Public edition · chemical structures withheld

Evidence behind the results

The source tables below make the headline results and nominated compounds inspectable. Download the rows, compare the populations, and follow the provenance manifest to the source analysis. Tables retain compound identifiers and measured results; structure images and SMILES/scaffold fields are withheld. Download hashes refer to the sanitized files; original source hashes are labeled separately.

Source release. The full pilot dataset is on Hugging Face (Zafrens/zscreen-pilot) with per-file SHA-256 manifests; an interactive Functional Drug Atlas covers a portion of it.

How to read the headline numbers

Scale

162,914 compounds and 190,699 compound–context profiles describe the pilot dataset analyzed in the white paper, not the platform’s capacity.

Learnability

On 7,544 HEK293 compounds whose building blocks were withheld from training, the additive recipe model reaches program r = 0.363 and 2.62× selection enrichment versus 0.337 and 2.53× for an end-to-end graph network trained on the same profiles. Retrospective replays recover the same responses with 33–52% fewer measurements.

Discovery

2,626 series–target matches occupy 178 clusters; 2,229 pass the lineage-support gate, which participates in selection. Chemical discovery gives 110 matches across 51 context–series labels; 33 have both internal-control and CRISPR support, including all ten highest-ranked chemical-twin series.

Calibration

CIGS 0.822 is a median matched-panel rank for 17 qualified controls. MZ1 AUROC 0.799 is chromatin-class recovery in H1650. CRISPR 0.580 is a compound median across 25 matched Tier-1 controls.

Controls that change how a claim should be read

Unseen building blocksOn the same 7,544 unseen-building-block HEK293 compounds, the additive recipe model reaches mean program r = 0.3628, R² = 0.1094 and 2.624× directional enrichment, versus Chemprop 2.3.1 at 0.3367, 0.0888 and 2.534×. Both use 5,721 training and 649 validation profiles. The shared-context three-seed transformer reaches r = 0.3978 and 2.861× enrichment using eight contexts, but its mean program R² is −0.9873: ranking improves while effect-size calibration worsens.

ActiveSeq population poolingPopulation recognition improves after pooling 32 held-out wells: joint image/RNA exceeds RNA alone by +2.86 percentage points for batch 1 → 2 and +10.48 for batch 2 → 1. Single-well classification shows no consistent gain; the pooled result is a post-fit aggregation under a fixed classifier, not a learned fusion model.

Source tables

Dataset	What it supports
platform_counts.json	Pilot-dataset scale
fold0_reference_model_summary.csv	Exact per-context simple and transformer results
cold_start_metrics.csv	Unseen-building-block head-to-head on 7,544 HEK293 compounds
replay_savings.csv	Retrospective measurement savings
activeseq_treatment_results.txt	ActiveSeq treatment-recognition and population-pooling results
cigs_qualified.csv	CIGS qualified-control denominator and panel metrics
cigs_calibration.csv	Full-atlas and salt-aware calibration metrics
chemical_leads.csv	The ten highest-ranked chemical-twin series and replication counts
chemical_qualified.csv	110 chemical matches reproduced by the source gate
crispr_calibration.csv	Compound-level and evaluation-row CRISPR calibration
genetic_leads.csv	The ten highest-ranked genetic-twin series and their selection fields
genetic_members.csv	Measured active/inactive analogs, wells and devices
provenance.json	Source paths and SHA-256 hashes for every released table

Next experiments

Prioritize fresh dose–response measurements and direct target assays for active/inactive compound pairs. Recheck single-well chemical series (ZS-SAR-01, -03, -06, -07) first.