Google's DeepMind company claims its AI model can decode the human 'dark genome'

According to Xinhua News Agency, DeepMind, a Google subsidiary, published a cover story in the journal Nature on the 29th introducing the AlphaGenome deep learning model. This model can decode 98% of the “dark genome” in the human genome that is crucial for health, and in the future, it can be used to gain deeper insights into genetic diseases, improve genetic testing, and provide information for the development of new therapies.

The human genome consists of approximately 3 billion DNA base pairs. The double helix structure of DNA is formed by the ordered arrangement of four bases: A, T, C, and G. Among these, only about 2% of the genome encodes proteins, while approximately 98% is non-coding regions. These regions, often referred to as the “dark genome,” do not directly encode proteins but can influence gene expression. A large number of disease-related variant sites are located within these poorly understood non-coding regions.

Traditional methods often require a trade-off between sequence length and prediction accuracy. The AlphaGenome model breaks through this technical bottleneck, enabling high-resolution predictions of long DNA sequences. The model is trained using the genomes of humans and mice and learns how DNA sequences influence various biological processes.

Research shows that the AlphaGenome model can predict the function of DNA sequences up to 1 million base pairs long. This tool can not only predict the location of genes but also forecast the impact of the “dark genome” on gene expression and splicing. Notably, the model can predict the effects of single-letter (single-base) changes in the genetic code.

The research team conducted comprehensive evaluations of the AlphaGenome model through 26 benchmark tests, demonstrating that the model achieved or surpassed the performance of the current state-of-the-art models in 25 tasks.

On the same day, the DeepMind team stated on social media that this tool can help researchers understand DNA, predict the molecular effects of gene variations, and promote new biological discoveries.

The AlphaGenome model was made available for non-commercial use last year, and since then, over 3,000 scientists have used the tool. Although the research team developing the model admits it is not perfect, some scientists have described it as “an incredible achievement” and “an important milestone.”