Meet Evo, the machine learning model that can decode DNA

Machine learning meets synthetic biology in Evo, a tool for precise genetic manipulation and biological research.

A new study by Eric Nguyen and colleagues introduces Evo, a machine learning model capable of predicting and designing DNA, RNA, and protein sequences at an unprecedented scale and accuracy. Evo, which uses 7 billion parameters, builds on advances in deep signal processing, positioning it as one of the most comprehensive tools yet in synthetic biology. Trained on 2.7 million diverse microbial genomes, Evo enables researchers to work across biological scales, from single molecules to whole genomes, and is poised to reshape approaches to DNA sequence design.

Christina Theodoris, a researcher at Stanford University’s Department of Genetics highlights the implications of this tool in her related perspective, noting that Evo’s ability “to predict the effects of mutations across all layers of regulation in the cell and to design DNA sequences to manipulate cell function would have tremendous diagnostic and therapeutic implications for disease.” This broad applicability marks a leap from existing generative models, which typically focus on individual molecules, to a tool capable of capturing the genomic interactions essential for understanding complex biological functions. Evo’s development represents a shift towards using machine learning not just to analyse biological data but to generate new sequences, potentially aiding the design of complex biological systems such as CRISPR-based therapies.

Harnessing the Power of a New Model Architecture

Built on the StripedHyena architecture, Evo differs from traditional DNA language models that often rely on dense Transformer architectures. While Transformers have found success in various fields, their application to biological sequences has faced limitations, including computational demands and a lack of fine-scale resolution. Evo overcomes these barriers by combining attention mechanisms with convolutional operators, allowing it to process and interpret sequences with greater efficiency and depth. This architecture allows Evo to achieve single-nucleotide precision across sequences up to 131 kilobases in length, a scale not previously feasible.

Testing of Evo revealed its capabilities across multiple biological contexts. It demonstrated high accuracy in zero-shot predictions, including mutation impacts on bacterial proteins and RNA, and it models gene regulation with precision. Evo’s training on large, diverse datasets also enables it to recognise complex genomic interactions, making it a versatile tool for researchers in the field. The potential applications are significant: from engineering proteins with specific functions to designing genetic constructs that may be employed in gene therapy. Future iterations could expand to eukaryotic genomes, incorporating even larger datasets to capture interactions on a broader biological scale.

Applications and Implications in Synthetic Biology

The Evo model’s scope includes functional biological systems, with researchers already exploring applications in synthetic biology, especially for designing CRISPR-Cas complexes and transposable elements. Evo’s unique feature of understanding coevolutionary links between coding and non-coding sequences offers synthetic biologists a new method to engineer DNA sequences that consider both functionality and stability within the broader genomic environment. In early tests, Evo successfully generated synthetic CRISPR-Cas complexes, demonstrating the model’s ability to support protein and RNA co-design—an advancement that may be instrumental for developing next-generation gene-editing technologies.

The integration of Evo into synthetic biology could also impact the speed of discoveries in genome engineering. The model’s capacity to generate megabase-length sequences with realistic genomic architecture suggests that it could streamline the design of complex constructs for therapeutic applications. This capability holds promise for researchers working on personalised medicine, as Evo could support the design of sequences tailored to modify specific cellular functions.

Evo’s developers foresee further advancements in both the model’s structure and its applications. Theodoris notes that “future models may learn from diverse human and other eukaryotic genomes, using larger context lengths to capture distant genomic interactions over larger genomic scales.” This vision hints at the long-term potential of Evo and similar models to deepen our understanding of genetic regulation across complex organisms and advance fields ranging from evolutionary biology to medical research.