Virtual Cell on Mike's Bloghttps://mikeogilvy.github.io/tags/virtual-cell/Recent content in Virtual Cell on Mike's BlogHugozh-cnFri, 05 Jun 2026 07:19:29 +0800GEARS Explained: How Graph Neural Networks Predict the Unseenhttps://mikeogilvy.github.io/posts/gears/gears_explained/Fri, 05 Jun 2026 07:19:29 +0800https://mikeogilvy.github.io/posts/gears/gears_explained/<p><em>A deep dive into the model that can forecast what happens when you knock out genes nobody has ever touched.</em></p> <hr> <h2 id="the-problem">The Problem</h2> <p>You&rsquo;re a biologist. You have a Perturb-seq dataset — 100,000 cells, each with one or two genes CRISPR-knocked out, plus single-cell RNA-seq readouts of the full transcriptome. You&rsquo;ve experimentally perturbed 100 single genes and 130 two-gene combos. But there are 4,950 untested pairwise combinations. Running them all would cost a fortune.</p>