Post by Charlie Lee, PhD

Charlie in the Loop #6: AlphaGenome on AWS You don't need a rack of A100s to run frontier genomics AI. Last weekend I self-hosted AlphaGenome, the sequence-to-function model, in my own AWS account. Ten output modalities, DNA windows up to half a megabase, on a single commodity GPU that costs a few dollars an hour and scales to zero when idle. Given nothing but raw DNA, it reconstructed the MYC gene's promoter: accessibility and transcription-initiation signals landing right on the transcription start site. It was never told where the gene is. When bigger windows failed, memory was the obvious suspect. I doubled it. Nothing moved. That's the fastest way to rule a theory out, and it sent me looking in the right place. I wrote up my testing on Amazon Web Services (AWS), including where the cheap GPU actually hits its limit, for anyone doing serious ML on a modest budget. Check out the article! Github repo on the CDK deployment - https://lnkd.in/gsAGpRuN Ankit Malhotra | Steven Malme | Hyunmin Kim | Daisuke Miyamoto | Sujaya Srinivasan | Nadeem Bulsara | Edwin Sandanaraj, PhD | Sikharin Kongpaiboon | ◻ Mike Lim | Noor K. | Vignesh Pillai | Nathaniel Ng | Aik Shin GOH | Kai Hui Ang | Brandon Ng Shan Yi | Khairul Azmi Mokhtar | MinSung Cho | Shig Okaya | Noon Ratchawan T. | Dr. Muhamad Yopan | Richard Goh | Keng-Hung (Leo) Lin | Max Lam #Genomics #MachineLearning #AWS #AlphaGenome #MLOps #OpenScience