Introducing AlphaGenome: Using AI to Understand the Genetic Code

Following Deerfield Academy’s fifth Deerfield Forum on April 21, hosted in partnership with Munk Debates on the motion “Be it Resolved, let’s engineer better human beings,” my classmates and I left the event not only having heard a deep dive into bioethics but also grappling with new, unfamiliar tools like AlphaGenome.

As speakers from the pro and con sides exchanged their arguments on genetic engineering, one idea in particular remained unelaborated throughout the debate: AlphaGenome. This term was included as part of the pro-engineering argument, used to suggest a future where humans could optimize their genetic potential, but neither side addressed what kind of technology it was.

What exactly is AlphaGenome?

AlphaGenome is a deep learning model developed by Google DeepMind, described as a powerful “swiss army knife” for predicting how non-coding regions of DNA influence molecular processes. Google DeepMind previously released deep learning tools like AlphaFold2 and AlphaFold3 that allowed for computational predictions of protein structure and protein-ligand interactions. Building upon this previous work, with AlphaGenome, the team has expanded to predicting patterns at the genomic level.

AlphaGenome uses artificial intelligence to recognize patterns within DNA that would be nearly impossible for humans to detect on their own. Its ability to analyze genetic information at the level of individual base pairs and still account for long-range interactions across the genome is what differentiates the technology from earlier models that solely focus on short DNA sequences. As over 98% of the human genome consists of non-coding regions which play critical and diverse roles in regulating gene expression and altering cellular function (e.g., by influencing transcription factor binding, altering chromatin accessibility, or affecting RNA splicing). Essentially, instead of studying one gene, mutation, or single function at a time, AlphaGenome can simultaneously predict how changes in DNA affect gene expression, protein binding, and even the three-dimentional structure of the genome.

Despite its promises, AlphaGenome is not yet publicly accessible like a normal Google tool. It is currently a research AI model and only accessible through a limited API (application interface), mainly for non-commercial scientific research. Moreover, there are broader limitations preclude this technology from actually being able to implement changes such as those discussed in the Deerfield Forum. Predicting changes in gene expression or chromatin structure is not the same as predicting human traits like intelligence or personality, which are still influenced by thousands of genetic variants interacting in complex ways that are not yet fully understood. In this sense, the idea of constructing an “optimal” genome masks the undiscovered intricacies of gene expression.

The versatile, “swiss army knife” nature of AlphaGenome represents a step toward the pro-engineering side’s future vision of the world. If we are developing tools that can analyze and even potentially predict the effects of nearly any genetic change, then we open up the possibility of intentionally shaping human traits. At the same time, we must acknowledge that engineering while relying on a “still-evolving” model risks oversimplifying the complexity of human traits, meaning the pursuit of an “ideal” genome may come at a cost we do not yet fully understand.