Zafrens/zscreen-rsi

Z-Screen RSI

Learning chemical recipes for useful biology

Drug discovery begins with a desired change in biology. The challenge is to find a chemical intervention that produces it reliably, in the right cells, with an acceptable exposure and safety profile. Z-Screen RSI explores an experimental route from that biological intention to chemical design: measure rich cellular responses across combinatorial chemistry, learn which chemical components shape those responses, and let each round of measurements improve the next choices.

The opportunity is a growing map between chemical recipes and cellular outcomes. Genetics supplies experimentally grounded response objectives; imaging and RNA describe what chemistry does; families of related molecules reveal reusable chemical patterns. Feedback then helps decide what to test next. A useful result is both a candidate intervention and a better experiment for distinguishing how it works.

This package supplies runnable components of that approach: chemical-family atlases, compact genetic references, known-control calibration, retrospective chemical-selection experiments, chemically unconditioned response clustering, matched recipe comparisons, and paired imaging/RNA measurement studies. All advertised examples use bundled inputs and run locally. The results connect each component to the next scientific decision.

Three findings that motivate expansion

1. Biology can provide a search objective. Known controls recover aspects of established pharmacology, with substantial dependence on cell context and baseline. Family-level responses retain genetic-ranking signal across member holdouts. Biological calibration
2. Chemical composition contains learnable information. Updating chemical-selection models beats fixed ranking and random acquisition for five of six objectives in the extension benchmark, across all three starts for those objectives. BRD4 is a boundary case. Adaptive discovery
3. The learned signal can be inspected and challenged. Selected compounds concentrate in interpretable building-block families. Matched substitutions, held-out measurement batches and imaging/RNA pairing test different explanations for those patterns. Chemical organization · Measurement design

Toward a therapeutic quine

Here, a therapeutic quine means a description of useful biology coupled to an executable chemical recipe—compound, dose, exposure and cell context—that can reproduce the intended response in fresh cells. The measurement of that response becomes another training example and helps refine both the recipe and the way the system chooses its next experiment.

The longer-term ambition is a transferable foundation model of chemical–biological response. Today's package contains fixed representations, compact reference scoring and recipe-based predictors. The next demonstrations require prospective chemical choices, fresh measurements, functional validation and transfer to held-out chemistry and cell contexts. The vision and experiment plan lays out those steps.

Companion data resource

This package is the analysis-ready release built on the Z-Screen pilot program-usage layer. The full pilot archive — complete 6,000-gene expression surfaces, imaging features, detection-level measurements and reference model weights — is published separately: Z-Screen Pilot on Hugging Face. Context and batch identifiers are anonymized in this release relative to the pilot (for example, the pilot's zel039_aec7 context appears here as aec7); all numerical results are unchanged.

Open the package

Reader	Start here
Understand the scientific case	Results and interpretation
Run a working example	Installation and four notebooks
Examine the statistical definitions	Methods
Use the data	Data guide
Understand the computation in plain language	How the discovery loop learns
Design the next experiment	Vision and next experiments

Zafrens, Inc. · Contact · Scientific references · Component licenses and attribution