How to build a virtual cell and biology scaling laws (letter.nikomc.com)

Markov Biosciences, a San Francisco-based startup, is poised to transform the biology sector through its development of "virtual cells," a concept likened to a potential GPT moment for the field. Founder Adam Green emphasizes the importance of large, unbiased datasets and innovative training objectives, arguing that traditional models reliant on hard-coded rules are becoming outdated. Instead of focusing solely on expensive perturbation data, Markov's approach leverages single-cell RNA-sequencing data, treating it as a ranking problem. Their findings suggest that pre-training virtual cells on observational data enables them to meet clean scaling laws, resulting in improved predictions for unseen perturbations over existing state-of-the-art models. This breakthrough holds significant implications for the AI/ML community in biology, challenging conventional wisdom around data handling and model training. By positioning virtual cells as specimens rather than simulators, Markov aims to derive deeper biological insights without relying heavily on biological priors. The company recently made a substantial prospective prediction related to antibody-drug conjugates in cancer treatment using their virtual cell model, which has garnered attention given the clinical stakes involved. This marks a potential paradigm shift in drug development and biological understanding, promising enhanced efficacy rates and deeper insights into the complexities of cellular behavior.