AlphaGenome useful but data gap limits it, says Bickmore

The genome is framed as something more dynamic than a static script. It notes that the Human Genome Project, completed between 1990 and 2003, showed that barely 2 percent of the human genome consists of actual genes. The harder question is how those genes are regulated, which involves much of the rest of the genome.

That regulatory picture includes transcription factors, enhancers, chromatin loops, epigenetic marks and post-transcriptional controls. These are described as context-dependent and combinatorial, which makes the genome less like a program and more like an open informational system. Genomic foundation models such as AlphaGenome are trained to learn correlations from sequence and treat this complexity as implicit in their training.

Bickmore's comments suggest the limits of that approach may lie in data rather than model design. She says she is sure AlphaGenome will be useful but with limitations, and she points to missing data on cell types and developmental change. A broader point is also raised: factors such as diet, environment and microbiome may matter, which could mean genomic models cannot predict all consequences of mutations from sequence alone.