AI Cracks DNA's Dark Matter: Two New Tools Decode Gene Regulation

The Challenge of Non-Coding DNA

More than two decades after the Human Genome Project mapped our genetic blueprint, scientists have struggled to understand the vast majority of our DNA. While protein-coding regions cover just two percent of the genome, the remaining ninety-eight percent has remained largely mysterious despite being crucial for orchestrating gene activity and harbouring many variants linked to diseases.

Two Breakthrough Models

Within days of each other, two artificial intelligence models have been published in Nature that address this challenge. Google DeepMind's AlphaGenome, released on January twenty-eighth, can analyse sequences of up to one million DNA base pairs and predict nearly six thousand genomic signals related to gene expression, splicing, and protein modification.

Meanwhile, a Dutch-developed model called PARM (Promoter Activity Regulatory Model), published on February third, takes a radically different approach. Requiring approximately one thousand times less computational power than comparable technology, PARM can be trained for specific cell types within a single day.

Cancer Research Applications

Both models address an urgent need in medical research: most cancer-related mutations occur in non-coding DNA regions, making interpretation extremely difficult. In one demonstration, researchers applied AlphaGenome to mutations identified in people with a type of leukaemia, and the model accurately predicted how non-coding mutations indirectly activated a nearby cancer-driving gene.

Global Adoption

AlphaGenome is now being used by nearly three thousand scientists across one hundred and sixty countries, processing around one million requests daily. Both models are freely available for academic research, promising to accelerate discoveries in cancer diagnostics, patient stratification, and therapeutic development.