Published on 9 min read
Journal Safari #7. DeepMind's AlphaGenome reads a megabase of DNA, predicts thousands of genomic tracks, and matches or beats the strongest competing model on 25 of 26 variant tests, mostly by a little. It calls the direction of a variant's effect far better than the size, which makes it a scanner rather than a verdict, and a decent preview of where AI in biology is right now.
Read more →
Published on 6 min read
Journal Safari #6. A Nature Genetics paper ran exome-style burden tests on 708,956 whole genomes for BMI and type 2 diabetes. Whole-genome sequencing turned out to be a better exome, worth about a fifth more signal, and the most interesting biology in it, a kidney effect of IRS2, rests on a few dozen carriers.
Read more →
Published on 14 min read
Journal Safari #5. Malva makes 74 million single cells searchable by raw sequence, answering 24-mer queries in milliseconds, and it found oral bacteria outside a tumour section that a reference-based pipeline discards by construction. The engineering is excellent. What decides whether it still exists in 2036 isn't in the paper, and three of its four referees spent ten months saying so.
Read more →
Published on 18 min read
Journal Safari #4. Eleven ESM protein language models were supposed to be near-interchangeable. They have different blind spots that do not overlap, and training each one on the others' best guesses puts sequence-only at the top of the ClinVar board. The gain comes off about 200 proteins, which is the part I keep chewing on.
Read more →
Published on 12 min read
Microbial ecology spent twenty years learning that in low-biomass samples, the thing you measure is partly your reagents. Single-cell has the same exposure, most of the same tooling gaps, and none of the same scar tissue — and the field is moving into exactly the regime where it starts to matter.
Read more →
Published on 17 min read
Journal Safari #3. Pan-human Azimuth puts 27 million cells from 23 tissues on one annotation tree, then classifies them with a model 70× smaller than the foundation model I read in July — and wins. The contribution is not the model. It is the labels.
Read more →
Published on 11 min read
Journal Safari #2. The UK Biobank just finished whole-genome sequencing all 490,640 participants. I read the resource paper to answer one question: what does WGS actually buy you over the array and exomes we already had — and who pays the storage bill afterward?
Read more →
Published on 8 min read
Kicking off Journal Safari, my biweekly close-read of one recent paper. First up: CellVQ, a 500-million-parameter single-cell foundation model that squashes every cell into a discrete code — and the peer reviews that quietly agreed with my doubts.
Read more →