Mining a genome-scale T-cell screen, together, across two continents
By Samuel Ahuno · 3 July 2026 · [draft — Samuel to review and make his own]
On Friday our team, together with the Kwarteng Lab in Ghana (Dr. Alexander Kwarteng, Humphrey P. K. Addy, Bernice Elorm Korsinah, Abdul Latif Koney Shardow, Joshua Yentumi, and many others), with Paz Polak on the call, spent the day mining a genome-scale CRISPR perturbation screen — a Perturb-seq — in human CD4+ T cells, using Anthropic’s Claude Science AI Workbench.
What we actually did
In plain terms: this kind of screen switches off one gene at a time across thousands of immune cells and measures the ripple effects on gene expression. Buried in that data are master-regulator genes — the ones that control broad T-cell programs, and, in some cases, can be targeted by drugs that already exist.
This is early, ongoing work. What it proves is not a biological result yet — it’s that a small team spanning continents can run genome-scale analysis together, in real time, with an AI collaborator handling the heavy lifting of the pipeline while the scientists focus on the biology.
The real bottleneck wasn’t the science — it was the hardware
Our biggest challenge was compute and storage. The full dataset was ~1.75–2 TB. Claude Science runs on your own machine by default, with the option to scale to the cloud or SSH into an HPC. Thanks to my PI Benjamin Greenbaum, I was able to run the session on a new MacBook on behalf of the Ghana team.
That detail is the whole point of AgenticGHX. Alongside investing in AI, we have to extend real hardware and compute to resource-constrained environments to make modern science genuinely collaborative. An AI collaborator narrows the expertise gap; it does not, on its own, close the infrastructure gap.
What’s next
This is exactly the kind of work our members can do — biology and engineering in the same room, across the diaspora and home. If it’s your kind of problem, come to a talk and say hello.
Ground rules we hold to: ship for real users, publish honestly, and never claim a result we haven’t earned. This post is an early field note, not a finding.