasked with new cell types or conditions.

GET applies the foundational model philosophy to the genomic landscape, and the results are pretty astonishing. The model can actually predict which regulatory elements are important for a particular gene with incredible accuracy. It’s like it can see the control panel for each gene and tell us which switches are flipped.

Extended Data Fig. 1 — GET’s model architecture

When the researchers put it to the test with something called lentiMPRA, a giant experiment where scientists test thousands of DNA sequences to see which ones act as on switches, GET predicted the results with surprising accuracy, without ever seeing the actual experimental data.

By relying on chromatin accessibility data and sequence information, GET achieves experimental-level accuracy in predicting gene expression—even for cell types it has never seen before. Think of it as a crystal ball for molecular biology, capable of anticipating how a gene will behave in a novel cellular environment. With GET, scientists may be able to predict the risk of certain diseases appearing or discover new regions in fetal cells which will help them understand and chart human evolution.

Extended Data Fig. 3 — **Transfer learning adapts GET to new platforms and cell types**

The model adapts to diverse data sources, sequencing platforms, and even non-physiological cell types like cancer cells. This adaptability makes GET a powerful tool for exploring the unknown. For example, the model identified long-range regulatory elements in fetal erythroblasts— those early cells that developed into red blood cells — that had eluded previous models probably because these regions are over a million base pairs away. In B cells, it unveiled a transcription factor interaction linked to leukemia risk, shedding light on genetic predispositions to disease.

GET also provides a comprehensive catalog of regulatory grammars, transcription factor interactions, and potential gene-regulation mechanisms for over 200 cell types. Such insights could guide everything from the study of noncoding genetic variants to the development of synthetic biology applications.

By incorporating additional biological layers such as three-dimensional chromatin architecture and transcription factor binding data, the model could achieve even greater precision. Researchers envision using GET to design cell-type-specific enhancers, transcription factors, or even therapeutic interventions for diseases like cancer.

The researchers behind GET have made their catalog publicly accessible, inviting the scientific community to explore this wealth of regulatory information. As genomic data continues to grow exponentially, tools like GET will be indispensable in translating this deluge of information into actionable insights.

These AI-powered advances in biology are a clear sign that we’re soon going to live in a future where computational genomics doesn’t just help us understand life’s blueprint but actively rewrites it, offering tailored solutions to genetic diseases or enabling the creation of synthetic cell types for biomedical applications.

And now, here are the week’s news:

❤️Computer loves

Our top news picks for the week - your essential reading from the world of AI

Keep reading with a 7-day free trial

Subscribe to Computerspeak by Alexandru Voica to keep reading this post and get 7 days of free access to the full post archives.