Zafrens/zscreen-pilot

Z-Screen Program Package

Version 1.5.1, 2026-09-04. A standalone screening-data release from the Z-Screen pilot combinatorial-chemistry screens. The data layer is per-compound 32-dimensional program-usage vectors and harmonized 6,000-gene response surfaces across 8 library × cell-line contexts, byproducts of Zafrens target-screening campaigns released for mining, defined against a pinned shared program basis. Five annexes (imaging, therapeutic hypotheses, chemistry, same-well, phenomimicry) carry the derived results, and a small reference model for usage prediction ships with its evaluation scores.

New in v1.5.0 (2026-08-31). A new annex_clusters/ ships the unsupervised census: 1,007 analog-family clusters recovered from the usage vectors across the 8 contexts, every one at q <= 0.01 against 200 size-matched nulls (median coherence z = 4.8), 855 of 1,007 carrying a significant Hallmark/KEGG/Reactome pathway at q < 0.05 (cluster_census.csv with per-cluster coherence and pathway annotations). The phenomimicry annex adds an ensemble rescoring (ens_min: median best percentile 0.018 vs 0.049 for full-gene cosine, 11 control pairs with at least half their cells in the top 20% vs 0 for cosine; ensemble_rescoring_panel.csv) with correlation-aware pair-level calibration. core/benchmark/ adds a fold-0 baseline comparison of transparent recipe-feature models against the reference transformer. A short platform summary now ships in docs/summary/, annex_hypotheses/LIMITS.md is renamed HOW_TO_READ.md, and consistency fixes run through the docs and annexes. If you downloaded an earlier version, re-fetch annex_clusters/, annex_phenomimicry/, annex_hypotheses/, core/benchmark/, and docs/.

New in v1.4.3 (2026-08-29). Anchor-lead external-triangulation figures are reconciled to the 1,000-draw null (annex_hypotheses/anchor_leads.csv), with minor text fixes in annex_phenomimicry/.

New in v1.4.2 (2026-08-29). The phenomimicry calibration now runs on a 1,000-draw random-target empirical null with a hub-matched companion null (validation_empirical_p.csv: p_emp, p_emp_hub), target-symbol normalization widens control coverage to all 35 named controls (55 scored control → target pairs, adding cobimetinib → MAP2K1 to the recovered set), and the ranked top-100 SAR-family table is rebuilt on families passing every annex guardrail (non-hub target, validated knockdown, non-promiscuous compounds, 3-200 recipe-mates), ranked by CRISPR-match strength, family size, and cross-context replication. If you downloaded an earlier version, re-fetch annex_phenomimicry/, annex_hypotheses/, and docs/.

New in v1.4.1 (2026-08-29). The CRISPR-knockout comparison has been expanded into a full annex, annex_phenomimicry/: every compound signature is now scored against 43 knockout signature sets from 10 public perturb-seq datasets (~18.8k target genes, ~8M perturbed cells), calibrated against a random-target empirical null (named-control pharmacology recovered across independent contexts and datasets, METTL3 and DOT1L inhibitor classes the strongest), with a per-target hub guardrail and a ranked top-100 SAR-family table. This replaces the previous single-atlas triage track, and the hypothesis ledger is correspondingly slimmer (1,027 rows).

The screens are a pilot relative to the chemical space the same libraries can generate. The objects that scale are the building-block grammar and the 32-program readout; that argument, and where this release sits next to LINCS, JUMP-CP, Tahoe-100M, Recursion, and DNA-encoded libraries, is in docs/WHY_THIS_MATTERS.md. Start there or with START_HERE.md. For the short platform summary and field position, see docs/summary/.

Code and documentation: https://github.com/Zafrens/zscreen-pilot
Full package (arrays and checkpoints): https://huggingface.co/datasets/Zafrens/zscreen-pilot
DOI: https://doi.org/10.5281/zenodo.22003566

Layout

START_HERE.md               tiered entry point: biology and data first, model-building second
docs/WHY_THIS_MATTERS.md    design argument, field comparison, recoverable biology
README.md                   this file
LICENSE.md                  software Apache-2.0; data and weights CC-BY-4.0
LICENSES/                   full Apache-2.0 and CC-BY-4.0 texts
NOTICE                      copyright notice (Zafrens, Inc.)
LICENSE_OR_DATA_USE.md      pointer to LICENSE.md (old name retained)
CITATION.cff                citation metadata
pyproject.toml              installable helper package (numpy/pandas/pyarrow; torch optional)
environment.lock            reference environment pins
verify.py                   one-command integrity + schema check -> src/zscreen_program_package
core/
  usages/                   per-context (compounds x 32) usage matrices + compound keys
  surfaces/                 per-context (compounds x 6,000) harmonized surfaces + panel
  basis/                    shared k=32 and k=12 bases + basis_registry.json (version pin)
  recipes.parquet           building-block grammar per public compound
  splits/                   fold_assignments.parquet (fold = SHA256(public_compound_id) mod 5)
  benchmark/                reference scores: per-context comparison, program-space primary,
                            k-resolution, correction arm, cross-context probe,
                            fold-0 baseline comparison (+ README)
models/                     reference model (evaluation grade): 3 checkpoints, model_def.py,
                            predict.py, bb_embedding_table.parquet, golden_predictions.json
annex_imaging/              per-compound image embeddings, marker intensities, zel039 latents,
                            reliability audits, fold-clean prediction dumps, decomposition
annex_hypotheses/           anchor_leads.csv, program_atlas.csv, sharp_sar_candidates.csv,
                            hypothesis_ledger_full.csv (1,027 rows, triage-grade), HOW_TO_READ.md
annex_phenomimicry/         calibrated compound x CRISPR-KO concordance: phenomimic pair
                            tables, top-100 SAR-family list, empirical-p validation,
                            ensemble rescoring panel, hub flags
annex_clusters/             unsupervised analog-family census: cluster_census.csv (1,007
                            null-calibrated clusters, pathway annotations) + README
annex_chemistry/            novel_bb_generalization.csv, attribution_certificate.csv,
                            activity_cliffs.csv, chemotype_series.csv, bb_effect_rankings.csv
annex_same_well/            same-well control study: 11,435 wells x 35 controls with paired
                            448-d image + 32-d RNA latents per well, evidence tables, README
docs/                       WHY_THIS_MATTERS, SCIENTIFIC_OVERVIEW, METHODS, DATA_DICTIONARY,
                            REPRODUCTION, terminology.json
docs/summary/               short platform summary and field position
src/zscreen_program_package/  thin loader + verification library (no model training code)
examples/                   4 notebooks: quickstart usages, reproduce benchmark,
                            browse hypotheses, join imaging
provenance/                 file manifest (frozen at release)

Two reading paths

• Guided readers / discovery: docs/WHY_THIS_MATTERS.md → START_HERE.md top half → annex_hypotheses/README.md → annex_imaging/README.md → annex_same_well/README.md → annex_chemistry/README.md → examples/03 and examples/04. Depth: docs/SCIENTIFIC_OVERVIEW.md.
• Model-builders: START_HERE.md bottom half → core/usages/ + core/basis/basis_registry.json → core/benchmark/ → models/README.md → examples/01 and examples/02. Depth: docs/METHODS.md, docs/DATA_DICTIONARY.md.

Benchmark scores throughout the package are evaluation-grade reference values: they describe the shipped data and reference configurations as measured, not tuning targets.

Install and quickstart

The data layers need only numpy, pandas, and pyarrow. The reference model additionally needs PyTorch (CPU is sufficient), available as the model extra.

pip install -e .          # helper loaders + verify CLI
pip install -e ".[model]" # adds torch for models/predict.py

or just use pandas/pyarrow directly. With the helpers installed:

from zscreen_program_package import data
usages, compounds = data.load_usages("zel039_aec7")   # (20813, 32), aligned keys
basis = data.load_basis(k=32)                         # (32, 6000), pinned v1
folds = data.load_folds()                             # context, public_compound_id, fold

Without installing, src/ can be put on sys.path (the root verify.py shim does exactly that). docs/REPRODUCTION.md has worked examples.

Verify

python verify.py           # fast: required files, schemas, shapes, ID formats
python verify.py --full    # additionally rechecks sha256 against provenance/file_manifest.csv

The fast check validates the row-alignment contract (every matrix against its compounds parquet), basis shapes, usage dimension 32, the 6,000-row panel, the 1,027-row hypothesis ledger, the same-well tables, and public-ID formats. --full rehashes every file once provenance/file_manifest.csv is present.

Citation and license

Cite per CITATION.cff (DOI 10.5281/zenodo.22003567; data at https://huggingface.co/datasets/Zafrens/zscreen-pilot). Software is Apache License 2.0. Data and model weights are CC BY 4.0. Chemical structures are not included. See LICENSE.md. Contact: hello@zafrens.com.