It is no longer an understatement that artificial intelligence has infiltrated everyday life. Now, it is being applied to one of the most challenging frontiers of biology: interpreting the regulatory genome. The human genome consists of over three billion base pairs, and only 1-2% of it encodes proteins. The remaining 98% is responsible for regulating gene expression and other vital biological functions. Google DeepMind has recently unveiled AlphaGenome , a deep learning model that is designed to help scientists identify how genetic variation disrupts the complicated networks controlling gene expression.

The human genome is frequently described as an “instruction manual for life,” and for good reason. It carries the molecular blueprint that constructs and maintains the human body. This instruction manual is written in a four-letter chemical code, organized into approximately 20,000 to 25,000 genes that encode proteins and control a vast array of biological functions. However, these genes only make up a minute fraction of the genome. The vast majority of the genome involves regulatory DNA that governs the way in which these genes are expressed. Understanding how this complicated regulatory landscape operates and how a single variation in the genome can lead to disease is a problem that decades of advances in computational technology have not quite addressed yet. It is exactly this challenge that AlphaGenome seeks to address.

AlphaGenome is the work of Google DeepMind , a United Kingdom-based artificial intelligence research laboratory that is renowned for creating the Nobel Prize-Winning AlphaFold system, a highly sophisticated computer tool for building molecular structures and designs of genes. Their newest creation, “AlphaGenome,” represents the latest extension of Google DeepMind’s ambitious goal to utilize artificial intelligence to solve fundamental scientific inquiries.

So, how does it work? Unlike previous computational models, which only accept short pieces of DNA as input, AlphaGenome is capable of parsing up to one million letters of DNA. By analyzing the properties of genetic material, it can make thousands of predictions about gene expression level, regulatory elements, transcription factor binding, and chromatin accessibility. Garnering impressive speed and accuracy, AlphaGenome was trained on publicly available human and mouse genome databases, such as ENCODE, 4D Nucleome and FANTOM5. By combining long-context sequence modeling with multimodal prediction, AlphaGenome matches, and in some cases, outperforms earlier regulatory models. In addition, beyond computational benchmarking, the model allows scientists to move easily from data sequencing to more complicated applications like synthetic biology and disease processes. For instance, the model can predict disruptions in RNA splicing related to the onset of cystic fibrosis. Further, it can assist in the design of regulatory motifs that selectively turn on or suppress gene expression.

Despite the experimental success the model has had, Google DeepMind made sure to acknowledge its shortcomings . For instance, the model struggles to identify regulatory sequences further than 100,000 base pairs away, and it is unable to fully capture the dynamic, context-dependent nature of cells. Additionally, AlphaGenome was not designed to comprehensively explain how variants influence complex diseases and can only predict their potential regulatory effects.

Beyond its technical accomplishments, AlphaGenome also invokes the broader question concerning the role of artificial intelligence in science. Can these machines help us interpret biological systems too complex for traditional computational analyses alone? There is no question about AI ever being able to replace experimentation, but developments like AlphaGenome demonstrate that machine learning can serve as a powerful ally in decoding the complexities of science.