EVO 2 Machine Learning Models Introduce the Power of AI in the Battle of Diseases

From the little wooden frogs to the towering redwoods, to you and me, DNA drives all life on Earth. The DNA embedded in every cell of every organism acts as a kind of biological instruction manual containing all the genetic information needed to create life.

The process begins with transcription: DNA creates a copy of a portion of the code to produce RNA, a type of molecule that can catalyze biological reactions that represent information embedded in the DNA. In these reactions, proteins are synthesized and become live cells. Overall, this is known as the central dogma of molecular biology: DNA makes RNA, and RNA makes proteins.

A single strand of DNA can contain millions of pairs of nucleotides, which are molecular building blocks that carry genetic information. Additionally, a single strand of RNA can contain tens of thousands. There are practically countless ways in which nucleotides can be combined to become life. And the complexity of the combination is simply too much for the human mind to understand. But that's where AI comes in.

“Machine learning can bring together higher-order patterns from large datasets,” says Patrick HSU, assistant professor of bioengineering. “AI already does this in natural language, vision and robotics. Now we do this in biology.”

In February 2025, HSU and his collaborators released a machine learning model trained with over 9.3 trillion nucleotides. biorxiv Preprint server. HSU, called EVO 2, can compare it to biological ChatGPT and analyze large-scale genetic data. Already the biggest AI model in biology, one day the EVO 2 was able to design new biological tools and treatments.

“We have a lot of observational data right now,” he says. “We know the correlation between genes and disease, but we don't know much about causality. Having something with the ability to predict causes and effects is really powerful.”

This type of prediction is a short-term vision for EVO 2. HSU shows an example of the breast cancer gene BRCA1. When women carry BRCA1 gene mutations, the lifetime risk of breast cancer increases dramatically. Over 60% of women with the BRCA1 gene mutation develop breast cancer at some point in their lifetime, compared to the overall 13% of women. Some BRCA1 mutations are known to be pathogenic, while others are known to be benign. However, most mutations are variations with unknown significance. We don't know what they're doing.

“If there is a pathogenic mutation, you will have a mastectomy. And if there is a benign mutation, you will get an annual mammogram. Ask HSU. “EVO 2 has an opinion on this and we found that the model is cutting edge in classifying the pathogenicity of BRCA1 mutations. We achieved over 90% accuracy in predicting potentially pathogenic mutations.”

Prediction of biological properties

EVO 2 is a product of a Bay Area independent nonprofit called the ARC Institute, and HSU co-founded with bioengineer and neuroscientist Silvana Konermann. The Institute aims to accelerate scientific advancement and deepen understanding of the underlying causes of disease, bringing together leading biomedical researchers from UC Berkeley, UCSF and Stanford.

The AI ​​model is based on its predecessor, EVO 1, launched in 2024 and fully trained in a single-cell organism. The Evo 2 will feature a few notches. This model was trained with a vast herd of biological information, containing from 100,000 to over 128,000 species of whole genomes and 9.3 trillion nucleotides from trees of life, including bacteria, plants and animals.

There are five base nucleotides that make up DNA and RNA: adenine (A), cytosine (C), guanine (G), thymine (T), and uracil (U). DNA contains A, C, G, and T, while RNA contains A, C, G, and U. Our genetic material is made from countless different sequences from these nucleotides, and EVO 2 uses this information to make stochastic predictions about what is most likely to come next within these sequences.

This model uses similar principles to those driving well-known major language models such as Openai's ChatGPT and Anthropic's Claude. And to build this cutting-edge model, researchers worked with industry-leading AI chip maker Nvidia.

“A machine learning model predicts the next token, which is the basic unit of data that the model processes,” says HSU. “ChatGpt predicts the next character and the next word. When you ask them to complete the sentence “Is it BE?”, they're a question. I have a question. I have a question. Additionally, AI models can capture complex biological properties based solely on sequence variation.

EVO 2 is a large linguistic model of a language that has never been spoken, and is expressed only in physical form. Its expression is the growth of cancerous tumors and the colour of the baby's eyes. EVO 2 can process up to 1 million nucleotides at once, allowing you to select patterns of data and identify relationships with other parts of the genome.

It does not only allow predictions about whether a genetic mutation is likely to be pathogenic. It also allows it to predict potential therapeutics that could potentially treat diseases and provide insight into the biological mechanisms that will progress. It could even help guide the direction that biomedical research takes.

“Researchers can generate data sets larger than ever before and perform bigger experiments, but it's not clear that this has led to more insights than ever,” says HSU. “Even the largest datasets are very small compared to the complexity of biology. That's where machine learning models come into play. You can employ large biological datasets, train your models, and find higher-order patterns of data that are more complex than you might imagine.”

“Efficiency is really important.”

In most cases, the science of biology was developed through a process of trial and error. Researchers formulate hypotheses, test them in scientific experiments, and analyze the results. The researcher then moves to the next hypothesis. And so on.

This approach takes time, but the results were obtained. Humans live longer than ever. Clinical trials for new treatments take years to implement and the vast majority of new treatments will not be available on the market. HSU compares this process to hiking in the mountains of California.

“Being a biomedical researcher can feel like walking in the wilderness,” says HSU. “You see a peak in the distance and you walk towards it. And after three hours of your walk, you realize you're not too close. And you need to decide if you're walking in the right direction.”

In biology, experiments tended to unfold on a lifetime scale for days, weeks, months, and years. And if you're heading in the wrong direction, you can get off course for quite some time.

“Efficiency is really important. You just have to work on the wrong things for years and get unlucky,” he says. “We really went to biology with something close to guessing and checking.”

One of the main objectives of EVO 2 researchers is to use AI to accelerate the development of discovery into actual treatments. The concept has its roots in the Covid-19 pandemic, with mRNA vaccines being deployed widely and quickly.

“That breakthrough has been around for 60 years,” says Howard Chang, senior vice president of global research for biotechnology firm Amgen and former ARC Institute researchers. “Messenger RNA was discovered in 1961 as a basic biological entity. It should take a long time.”

According to Chang, EVO 2 can already do things that help speed up the process. It allows you to accurately predict which RNA genes are essential for cell function and which RNA genes are essential. It shows the genes involved in controlling the behavior of cells that lead to disease. This allows researchers to follow the right path before.

“When you track individual families who are prone to a particular disease, there are many genetic differences that are mapped to locations on the genome where information changes can cause disease, but we don't know what they are. EVO2 can identify that,” says Chan.

“If Evo 2 tells you that diseases occur because the protein is too active, you know what the problem is and try to make a drug that addresses it. These are the kind of possibilities you have with Evo 2,” he adds. “It's a new kind of oracle.”

HSU argues that this type of advancement is particularly transformative in molecular biology. A study can take years to complete, and the vast majority of clinical trials fail.

“The failure rate for clinical trials is 90%. So in many cases we are just tackling the wrong drug target,” says HSU. “AI helps you find the right target more effectively.”

Towards a healthier future

For HSU, pursuing treatments for complex diseases is a profound and personal effort. When he was one year old, his grandfather was diagnosed with Alzheimer's disease. His grandfather lives with his family and HSU witnessed his inevitable decline. Slowly, he realized he wouldn't come back. Neurodegenerative conditions are incurable and ultimately fatal.

The experience was formative. As a teenager, HSU worked at the University Institute of Neuroscience at Stanford University. He studied Alzheimer's disease during his graduation research at Harvard University, and the disease remains the focus of his work at Berkeley and Ark Institute.

“If you look at the list of top killers in the US for 30 years, you can see that heart disease, cancer, Alzheimer's disease and more are the same as today,” says HSU. “This is a rather dire situation. It means that despite more and more biomedical research and more and more money is being spent, there is no progress in curing these diseases.”

AI is essential to improving things, HSU argues. The complexity of biology is simply too much for the human mind to fully tackle. And analyzing huge amounts of data is exactly what AI is great. HSU envisions a future in which AI can make biomolecular research more efficient and enable treatment tailored to patient health outcomes.

“We just don't want to understand the effects of certain genetic variations and whether they are a pathway to disease,” says HSU. “We want to use EVO 2 to conduct genome-wide association studies that sequence both healthy and unhealthy people to determine which genetic mutations are associated with the disease and communicate more specific about your own risks.