Mouse models, cell lines and crop genomes all fail the same test: transfer

The same gap, four ways

Four preprints posted this week measure the same problem from four directions: a result obtained in one biological system does not automatically hold in another. A framework called GPAT scores how well mouse gene knockouts predict human traits. A pan-reference genome gives wheat researchers one coordinate system for 37 different assemblies. A pipeline called NanoPAMDA characterises new CRISPR enzymes mined from metagenomes. And a model called MODAE+ remaps cancer cell-line drug data so it matches patient tumours. Each is, in its own way, a tool for measuring transfer rather than assuming it. m:…7532 m:…2957 m:…5164 m:…5163

The mouse study is the bluntest. The GPAT framework quantifies how well mouse gene perturbations carry to human phenotypes, and the answer is uneven: 12 of 30 matched mouse-human phenotype pairs showed evidence of transportability. Genome-wide perturbation effects were positively correlated between the species, with a meta-analysis p-value of 2.3 x 10-27, but the correlation was uneven, and for fat mass and BMI, the mouse models showed limited evidence of carrying over to human phenotypes at all. m:…7532

References that carry

Wheat has a reference problem of a different kind: its genome is 18.87 billion bases, assembled differently by every group that works on it. WheatPanRef folds sequence data from 37 accessions into one linear pan-reference, with a bidirectional coordinate system that lets researchers project between the shared reference and individual assemblies without remapping. The preprint reports improved read alignment and the recovery of wild-relative introgressions and sequences missing from conventional references, the kind of variation a single reference simply cannot see. m:…2957

NanoPAMDA attacks the same class of problem for gene editing. Metagenomes hold vast numbers of uncharacterised Cas9 enzymes, but each one needs its PAM, the short DNA sequence it requires beside its target, measured before it can be used. The pipeline uses Nanopore long-read sequencing to profile PAMs kinetically and in parallel, and in the preprint it identified 14 new Cas9 candidates from 660,185 metagenome-assembled genomes and raw human microbiome reads. One of them, E10, was validated for PAM-specific activity in vitro and shown to work in E. coli. m:…5164

From cell lines to patients

The cancer study takes the transfer problem into the clinic. Drug sensitivity measured in cell lines often fails to match patient tumours, because the two sit in different distributions. MODAE+ improves cross-domain alignment between preclinical models and patient tumours while preserving cancer lineage and patient heterogeneity, and its predicted sensitivities were associated with overall survival in 46.3% of evaluated cancer-drug pairs in TCGA data. The drug-class-specific patterns in its predictions were driven by resistance-related genes involved in transport, metabolism and detoxification. m:…5163

All four are preprints, and all four stop short of the proof that would settle the question. What the week establishes is narrower but still useful: transfer can be scored, and the scores are uneven. The wheat study shows the same pattern in a different register: a single reference cannot carry variation it was never built to hold, so the fix is a reference that holds more of it. For drug discovery teams, the practical read is to treat mouse and cell-line evidence as a hypothesis with a measured confidence, not as a result, and to check, before trusting a model, whether anyone has measured how far it travels. m:…5163 m:…7532 m:…5164