ZafrensInvestor & collaborator portalZ-Screen discovery platform
Zafrens · Z-Screen discovery platform · Pilot data package V2

Teaching biology to read chemistry, one reusable part at a time

Z-Screen pairs combinatorial small molecules built from reusable building blocks with whole-transcriptome RNA and high-content imaging readouts in microfluidic nanowells. This portal presents the Z-Screen pilot dataset (162,914 compounds, 190,699 profiles, 11,435 same-well ActiveSeq wells) and what it shows: the measurements are learnable and they match chemical and genetic reference responses.

Read the technical white paper Download PDF Pilot dataset on Hugging Face ↗ Interactive drug atlas ↗ Companion RSI white paperEvidence register →
162,914
pilot compounds
190,699 compound–context RNA profiles across 8 contexts; 627 building blocks
r = 0.89–0.94
block-level modularity
in silico building-block knockout reproduces measured block effects; fragment sums reach 0.97–0.99
0.822
chemical-twin calibration
median matched-panel rank on 17 CIGS reference controls; MZ1 chromatin-class AUROC 0.799
2,626
genetic-twin matches
strict series–knockout matches in 178 clusters; 2,229 pass the lineage-support gate

Pilot, not ceiling. The counts on this page describe the dataset analyzed in the white paper. Z-Screen itself is designed to scale to millions of compounds and beyond.

From reusable chemistry to mechanism hypothesesNumbers describe the pilot dataset analyzed in this paper; the platform itself is designed to scale to millions of compounds and beyond.1 · CHEMISTRYCombinatorial library627 reusable building blocks5 positional slots (bb0–bb4)162,914 pilot compounds722,397-recipe design space2 · MEASUREMENTZ-Screen nanowell profiling2–10 cells per nanowellUp to 50,000 wells / run6,000-gene surface → 32programsImaging on 33,575 compounds;ActiveSeq same-well co-assayon 11,435 wells3 · LEARNINGModels & benchmarksBlock knockout r = 0.89–0.94Recipe, transformer & GNNcomparisonsGraph features vsfingerprints: 40/40 splitsActive-learning replays:33–52% fewer measurements4 · HYPOTHESESPhenomimicry & hypothesesChemical twins (LINCS,Tahoe-100M, CIGS): 110 matches/ 51 seriesGenetic twins (43 Perturb-seqsets, ~18,800 genes): 2,626matches / 178 clusters14 mechanism hypotheses from20 seriesMeasure cellular responses to reusable chemical parts → learn transferable rules → test the strongest mechanism hypotheses.
The Z-Screen discovery architecture: measure, learn, prioritize. Click to enlarge.
Pillar 1 · Platform and combinatorial grammar

A reusable chemical language for learning cellular responses

A conventional screen tests a pile of unrelated molecules for a single yes/no signal. Z-Screen builds molecules like sentences from recurring chemical words, then reads out 6,000 genes per nanowell, with imaging paired on 33,575 compounds and measured inside the same well on 11,435 ActiveSeq wells.

Investor lens

The “Lego grammar” of molecules

In the pilot dataset, 162,914 molecules are assembled from 627 reusable building blocks in up to five positions. Each block appears in hundreds of molecules (median 461), and the dataset contains 5.08 million pairs of molecules that differ by a single piece.

  • Cells read individual parts. Swapping one piece predictably switches a cell from a heat-shock program to an ER-stress program.
  • Throughput. Up to 50,000 conditions per run, plus the ActiveSeq co-assay that reads microscopy and RNA from the exact same nanowell.
  • Scale. The pilot is a sample; the platform is designed for millions of compounds and beyond.
Technical lens

Compositional factorization and the ActiveSeq co-assay

Across 4 libraries (ZEL024, ZEL028, ZEL031, ZEL039) and 5 cell backgrounds (8 contexts), 46,944 raw genes are depth-normalized onto a 6,000-gene surface and projected onto a pinned k = 32 semi-NMF basis (W ∈ ℝ32×6000).

  • In silico building-block knockout. Removing a synthon from the trained model reproduces its empirical marginal effect at median Spearman r = 0.89–0.94 (z = 13.8–19.2 versus shuffled null); Morgan-substructure decomposition reconstructs whole-block effects at r = 0.97–0.99.
  • Chemistry-blind recurrence. Clustering 2,394 HEK293 compounds on RNA alone yields batch-recurrent directions (cosine advantage 0.581, p = 0.005) with 57 of 4,308 post-hoc synthon enrichments up to 8.3×.
  • Exact-well pairing. QC-adjusted within-treatment image-to-RNA correlation is 0.109 with real pairs versus −0.001 with broken pairs across 2,287 held-out wells (35/35 compound averages, 64/70 strata). Pooling 32 held-out wells, joint image/RNA exceeds RNA alone by +2.86 and +10.48 percentage points in the two batch directions; single-well classification shows no consistent gain.
A compound is a recipe of positional building blocksThe same block recurs across hundreds of molecules and across libraries, so its cellular effect can be isolated. Structures are withheld inthis public edition; block identities are schematic.bb0bb1bb2ZEL028A3B7C1ZEL028A9B7C4ZEL031A2B7C6ZEL031A2B7C8Block B7 recurs acrossrecipes and libraries(median block: 461molecules)Single-swap pair: onlybb2 differs (5,077,542measured pairs)461median molecules perbuilding block~1,100other molecules sharing acompound’s rarest block5.08 Mmeasuredsingle-substitutioncompound pairs722,397unique recipes · 2,079,661nominal compound–contextstatesMeasured coverage of the nominal design grid190,699 measured profiles ≈ 9.2% of 2,079,661 nominalcompound–context states across 104,385 Bemis–Murcko scaffolds. Theunmeasured 90.8% is the space the models are asked to predict.
Recurring building blocks across recipes and libraries
A compound is a recipe of positional blocks. Shared blocks isolate a part’s effect; single-swap pairs expose activity cliffs. Structures are withheld in this public edition.

In the white paper: Sections 1–2 · Platform, grammar and the ActiveSeq co-assay · Figures 2–4

Pillar 2 · Internal and external benchmarks

Better prediction on unseen chemistry

Two internally generated benchmarking workstreams tested unseen recipes, unseen building blocks and active-learning screening compression, with a coverage check of the program basis on public primary cells (NeurIPS 2023 PBMCs).

Investor lens

Why these benchmarks matter for capital efficiency

  • Predicting molecules before making them. Models predict the cellular effect of unmade combinations up to r = 0.58, and when a molecule uses a chemical part never seen in training, the additive recipe model edges the industry-standard graph neural network (Chemprop v2.3.1).
  • Fewer measurements. In retrospective replays across five of six objectives, model-directed selection matched the response recovery of 1,400 random measurements using 667–933 selections: 33–52% fewer. HSPA5 recovery reached 42.8% versus 30.7% frozen and 14.8% random. Prospective campaigns will test the laboratory savings.
Technical lens

Quantitative summary and scope boundaries

  • Unseen building blocks (7,544 HEK293 compounds). Additive recipe model: program r = 0.363, R² = 0.109, 2.62× directional enrichment; Chemprop 2.3.1: 0.337, 0.089, 2.53×. Shared-context three-seed transformer: r = 0.398 and 2.86× using eight contexts, but program R² = −0.99 (ranking improves, calibration worsens).
  • Public primary cells (NeurIPS 2023 PBMCs). On 614 primary PBMC profiles (146 drugs × 6 cell types) the pinned k = 32 basis captures 17.59% of DE variance (32.29× random, 6.40× a permuted basis); a coverage check, not a prediction benchmark.

Head-to-head on unseen building blocks (HEK293 · ZEL024 fold 0 · 7,544 test molecules)

Every test molecule contains at least one building block withheld from training. Across all 40 context-fold splits, pretrained CheMeleon-2048 features also beat Morgan-2048 fingerprints 40/40. Bars span 0–0.5.

Chemprop v2.3.1 D-MPNNend-to-end graph network
r = 0.337 · 2.53×
Additive recipe modelbuilding-block identities only
r = 0.363 · 2.62×
Shared-context transformer3-seed ensemble, eight contexts
r = 0.398 · 2.86×

Second value: directional enrichment at a 5% selection budget. The transformer’s program R² is −0.99: ranking improves while effect-size calibration worsens. Source results.

Learn, select and improve the next batch (AEC7 · HSPA5 objective · 1,400-compound budget)

Recovery of the top-5% measured-response set after 200 initial compounds plus six rounds of 200. Bars span 0–50%.

Updated modelretrained after each round
42.8% recovery
Frozen modelinitial model, never updated
30.7% recovery
Random selection
14.8% recovery

Across five of six objectives, model-directed selection matched 1,400 random measurements with 667–933 selections (33–52% fewer). Replays use previously measured compounds. Source results.

In the white paper: Section 3 · Internal and external benchmarking · Figures 5–7

Pillars 3 & 4 · Cross-atlas chemical and CRISPR phenomimicry

Chemical twins and genetic twins: turning uncharacterized chemistry into testable target programs

Z-Screen’s phenomimicry engine matches pilot-dataset profiles against 13,046 clinical and preclinical drugs (LINCS L1000, Tahoe-100M, CIGS) and 43 CRISPR Perturb-seq signature sets (~18,800 genes, ~8 million single cells), with calibration on reference controls at every step.

Investor lens

Facial recognition for cellular states

When a molecule enters a cell it changes thousands of genes, producing a high-dimensional barcode. We compare that barcode against two reference libraries:

  • Chemical twins. Series whose responses resemble established drugs provide an immediate mechanism hypothesis. Compound identifiers and response evidence are public; structures are withheld.
  • Genetic twins. When a disease target has no existing drug, matching against cells where a gene was silenced by CRISPR finds small-molecule twins for historically undruggable targets across the genome.
Technical lens

Calibrated scoring and bipartite mutual specificity

  • Three calibration choices for chemical references. Signed contrastive projections (unsigned-rank averaging would cancel a perfect match to 50%); vehicle-based centering against 72 pure DMSO wells rather than the 90.7%-active ZIC008 panel mean; and plate-level standardization of CIGS (replicate cosine 0.088 → 0.168). Together they give a 0.822 median matched-panel rank on 17 qualified controls (p = 0.0005) and zero-shot MZ1 chromatin-class AUROC 0.799 (ARV-825 ranked #2 of 1,940).
  • CRISPR scoring calibration. About 45% of genes in 25,000-gene Perturb-seq pseudobulks have MAD = 0, so z-scores use a positive MAD floor with a pure-vehicle baseline (+0.087 paired compound lift over an unfloored cosine, p = 0.045); the selected control-contrast bimodal ensemble reaches a 0.580 compound median (p = 0.043) on 25 matched Tier-1 compounds.
  • Selected target recoveries. STM2457 → METTL3 (#4 of 18,293), STC-15 → METTL3 (#1 of 2,393), BGT226 → PIK3CA (#1 of 18,293), AZD-7624 → MAPK14 (#15 of 2,620). These are selected atlas-specific ranks; the compound-level benchmark summarizes performance across the matched control set.
Genetic-reference calibration and genome-wide discoveryA: compound-median rank of the direct target across genome-wide and essential-gene Perturb-seq references (25 matched Tier-1 intracellularcompounds). B: the strict discovery funnel.AScoring-method calibration (0.5 = chance)0.400.450.500.550.600.65chanceUnfloored cosineclean z, self-masked, no MAD floor0.4853Full cosineclean z across all genes0.5385MAD floor + vehicle baselinepositive MAD floor, vehicle or library baseline0.5408 *Deconfounded IDF cosinetop-PC stress deconfounding0.5581Bimodal Z-mutual ensemblecalibrated five-channel ensemble0.5486Control-contrast ensembleunblended control contrast (selected method)0.5801 ** Wilcoxon p < 0.05 versus 0.5. MAD floor + vehicle baseline versus unfloored cosine: pairedmedian change +0.0867 (p = 0.0452). Source: evidence/crispr_calibration.csv.BStrict discovery funnel (8 contexts)55,227,900(series, target) pairs tested2,626strict-gate discoveries; 0 syntheticfalse discoveries at FDR < 0.052,229with lineage support (84.9%, aselected-set fraction)178disjoint chemical program clusters10highest-ranked genetic-twin seriesChemical–genetic convergence among the 2,626:168Tier 1 · exact target138Tier 2 · complex455Tier 3 · pathway
Genetic-reference calibration and the discovery funnel
Scoring-method calibration on 25 matched compounds, and the path from 55.2 million tested pairs to the ten highest-ranked genetic-twin series.

In the white paper: Sections 4–5 · Chemical twins and genetic twins · Figures 8–11

Pillar 5 · High-confidence hypotheses from the pilot dataset

High-confidence mechanism hypotheses: 20 convergent series, 14 distinct mechanisms

Grouped by mechanism, the 20 series describe 14 distinct hypotheses: ten with a precedented drug class or with none at all, and four cytotoxic or stress-response mechanisms the engine recovers but does not rank. Three mechanisms are reached by more than one series, and the BET core BB_3717371053 is found by both tracks. These are high-confidence hypotheses from the pilot dataset, offered for collaborators to test. They are not Zafrens pipeline programs, and their order is not a statement of which targets Zafrens is advancing.

Ranked mechanism hypotheses

Ordered by convergent evidence, then by how crowded the precedented drug class is; the last four rows are recovered mechanisms that are not ranked. The series views list the underlying chemistry. “Experiment” is the assay that would test each hypothesis, not a planned study; statistics are in the white paper.
RankMechanism hypothesis · therapeutic contextSeries · context · core blockConvergent evidenceExperiment that would test it
1METTL3 / METTL14 / WTAP m6A writerAML and solid tumours; one clinical compound in the classCRISPR-SAR #9 · H1650 · ZEL024 · BB_9434530945 (17 active, 9 multi-well, 54 cliff controls); supporting series ZS-SAR-06 · H1650 · ZEL028 · BB_8596154155 (15 active, 74 cliffs)m6A-writer complex #3 of 1,225 targets; partner METTL14 Z 5.63; regioisomer cliff 0.82. ZS-SAR-06 mimics GSK3368715; STM2457 control Z 3.58; EZH2 KO Z 4.28METTL3/14 MTase-Glo; m6A/A LC-MS/MS
2BRD2 / BRD4 BET bromodomainhaematological malignancies; crowded precedented classCRISPR-SAR #1 · A549 · ZEL031 · BB_3717371053 (34 active, 33 multi-well, 19 cliffs); ZS-SAR-04 · THP1 · ZEL031 · same core (5 active, 3 multi-well)BRD2 KO #2 of 1,225, Z 16.4; partner ELL2; AZD5153 mimic with MZ1 control Z 4.10; one-carbon linker cliff 0.87; the only mechanism found by both tracksBRD2/4 BD1/BD2 TR-FRET; MYC repression
3CDK9 / P-TEFb transcriptional kinaseMYC- and MCL1-dependent leukaemia and lymphomaZS-SAR-01 · A549 · ZEL028 (15 active, 47 cliffs); ZS-SAR-07 · AEC7 · ZEL039 (5 active); CRISPR-SAR #6 · AEC7 · ZEL039 (6 active, 5 multi-well)roniciclib and LY2857785 mimics; palbociclib control Z 5.14 and 4.16; CDK9 KO Z 5.27 and 6.70; SUPT5H partner Z 7.85; three series in two cell linesCDK9/cyclin T1 NanoBRET; p-RNAPII Ser2
4TYK2 / STAT5Boral immunology; the only non-oncology hypothesis in the setCRISPR-SAR #5 · HEK293 · ZEL024 · BB_1193003673 with linker BB_5290020531 (12 active, 12 multi-well, 34 cliffs)TYK2 KO #6; STAT5B partner Z 2.91; three orthogonal channels; cliff 0.80TYK2 JH1/JH2 TR-FRET; p-STAT1/5 flow
5AR / ER axis with the p23/HSP90 co-chaperonebreast and prostate cancer; resistance mechanism rather than another antagonistZS-SAR-02 · A549 · ZEL031 · BB_3103020399 (14 active, 13 multi-well); ZS-SAR-08 · A549 · ZEL031 · BB_5574910360 (17 active, 16 multi-well)endoxifen and ailanthone mimics; endoxifen control Z 6.40 and 6.70; AR KO Z 6.43 and 6.83; best multi-well replication among the chemical twinsERα/AR PolarScreen; p23/HSP90 client stability; FKBP5/TFF1 qPCR
6RAF1 / MEK MAPK signallingKRAS- and BRAF-driven tumours; crowded precedented classZS-SAR-09 and ZS-SAR-10 · H1650 · ZEL024 (17 and 23 active; 17 and 20 multi-well)ZM 336372 mimic; ZM336372 and sorafenib controls; MAP2K1 KO Z 3.99; RAF1 KO Z 4.49RAF1/MEK1 TR-FRET; p-ERK1/2 AlphaLISA
7ACTL6A (BAF53A) / SWI–SNF chromatin remodellingACTL6A-amplified squamous cancers; no precedented drug classCRISPR-SAR #4 · A549 · ZEL031 · BB_7209113661 (13 active, 13 multi-well, 18 cliffs)ACTL6A KO #1 of 1,225, Z 11.9; cliff 0.87; partner ARID1A weak (Z 1.09), pathway-level supportSWI/SNF remodelling ATPase; ATAC-seq
8NUAK1 / ROCK cytoskeletal kinaseLKB1- and KRAS-mutant tumours, fibrosis; tool compounds only in the classZS-SAR-05 · H1650 · ZEL024 · BB_5272889191 (29 active, 11 multi-well, 33 cliffs)HTH-01-015 mimic and control Z 2.08; NUAK1 KO Z 3.11; cliff 0.90NUAK1/ROCK1 radiometric IC50; p-MYPT1
9ALK / MET / IGF1R receptor tyrosine kinasefusion-driven NSCLC; precedented classZS-SAR-03 · H1650 · ZEL028 · BB_3467217614 (11 active, no multi-well active, 87 cliffs)ceritinib mimic; AZD-3463 control Z 6.26; ALK KO Z 3.56; rescreen before any target assaydose–response rescreen, then ALK/IGF1R ADP-Glo
10ATR / CHK1 / WEE1 checkpointDDR-deficient tumours; crowded precedented classCRISPR-SAR #3 · HEK293 · ZEL024 · BB_6048600856 (12 active, 12 multi-well, 74 cliffs)DDR complex #6; Z 11.1; ATR partner Z 7.94; cliff 0.91ATR/CHK1 kinase IC50; p-CHK1 Ser345, γ-H2AX
—26S proteasome PSMA3 / PSMB5recovered, not ranked · established cytotoxic class (bortezomib, ixazomib)CRISPR-SAR #2 · AEC7 · ZEL039 (15 active, 15 multi-well)PSMA3 KO #2, Z 10.7; PSMB5 partner Z 7.9420S chymotrypsin-like activity
—Ribosomal stress → ATF4 (RPL28)recovered, not ranked · stress response at the ribosomeCRISPR-SAR #7 · A549 · ZEL031 (4 active, 4 multi-well)RPL28 KO #23; ATF4 partner Z 1.81OP-Puro translation; p-eIF2α / ATF4
—Mitotic spindle PLK1 / AURKA / CDK1recovered, not ranked · generic antimitoticCRISPR-SAR #8 · THP1 · ZEL031 (6 active, 3 multi-well)complex #11; CDK1 partner Z 5.04; cliff 0.98mitotic kinase panel; p-histone H3
—TOP1recovered, not ranked · payload-class cytotoxic mechanismCRISPR-SAR #10 · THP1 · ZEL031 (5 active, 4 multi-well)TOP1 KO #10; one orthogonal channel; cliff 0.59TOP1 relaxation; ICE cleavage complex

In the white paper: Section 6 · High-confidence hypotheses from the pilot dataset

Technical documentation and downloads

Companion technical white papers

The technical white paper explains the methods, principal benchmarks, nominated chemistry and next experiments, with all eleven figures and their captions. The evidence register links every underlying table to its source file and hash.

Open data. The pilot dataset is released on Hugging Face as Zafrens/zscreen-pilot, with per-file manifests and SHA-256 hashes. An interactive Functional Drug Atlas lets you browse a portion of the measured responses.

Volume II · Technical white paper · October 2026

Z-Screen: Combinatorial Perturbation Transcriptomics, Model Benchmarking and Multi-Atlas Phenomimicry

Brings together the pilot dataset, same-well image and RNA measurements, internal and public model benchmarks, 110 chemical-reference matches across 51 series, and 2,626 genetic-reference matches across 178 clusters, with the ranked mechanism hypotheses and the experiments that would test them.

Read the web editionDownload PDFEvidence register
Volume I · Companion architecture paper · September 2026

Recursive Self-Improvement in Therapeutics: Chemistry That Biology Can Read

Establishes the conceptual and experimental architecture of Zafrens’ dual learning loops: the candidate loop, which updates building-block weights after each measured batch (recovering 42.8% of top-5% chemistry versus 30.7% frozen and 14.8% random), and the procedure loop, which optimizes normalization, representations and acquisition rules on held-out biological objectives (known-target rank from 29.4% chance to 14.4%, p = 0.001).

Open the RSI white paper (PDF)