Screen statistics, published context and inspectable evidence.
Public workspace access is not enabled. The analysis engine and research tools are implemented locally; workspace and demonstration features have different availability, stated below.
Explore published CRISPR screen results and their source metadata. Atlas records retain the original studies' analysis methods and hit definitions.
Status: Explorer implemented over BioGRID ORCS human screens; public console access disabled
Inspect effect sizes, statistical results, guide support and available Atlas evidence. A calibrated probability of independent validation is not available yet.
Status: Recorded-results report with CSV and JSON export; no calibration
Explore how library size and coverage change guide, cell and read requirements, with approximate cost and timing assumptions.
Status: Design arithmetic implemented, no account needed; no power estimate
Record follow-up outcomes alongside a screen's evidence. Validation records support review; automatic model retraining is not implemented.
Status: Outcome entry and review implemented; retraining not implemented
Read screen hit results through the REST API using a scoped organization key. A remote MCP server is not available yet.
Status: Read-only REST code; public workspace access disabled
A conceptual way to separate experimental support from how well a gene is studied. These categories guide review; they do not establish that a hit is real or provide a calibrated validation probability.
SplicR, built on published rankings, ties the frontier ensemble on the same library, 0.220 to 0.219. After the screen, it adds published-screen history to effect size: 9.2 of its top ten reproduce in an independent screen, up from 7.0 on effect size alone.
Chart: AssayBench test set, 334 screens published after 2021
9.2 of 10: 124 held-out screen pairs, a separate research-only task
Measured benchmark and real-data checks, including a failed research hypothesis. Inspect downloadable evidence, data provenance and uncertainty on the Evidence page.
In the archived olaparib reanalysis from raw reads, CHD1L ranked fourth. This is a check on one published screen, not a blind prospective validation.
The archived check reports a median per-guide CPM ratio of 0.97 across four samples. A ratio near one checks scale; it does not establish perfect guide-level agreement.
The archived check reproduces Fortin et al.'s per-gene counts for 20,872 genes. Annotation agreement does not measure artifact-detection accuracy in a new screen.
In the archived 500-guide fingerprint check, the source library matched fully and decoys stayed below 2%. This does not guarantee detection on every uploaded library.
Discuss a blinded evaluation with us. Predictions, outcome definitions and the analysis plan must be frozen before independent validation outcomes are revealed. This workflow still requires a laboratory partner.