They were learned during training along with other parameters. See page 31 of the manuscript detailing where they were incorporated.

Thanks to the large consortia like ENCODE, FANTOM5, GTEx, 4DNucleome for generating key training data. And finally, thanks to numerous researchers developing inspiring DNA sequence-to-function models, informative evaluations, and interpretation tools.

Big thanks to the amazing team! They went above and beyond to get us here. It's a big privilege to work with them. Also thanks to our early testers for their feedback.

This model represents a start, and not the end, of our AlphaGenome named research and release journey aiming to help researchers better understand the genome using AI. We’d love to learn about how you’re using it. You can share projects & feedback in alphagenomecommunity.com forum.

A single model can be more convenient to use and deploy - saying this as someone who worked on a model zoo before. SOTA performance across diverse tasks suggests that fine-tuning it on new data should work quite well. One could also extend it with new assays in the future.

This all-in-one model was achieved by lots of amazing engineering that unlocked training on base-resolution data with long sequence context without increasing the total training compute - training a single model actually used half of what we used for Enformer on the same (older generation) hardware.

I see this model as a fusion of many different DNA sequence-to-function models tackling different steps of gene expression. It also has important new features like splice junction prediction to better capture the complexity of different splicing events.