CENO evaluates genomic variants and designs regulatory sequences using contexts of up to one million nucleotides

CENO evaluates genomic variants and designs regulatory sequences using contexts of up to one million nucleotides

On 30 July, a bioRxiv preprint introduced CENO, a family of DNA models with up to one billion parameters. The authors trained the models on DNA segments containing up to one million nucleotides. They then applied the same modeling framework to variant effect prediction and the computational design of enhancers for mouse cortical cells.

Cells in the brain, liver, and skin contain nearly identical DNA. What differs is the activity of their regulatory regions, which determine which genes a cell activates and how strongly it activates them. An enhancer is a DNA region that helps activate a gene. Its activity depends both on nearby nucleotides and on the broader chromosomal context.

In the CENO preprint, the authors first trained the model on segments of 8 thousand nucleotides, then continued training with contexts of 131 thousand and one million nucleotides. A long DNA segment gives the model both the local sequence and the surrounding regulatory context of a gene.

The March Evo 2 model could already read and generate DNA using contexts of this size. CENO brings together two tasks in a single research system that are usually performed in sequence: evaluating a DNA substitution and finding a regulatory sequence for a specified cell type.

For each substitution, CENO compares the probabilities of the original and altered sequences within evolutionary alignments. The authors fine-tuned the model on alignments of related sequences from different species. The difference between the probabilities becomes an estimate of the substitution's effect. In the authors' tests, this estimate agreed more closely with known variant effects. The result is a ranked list of sequences for further testing.

For enhancer design, the authors built a separate computational workflow around CENO. A predictor estimates chromatin accessibility, meaning whether a DNA region can be reached by proteins that regulate gene activity. Conditional fine-tuning and reinforcement learning then modify the sequence for a selected mouse cortical cell type. This calculation produces candidates for laboratory testing.

CENO uses one modeling framework to connect the interpretation of a substitution with the design of the next sequence. The first task ranks variants by their effects, while the second proposes enhancers for a specific cell type. The authors use both components as successive steps in selecting candidates for experimental testing.