Screen UMAP explorer

A 2-D projection of every AssayBench screen, computed from the model-agnostic ground-truth transfer matrix. Pan, zoom, search, and recolor by split, phenotype, library type, cell type, or publication year.

How the transfer matrix is built: for each ordered pair of screens (source → target), we take the top-100 genes of the source's ground-truth ranking and score them against the target's ground-truth ranking using AnDCG@100. The full N×N matrix is then symmetrized and converted to a distance via 1 − ((T + Tᵀ)/2). UMAP is fit on that precomputed distance with n_neighbors=15, min_dist=0.1. The Train-Fitted layout fits UMAP on training screens only and projects val / test / LaTest via a 10-NN soft assignment so layouts are comparable across the year split.

Why look at this?

The phenotype heatmap shows that some phenotype classes are far easier than others. The UMAP shows the mirror image of that observation: viability screens form a dense, well-mixed cloud (you can predict any of them well from the others), while reporter-activity and host-pathogen screens form sparser clusters that don't transfer as well. The Train-Fitted projection is useful for asking is my held-out screen in a region where the training set is dense? A useful proxy for OOD risk.

Caveat

The transfer matrix is built from ground-truth rankings, so closeness here reflects overlap in hit-gene structure, not necessarily semantic similarity of the screen descriptions. Two mechanistically related screens with disjoint hit sets can still appear far apart. Treat the UMAP as a complement to the phenotype heatmap, not a substitute.