Recently, artificial intelligence has gained a negative reputation, often with good reason. However, a group of researchers at Google’s DeepMind claims to have discovered an innovative application for AI: assisting humanity in understanding the “dark matter” of our genome more efficiently than ever before.

In a recent study published in Nature, DeepMind scientists introduced their advanced deep learning model, AlphaGenome. They assert that AlphaGenome can forecast the function of longer DNA sequences compared to existing models while maintaining a similar level of accuracy. The team is optimistic that their model can serve as a valuable tool for examining how subtle variations in human DNA impact our health and biology, particularly in the majority of the genome that operates silently in the background.

Pushmeet Kohli, vice president of research at Google DeepMind, expressed enthusiasm about AlphaGenome during a press briefing on Tuesday. Our DNA contains instructions for constructing and regulating every biological aspect of ourselves. Only a small fraction of our genes – around 2% – encodes proteins necessary for bodily functions like insulin or collagen. The remaining 98% consists of non-coding regions known as the dark matter of our genome. Previously considered junk DNA, this genetic dark matter is now understood to contain crucial sequences for regulating protein-making genes.

While most of the human genome has been mapped out by scientists, there remains limited knowledge about the functionality of many genes, especially those within non-coding regions. DeepMind researchers claim that AlphaGenome is the most comprehensive and accurate DNA sequence model available to date.

The researchers trained AlphaGenome on both human and mouse genomes. It can analyze up to 1 million DNA letters at once, surpassing older models which could analyze around 500 kilobases but with some trade-offs. Additionally, the model can predict various functional genomic tracks beyond gene expression, such as interactions between coding and non-coding regions or chromatin structure.

In comparisons with other AI models across 26 tests measuring genetic variant effects prediction, AlphaGenome matched or exceeded performance in 25 tests. Not only is it accurate, but it can predict nearly 6,000 human genetic signals associated with specific functions simultaneously.

While some external scientists have praised AlphaGenome’s capabilities, they acknowledge that it does not solve all mysteries surrounding our genetic code yet.

Ben Lehner from the University of Cambridge’s Wellcome Sanger Institute mentioned that AlphaGenome performs well based on extensive testing but still requires improvement given limitations in biological data quality and quantity.

Despite its imperfections, experts see AlphaGenome as a significant advancement in AI genomics with potential applications in diagnosing rare genetic diseases, identifying cancer-driving mutations, and discovering new drug targets.