This one is different from the others. We are not reading a lab discovery. We are reading a tool: AlphaGenome, the model Google DeepMind released this year that takes raw DNA sequence and predicts thousands of things about what it does, at single-letter resolution. On paper it sounds like the genome is solved. It is not, and the reason why is the most useful part of the session. Here is the detail worth sitting with before you watch. The model hands you a score. This variant probably changes gene expression, by this much, with this confidence. That score is real, and it is often right. But a score is not the same thing as knowing what a change in the genome actually does. The model can tell you that something shifts. It cannot tell you why, or what the biological story behind it is. That gap between prediction and understanding is not a footnote. It is the whole reason biological fluency still matters even when the AI is very good. We cover what AlphaGenome gets right, where it quietly breaks, and what you can and cannot trust it for if you work with genomic data yourself. No background needed. Recording goes up here at 7pm Sydney, free to watch. Just head over the to Classroom and click on “Journal Club” or click on this link (https://www.skool.com/biology-unlocked-7569/classroom/becc80e4?md=13cb2db885b94e94b7360e549df8f738) to watch! Paper: Avsec et al. (2026). Advancing regulatory variant effect prediction with AlphaGenome. Nature 649, 1206–1218. DOI: 10.1038/s41586-025-10014-0

AlphaFold read proteins. AlphaGenome is trying to read the genome. Google DeepMind just published a model that takes one million base pairs of DNA as input and predicts thousands of functional outputs at single base-pair resolution. Gene expression. Splicing. Chromatin accessibility. Transcription factor binding. All from sequence alone. It matched or outperformed every existing model on 25 out of 26 benchmark tasks. It can simulate known mutations responsible for leukaemia and predict what goes wrong at the regulatory level. 98% of genetic variants that affect disease risk sit in non-coding DNA, the parts of the genome that don't make proteins but control how genes are switched on and off. AlphaGenome was built specifically to read that part. This is what we are breaking down tomorrow in Journal Club. Not just what AlphaGenome can do, but what it still cannot do, and what that tells us about how much of the regulatory genome we actually understand. Tomorrow. Saturday. 7pm Sydney, 2:30pm IST. Biology Unlocked Journal Club. Free to join. Link in the first comment. #molecularbiology #genomics #bioinformatics #AlphaGenome

A drug showing 15-fold downregulation of an oncogene sounds like a breakthrough. It wasn't. And only a biologist caught it. Jan was testing whether her drug could reduce expression of two oncogenes in a cancer cell line. Her positive controls were showing roughly two-fold downregulation. Her drug was showing fifteen-fold. She sent the data to Jon, the bioinformatician in the collaborating lab. The numbers held up. They agreed: this drug had potential. Except something was bothering Jan. She wasn't seeing the protein-level changes you would expect from a 15-fold shift. And the cells weren't dying the way they should be. So she went back to the raw Ct values. She found it. One of her three reference genes had inconsistent expression across samples. Reference genes are used in qPCR to normalise data across samples. The assumption is that they are stably expressed regardless of condition. But reference gene expression varies across tissues and cell types. A gene that works in one tissue may be completely unreliable in another. When one reference gene is unstable, your normalisation is wrong. When your normalisation is wrong, every fold-change built on it is wrong. The 15-fold downregulation was not a drug effect. It was a maths problem built on a biological assumption nobody had checked. Jon never saw it. The analysis was clean because the input looked clean. An AI would have called this a finding. Jan caught it because she knew what raw Ct values meant. Because she knew that reference genes are not always reference genes. If she hadn't, this would have been submitted. Possibly published. Resources spent chasing a result that was never real. The difference between a finding and an artefact is not always in the statistics. Sometimes it is in knowing which biological assumptions your numbers are resting on. Would your pipeline have caught this? #bioinformatics #datascience #molecularbiology #qPCR

You are handed a scenario, not a real dataset. I am sure this type of brief has landed in your inboxes: "Here's RNA-seq from 40 tumour samples and 40 matched normal tissue. Samples were collected over 8 months across two clinical sites. Build a classifier." Before any modelling: what's the first question you'd ask the person who handed you this? Reply with your one question. I'll pull the most common threads in a follow-up post and run through what the methods section would need to say for each one to be answerable, the way I'd flag it in an editorial review. (No wrong answers here. The ones that name a specific confound, site, batch, collection date, will probably crowd the leaderboard.)

Whether you joined live, dropped a question in the chat, or are catching the recording now, thank you for spending your Saturday evening reading a real paper with us. Tonight we covered a lot. RAS, one of the most mutated proteins in human cancer, was considered undruggable for forty years. A new class of molecular glue finally cracked it by not trying to find a pocket in RAS at all, but by recruiting a protein the cell already makes and landing it on the surface RAS uses to signal. The drug works. The ASCO data is striking. And then the tumour finds two completely different ways to restore the exact interaction the drug broke. Different mutations, same escape. That last part is the idea worth carrying: convergent resistance. Cancer does not need to be creative. It just needs to find any road back to the same answer. If you have a question that did not make it into the chat, leave it below. I will work through everything posted here during the week. The full paper is linked below. It is open access. Sang et al. (2026). Disrupted molecular glue complex drives RAS inhibitor resistance. Cell 189, 2918-2933.DOI: 10.1016/j.cell.2026.03.031 Journal Club No. 05 is already locked. We are reading the AlphaGenome paper from Google DeepMind, a model that takes a raw DNA sequence and predicts thousands of functional genomic outputs at single-base resolution. If you work with genomic data, build models on biological data, or just want to understand what AI can and cannot actually do with a genome, this one is for you. See you next Saturday.