Z-Screen makes new molecules and immediately asks living cells what each one did. There is no separate plate-prep, robotics line, or downstream sequencing workflow: the chip keeps the molecule, the image, and the RNA readout tied together in a single microwell. The five preprints show what that already produces at pilot scale.
A chip, not a robotics line. A Z-Screen chip holds tens of thousands of microwells. Each well contains a small group of cells and one bead carrying one compound. Once the chip is loaded, everything happens in place.
Two readouts from the same well. The chip images the cells, then reads their RNA. The image shows what changed. The RNA points to which biological programs moved. Paper 01 tests whether those two views work better together than apart.
Chemistry with memory. Each Z-Screen compound is assembled from known chemical building blocks. Every hit can be traced back to the pieces that produced the response, so the screen accumulates a record of which chemistry drives which biology. Papers 02 and 03 use that record to predict new combinations and to check whether the model is really learning.
Scale by adding chips. The chip is the throughput unit. Bigger campaigns mean more chips, not a custom robotics buildout for every program. The pilot used a thin slice of that capacity, which is why the same loop should yield more, not the same, when the campaign sizes grow.
Most adjacent platforms are strong on one axis: very large imaging screens, deep single-cell perturbation atlases, or compound-reference catalogs. The thing Z-Screen does that none of these do is keep the molecule's recipe, the image, and the RNA state attached to the same microwell.
These platforms see morphology at enormous scale. The follow-up question is usually what the morphology actually means biologically. Z-Screen keeps the image and adds a paired RNA readout from the same well.
These atlases are deep references for how cells respond to genetic perturbations. The thing Z-Screen adds is chemistry resolution: each measured response is tied to the building blocks that made the compound.
LINCS made compound-to-gene-program search useful for the whole field. Z-Screen extends that idea backward into design: because every compound has a known combinatorial recipe, a response can be traced to specific chemical choices.
DEL datasets ask whether a molecule binds a target. Z-Screen asks what the molecule does to a living cell. Paper 03 takes the DEL community's hard-won lesson about generalization seriously and separates real prediction from close-analog lookup.
Zafrens uses Z-Screen as a discovery engine across modalities, including small molecules, peptides, macrocycles, PROTACs, nucleic-acid delivery, cell therapies, ADC optimization, and patient-cell models. The common pattern is the same in all of them: make or present many candidates, read the cell response, and let each experiment inform the next one.
These preprints describe the measurement and modeling layer that sits underneath those active programs. They are also a fair test of whether Z-Screen can serve as a data-generation engine for outside partners.
A platform story should not live only inside a deck. These five preprints put early evidence in writing, backed by one public dataset and reproducible analysis code. The pilot is small on purpose, and each paper answers one underlying question:
Across the five papers the answer is yes, with boundaries. Some results are ready to scale. Others define the next decisive experiment. Publishing the pilot now puts the signal, the caveats, and the next step on the same page.
The papers are public and each stands on its own. Read them in order for the full platform story, or start with Paper 05 if you mainly care about connecting chemistry to genetic mechanism. The dataset and reproduction scripts are on Zenodo with a citable DOI.