VCC 2026 Chronicle · 2026-09-18
The edge of the map
Borrowed biology kept winning — until we asked which cells we were actually predicting.
immutable copy · lens://079c81cf04360d44e0de39eb9f74dc387746a6d3e78e1b20584fc3af65911074
In plain words
Our best score (+0.0596, rank 486 of 984) came from borrowing real perturbation signatures — but a lineage check showed the challenge cells are not the cells those signatures came from: context A looks Jurkat-like, B weakly RPE1, C unresolved. Three honest attempts to fix the mismatch all failed: moving K562 signatures keeps only 13% of their direction in a new cell type, lineage-matched scaling scored worse than the uniform guess, and no public screen of the right cells covers any of the 300 targets. The lookup has a ceiling — so we stopped borrowing and started training our own model on a rented GPU.New words
- CRISPRi — Turning a gene's volume down with a guide — a dimmer, not a switch.
- Perturb-seq — Silence one gene per cell across millions of cells, and read what changed — a catalog of cause and effect.
Where this comes from
k011 delta-scale champion — +0.0596, rank 486 k011-delta-scale-x1p7-validation
Kytos Observatory
k012 lineage score — context identity k012-lineage-score
Kytos Observatory
k013 context-scale — the failed fix, published k013-context-scale-validation
Kytos Observatory
Deep science
Track 1 hit its ceiling across three axes. (1) Paired transfer (k012): leave-one-out on 47 K562↔hESC target pairs gives mean cosine 0.126 — borrowed deltas keep ~13% direction cross-lineage. (2) Per-context amplitude (k013): lineage-measured scale factors {A:1.7, B:2.65, C:0.75} scored +0.0312 vs the uniform ×1.7 champion at +0.0596. (3) Coverage: figshare GWPS is K562-only — the Jurkat/RPE1/HepG2 public files are a 2,393-gene essential screen with 0/300 panel overlap. Lineage scoring on top-2000 discriminative genes: A→Jurkat r=0.649, B→RPE1 0.369, C unresolved 0.379. Track 2 (Nebius GPU, trained conditional model) is now the primary path.