Phenotypic screens find interesting compounds before anyone knows what they do. A small molecule can influence several targets or pathways at once. These layered effects allow drugging of complex biology, but follow-up becomes challenging without a ranked view of possible mechanism axes. Paper 05 compares small-molecule transcriptome fingerprints with public CRISPR states, including paired CRISPR edits, to generate calibrated mechanism hypotheses for validation.
Each compound produces an RNA fingerprint. Each gene perturbation in a public CRISPR atlas produces an RNA fingerprint too. Paper 05 puts those two dictionaries in the same search frame. The headline scan returned calibrated CRISPR-like hypotheses for 1,276 chemical tuples. Known controls behaved as they should, and two independent observations pointed to a shared V-ATPase/trafficking state hypothesis.
Phenotypic discovery is powerful because it can find compounds from function alone, even when the target is unknown. The cost of that strength is the next bottleneck: once a hit appears, teams can spend months figuring out what the compound is actually doing.
A CRISPR atlas is essentially a dictionary of cell states. Each entry describes what the cell looks like when a gene is knocked out or activated. If a compound produces a similar RNA state, that entry is a candidate mechanism axis. Z-Screen makes the search practical because the hypothesis lands attached to a specific combinatorial molecule and its building-block recipe.
We converted Z-Screen compound responses and public CRISPR perturbations into comparable ranked gene signatures, then asked which CRISPR state each compound most resembled.
The headline scan compared 8,599 ZEL024 / HEK293 chemistry tuples against 8,603 Replogle K562 CRISPR knockout signatures, with the top matches calibrated against matched random gene sets. Additional checks used Norman CRISPR-activation doubles and matched THP1 CRISPR controls.
The underlying questions were practical. Does full-molecule resolution actually matter. Do known controls recover known biology. Is the leaderboard dominated by a few generic hubs. Do unannotated compounds produce plausible hypotheses. Does the chemistry follow the logic of genetic combinations.
The ZEL024 / HEK293 scan linked 1,276 distinct chemical tuples to 1,714 distinct CRISPR programs at the calibrated threshold. The matches concentrated on interpretable biology: DNA replication, mitochondrial function, trafficking, and RNA surveillance.
MZ1 is designed to degrade BRD4. Against a matched THP1 CRISPR panel, BRD4 came back as the top calibrated match, with neighboring programs consistent with BRD4 biology.
BAY-293 is annotated as an SOS1 inhibitor. Its Z-Screen RNA state pointed to a coherent V-ATPase and vesicle-trafficking neighborhood. Surfacing this kind of secondary biology is one of the things a phenotypic platform should be doing.
The Norman atlas contains paired gene perturbations. Z-Screen tuple scores tracked the expected relationship between each double and its two singles, and the pattern repeated in a second library across three cell lines. The map is reading structured biology, not noise.
Some CRISPR perturbations match many states because they are broadly responsive. After removing the 50 most recurrent programs, the table still showed broad chemistry and CRISPR diversity. The headline is not being carried by a handful of generic hubs.
Documented targets were enriched in genome-wide rankings, but not universally recovered. The misses are informative: a drug can act through a target without producing a knockout-like RNA state in that particular cell context.
ZEL028-2 had shallower per-tuple coverage, so the analysis leaned on independent chemistry subsets as supporting evidence. Two non-hub candidates cleared the strictest corroborated tier (HEK293 ATP6V1A and A549 FKBP9), with a broader follow-up queue behind them.
This is the strongest internal validation signal. BAY-293 in A549 pointed to V-ATPase and trafficking biology. Separately, an unrelated ZEL028-2 tuple in HEK293 surfaced ATP6V1A with independent subset support. Different chemistry, different cell context, same biological neighborhood.
ZEL028-2 reproduced several themes from the headline library across three cell lines, including RNA-processing and mitochondrial-translation programs. Tuple coverage was shallower here, so we treat these as a ranked chemistry-resolved leaderboard, not finished mechanism calls.
A team starting from functional genomics can query a desired genetic state and ask for chemistry that pushes cells toward the same RNA program. A team starting from a phenotypic hit can ask the opposite question: which gene perturbations does this hit most resemble. The pilot does not replace target validation. It does turn the first day of mechanism work into a calibrated, chemistry-resolved hypothesis list instead of a blank page.
Public release. The chemistry-to-CRISPR match table and full reproduction pipeline ship with the data bundle on Zenodo. Reproduction time on a laptop is roughly 35 to 40 minutes for the headline ZEL024 / HEK293 scan.