Drug teams routinely re-run the same chemistry across many cell types to find the biology that matters. This paper is a first attempt to cut down that work: learn how chemical response programs move between paired cell types, then use that map to prioritize the next target-cell experiment.
Just reusing a compound's source-cell signature in another cell type is not good enough. A learned transfer map does better. The unit that travels best is the chemical program, not the isolated molecule, which lines up with how early discovery teams already make decisions: at the level of a chemical series.
A lot of preclinical cost comes from running the same experiment again in every cell type that might matter. Skipping the wrong one can miss efficacy, toxicity, or selectivity. Running all of them is expensive.
This paper starts to bend that curve. With paired examples across two cell types, Z-Screen can learn a map from one response space to the other, and a new source-cell measurement can help prioritize target-cell follow-up. We are careful about the scope: this is the cell-intrinsic RNA response, not whole-body pharmacology or toxicology.
Even within that scope, the implication for ops is real. A small paired landmark atlas can make every future source-cell measurement more useful for whatever the next cell type of interest is.
We used Z-Screen libraries with enough shared chemistry across paired cell lines to make the comparison fair. The clean molecule-level comparisons were ZEL024 between HEK293 and H1650, and ZEL031 between A549 and THP1.
We tested two levels. Can a learned map project individual molecule responses better than simply reusing the source-cell response. And do grouped chemical programs transfer more cleanly when the groups still preserve chemical meaning.
Aggregation can be misleading. Broad averages can look strong while losing the chemistry the groups were supposed to represent. The headline grouped results therefore use chemistry-resolved groups, not single-building-block averages that pool too much.
When molecules were grouped along meaningful chemical axes, learned transfer was much stronger than direct reuse. Early discovery decisions are usually made at the chemical-series level, so this is the more useful unit anyway.
A compound's source-cell response cannot just be carried over into another cell type. Learned maps recovered more target-cell information than identity reuse, so cross-cell projection is a real prediction problem and not a relabeling exercise.
Accuracy climbed steadily as more shared compounds were available for training. The scaling path is straightforward: broaden paired landmark coverage and more cell-type projections become useful.
Some library-to-library projections came back positive, others were weak or asymmetric. The current data tell us this is worth designing prospectively, not that it is already solved.
An exploratory gene-level decoder partially recovered target-cell gene-rank structure. We treat this as a secondary readout for now. It points toward producing ranked gene and pathway hypotheses from transferred responses as the paired-landmark atlas grows.
Public atlases such as LINCS L1000 and Tahoe-100M are valuable references for comparing compounds across cell lines. Z-Screen brings combinatorial chemistry provenance to the same comparison: every response is tied back to specific building blocks, so the platform can transfer chemistry-linked programs alongside individual compound signatures. Paper 05 picks those programs up and matches them against CRISPR perturbation states.
Public release. The transfer maps, paired-landmark splits, and grouped-program benchmarks ship with the data bundle on Zenodo.