DeepMind Launches AlphaGenome: AI for Decoding Life to Become a Key Tool

On June 26th, it was reported that DeepMind, an artificial intelligence research institution under Alphabet, is bringing genomics into a new era. By launching a brand - new AI model, AlphaGenome, DeepMind aims to answer a question that has puzzled biologists for decades: which part of human DNA plays a key role in disease formation and how does it work?

Five years ago, Google launched AlphaFold, an AI model for predicting the three - dimensional structure of proteins. Due to its revolutionary contribution to biology, this technology won the Nobel Prize last year. It also gave rise to Isomorphic Labs, a spin - off company focused on drug development, and spurred a wave of AI - driven pharmaceutical startups. Now, AlphaGenome attempts to answer another more fundamental yet equally important question: when a single letter in DNA changes, what impact will it have on gene expression? Is this impact related to health or disease?

Different from traditional models that have focused on short sequences or single tasks in the past, AlphaGenome can handle DNA fragments up to one million base pairs long and predict multiple biological properties related to gene regulation in real - time, including gene start positions, splicing patterns, RNA expression levels, and even the likelihood of protein binding.

This model doesn't just focus on the known protein - coding regions, which only account for 2% of the genome. For the first time, it comprehensively explores the "dark matter" of the genome - the vast but long - neglected non - coding regulatory regions. These regions are considered crucial for regulating when and where genes are turned on or off, and it is often in these locations that mutations closely related to cancer, rare diseases, and even neurological disorders are hidden. In the future, diseases such as cancer or Alzheimer's can be detected earlier, better understood, and treated more personalized.

A Model for Comprehensive Prediction

DeepMind states that AlphaGenome is currently the first AI system capable of integrating long - context and single - base resolution prediction capabilities in a single architecture. By introducing an architecture that combines convolutional networks and Transformers, the model achieves unprecedented accuracy and breadth - it not only makes accurate predictions but also predicts a more comprehensive range of content.

In practical applications, researchers can submit a DNA sequence to the model and quickly obtain an assessment of the regulatory activity of this sequence in different tissues and cells. This speed and efficiency are of direct significance for promoting research in fields such as rare diseases and cancer.

In a case study, AlphaGenome successfully predicted that a non - coding mutation in the genome of a leukemia patient might lead to the abnormal activation of the oncogene TAL1 by introducing a new MYB binding site. This prediction is highly consistent with the known pathogenic mechanism, demonstrating the potential of AlphaGenome in revealing the causal relationship between mutations and diseases.

Leap in Efficiency and Performance

According to DeepMind, AlphaGenome outperformed the existing best models in 22 out of 24 standard tests in the field of genome prediction. In the task of predicting mutation effects, it either matched or outperformed specialized models in 24 out of 26 cases.

Notably, AlphaGenome is the only model that can achieve joint prediction across tasks and modalities. Previously, researchers often had to use multiple models to complete these tasks. Now, a single API call can provide a full set of prediction results, greatly improving research efficiency.

More importantly, without sacrificing performance, the training cost of AlphaGenome has been significantly reduced - the training time is only 4 hours, and the computing resources it occupies are half of those of its predecessor, the Enformer model.

Towards Personalized Medicine

Although the current version is only for non - commercial scientific research and has not been used for individual genetic diagnosis, its potential significance is self - evident. The prediction ability of AlphaGenome will enable scientists to identify key mutations more quickly and enhance the early screening and targeted treatment capabilities for complex diseases. "This work lays the foundation for precision medicine," said Marc Mansour, a professor of cancer genomics at University College London. "We finally have a tool that can assess the impact of non - coding mutations on a large scale, which is the key to cracking the mechanisms of complex diseases."

DeepMind also admits that AlphaGenome is not all - powerful. Currently, the model still has difficulty capturing long - distance regulatory signals more than 100,000 base pairs away from the target gene. In addition, the ability to capture differences between different cell and tissue types is still being optimized. More importantly, it cannot replace medical diagnosis - complex traits and diseases often involve developmental, physiological, and environmental factors, which are not yet within the scope of AlphaGenome's modeling.

However, for the scientific research community, AlphaGenome provides a unified, powerful, and scalable tool framework. With the addition of more data, it is expected to be extended to other species and even support clinical applications in the future.

This article is from "Tencent Technology", author: Wuji. It is published by 36Kr with authorization.