Mendel, a foundation model of human genetic variation, prioritizes regulatory variants and improves gene expression prediction
Mendel is a novel DNA foundation model trained with a variant-centric objective that outperforms existing models like Evo 2 in predicting allelic events, effectively prioritizing regulatory variants and significantly improving gene expression prediction across human tissues.
Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer
The human body is built from a vast, intricate instruction manual written in a four-letter chemical code. For decades, scientists have been reading this manual to find the tiny typos—single letter changes in the DNA sequence—that are linked to diseases, physical traits, and how our bodies respond to the world. These discoveries have been made by comparing the DNA of millions of people, spotting patterns where a specific letter change often appears alongside a specific health outcome. However, finding the link is only the beginning. The real challenge lies in understanding what that letter change actually does. Does it turn a gene on or off? Does it break a machine? Most of these typos sit in the vast, non-coding regions of the genome, areas that do not directly build proteins but instead act as the switches and dimmers that control when and where genes are active. Figuring out the mechanical consequences of these regulatory typos has remained a stubborn bottleneck in genetics, leaving researchers with a long list of suspects but no clear picture of their crimes.
To solve this, a team of researchers has developed a new type of artificial intelligence called Mendel. Unlike previous models that were trained to understand the general grammar of DNA by reading the entire sequence of many different species, Mendel was built with a different goal in mind. It was trained from scratch specifically on the DNA of human beings, with a focus entirely on the differences between individuals. Imagine a student who has memorized a dictionary and the rules of grammar but struggles to understand why two people might spell the same word differently. Mendel is like a student who has spent years studying those specific spelling differences, learning how the surrounding letters influence whether a change is harmless or significant. By concentrating its learning on the spots where human DNA actually varies, Mendel learned to predict what genetic code a person should have at a specific location based on the sequence around it. When tested on people it had never seen before, the model predicted these genetic variations with 98.4% accuracy, a massive leap over existing tools that achieved only about 31% accuracy on the same task.
The true power of Mendel, however, lies not in its ability to predict the common, expected genetic code, but in the mistakes it makes. When the model gets a prediction wrong, that error is not random noise; it is a signal. The researchers found that the model struggled most at locations in the genome that are known to regulate gene activity, specifically at sites called expression quantitative trait loci, or eQTLs. These are the very spots where a genetic change is known to alter how much of a protein a cell produces. While older models treated these important regulatory sites just like any other part of the DNA, Mendel's difficulty in predicting them revealed their functional importance. The harder the model found it to guess the correct genetic code at a specific spot, the more likely that spot was to be a key regulator of gene expression. This "residual difficulty" acted as a natural highlighter, pointing directly to the variants that matter most for human biology.
The researchers then put this insight to work in a practical test. They used the model's difficulty scores as a guide to improve predictions of gene expression across 50 different human tissues, such as the liver, brain, and heart. In every single tissue, weighting the genetic variants based on how hard Mendel found them to predict resulted in more accurate predictions of gene activity than using standard methods. This improvement held true even when compared against other sophisticated tools that rely on pre-existing databases of harmful mutations. The study suggests that by training an AI to focus on the sparse, individual differences that make us unique, rather than the shared background of our species, scientists can uncover a new layer of functional information. Mendel does not replace the existing models that understand the broad language of life; instead, it complements them by offering a specialized lens tuned specifically to the variations that define human diversity and disease.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.