Unlocking nature's code: innovative research draws parallels between AI models and genetic encoding

Posted on: 02 April 2025

This collaboration between Nick Cheney, Professor of Computer Science and Complex Systems Science at UVM, and Kevin Mitchell, Professor of Genetics and Neuroscience at Trinity, offers a fresh perspective on how the vast complexity of organisms is encoded in their DNA. 

By comparing current generative AI models that produce novel instances of images and text by distilling many prior examples into their most essential features and recombining those attributes in new ways, the paper argues that genomes similarly represent the most important features of an organism, distilled over countless generations of evolution into the attributes that most effectively represent it and stored in the form of DNA.   

The research considers one of the most enduring questions in biology: how the form and nature of organisms is encoded in their DNA?

“Why do cats have kittens, while dogs have puppies? This naïve question captures one of the core phenomena of life – that like begets like,” said Prof. Mitchell, Associate Professor in Developmental Neurobiology.

“How could the shape of an organism’s ears or the length of its legs or the tendency to knock objects off tables or bark at the mailman possibly be ‘encoded’ in a molecule like DNA?” 

Prof. Kevin Mitchell in Trinity.

Multiple metaphors have been employed to try to capture this relationship – from a blueprint to a program to a recipe – but these metaphors are often lacking and, sometimes, deeply misleading. In this new paper, Mitchell and Cheney – a geneticist and a computer scientist – draw an analogy with machine learning systems and set out a new way of thinking about how a species’ DNA encodes information about its physical form and nature. 

With the proliferation of powerful generative AI tools available online, even casual users are familiar with how these systems can generate an image – of a cat or a dog, for example. These systems learn, from training on hundreds of thousands of labeled images, the general, abstract characteristics that typify cats and dogs. More specifically, their neural networks encode a generative model – a highly compressed representation of these features, which can be decoded to produce a new image.  

In their paper, Mitchell and Cheney propose—by analogy—that the genome encodes a generative model of the organism – a compressed representation embodied in the chemical sequence of the DNA itself. In this framework, evolution acts as an “encoder” – an algorithm that learns and constantly tweaks the weights in the genetic network, shaping the variables of the generative model. The processes of development act as the “decoder” – decompressing the model through the stages of embryogenesis to produce a new individual of a given species with that species’ typical form and nature. 

By working at the intersection of AI and biology, Cheney suggests their research and resulting analogy provides a new perspective on the age-old question of how the vast complexity of organisms like plants, animals, and humans can be stored in just a single cell. “Our analogy of the genome as a generative model provides a platform to study the open questions in biology within a more tractable computational model that more easily lends itself to experimentation and analysis,” said Prof. Cheney. 

This theory suggests a mathematical framework to help make sense of the huge amounts of data that experimentalists can now gather on patterns of gene expression in developing embryos.  

The resulting models have profound implications for our understanding of how genetic variation influences traits and how species evolve over time. They offer the means of implementing these same principles in artificial systems, where artificial life and even artificial intelligence could evolve under the same dynamics as biological organisms. 

The published journal article can be read on the publisher's website. 

• Tags:
• Research
• Science