Protocol

AssayBench-Loop adaptive hit discovery: 10 rounds of 100 genes on 20 held-out CRISPR screens

Retrospective replay in which a policy picks 100 genes per round for 10 rounds from a shared pool of about 20,000 genes, sees each screen's historical hit labels for its picks, and is scored on hits found against random selection.

53 evaluations · 530 results

Overview

An explanatory profile has not yet been reviewed for this record.

Consult the linked sources for architecture or protocol details. Missing evidence is not evidence of a missing capability.

53 recorded evaluations, 530 metric rows. A comparison chart has not yet been validated for these results. The table retains the individual findings and their sources.

View coverage and remaining gaps across all benchmarks

Results

Results are available, but no reviewed comparison panel is linked in this release.

All evaluations

53 evaluations · 530 results. Different protocols are not a single leaderboard.

Filter evaluations

Tested model or methodProtocolDatasetEvaluation setupMetricEvidence origin

Applied filters: All linked evaluations

Source checking is not independent reproduction. Release 2026-10-23-5347695500ed.

Methods and evaluation design

Procedure, tasks and evaluated configurations

Recorded evaluations

Each evaluation records what was tested and under which conditions.

Explore all linked results

Baseline coverage

Reference methods help show what a model adds beyond simple controls. We track a null control and a conventional method for each protocol.

0 of 2 active baseline roles have published Rewire measurements in this release. Measurements on a selected protocol do not establish coverage of an entire suite.

No execution recipe linked to this protocol. Recipe availability does not establish a completed evaluation.

Author-reported evaluations: 30
External evaluations: 23

Literature evidence is not a Rewire measurement. Executed but unpublished runs and private review status are not included.

Null control

Proposed control: requires review

Select a task-valid null control after reviewing inputs and metric

Protocol-specific applicability, permitted inputs, access, split, evaluator and execution requirements need review before implementation or execution.

This is a suggested selection rule, not a validated method or a measured score.

Conventional reference

Proposed control: requires review

Select an upstream conventional reference after reviewing the full protocol

Protocol-specific applicability, permitted inputs, access, split, evaluator and execution requirements need review before implementation or execution.

This is a suggested selection rule, not a validated method or a measured score.

Protocol coverage CSV (gzip) · Model evaluation matrix (gzip) · Source table (gzip) · Release and checksums (gzip)

Coverage is derived from release 2026-10-23-5347695500ed. Source citations describe the original records; they do not validate an unreviewed baseline proposal. No results have been generated by this audit.

Run instructions

No runnable recipe has been reviewed for this protocol. Dataset access, model requirements, licences and compute requirements must be checked against its sources before execution.

Applicable tests and references

Applicability is distinct from a completed evaluation.

Random policy on AssayBench-Loop (one draw) · source_supported

Evidence

Source checking verifies the cited claim or transcription. It does not establish independent reproduction.

Evidence table

Inspect claims, sources and review details

Trace each statement to its source and review. A context-only reference supports the record generally; it does not verify an individual field. Source checking does not reproduce an experiment.

One row per statement and cited source. Multiple citations are not independent evaluations. Shared locators are labelled explicitly.

0 evidence rows matching the loaded filters

Claims, original sources and review scope · Release 2026-10-23-5347695500ed
Property and statement	Original source and location	Review and provenance

No evidence rows match these filters. Choose another scope or clear the search.

How to read the evidence

Sources and history

Release 2026-10-23-5347695500ed · Record review: source checked

1 source record and release history

Biology-in-the-loop: Amortized Adaptive Hit Discovery in CRISPR Screens · Original source · arXiv:2609.11877 version 1, submitted 2026-09-10; not peer reviewed

Download this release (gzip)

Technical metadata and extraction receipts

Stable ID: ucpheno-20261022-protocol-edwards2026-assaybench-loop-test-10x100

areas: cells-tissues
contexts: research
protocol: For each test screen the policy receives a natural-language description of the screen. At each of T = 10 rounds it nominates b = 100 genes from the candidate pool G, observes their binary hit labels from the historical screen, and conditions the next batch on the accumulated history. Budget 1,000 acquisitions per screen (about 5% of the library). Methods may use the 1,349 training screens. Metrics are computed per screen and averaged equally over the 20 test screens.
version: arXiv:2609.11877v1, Section 2.1, Methods 4.1-4.3, Appendix H; Table 1 and Supplementary Table 4
metric: hit-enrichment-factor
metric direction: higher
unit: unitless
metric definition: EF = |G_T ∩ H| / ((N_L + N_Ḡ + N_miss) · h_rand), where G_T is the set of acquired genes, H the screen's hits, N_L the acquired genes in the screen library, N_Ḡ the hallucinated (invalid) gene names, N_miss the unfilled slots and h_rand = |H|/|L| the screen's hit rate. Valid genes outside the screen library are not counted in the denominator. Secondary: nAUC, fraction of hits found, shortfall, %ess, Vendi ratio, pathway overlap, EP-B/EP-S/EP-D (Appendix H).
aggregation: Each metric computed per screen, then averaged with equal weight over the 20 test screens (EP-D is computed over the pooled test set).
allowed information: Screen description (cell line, perturbation modality, treatment, readout, hit definition); the genes acquired in earlier rounds and their hit labels; the 1,349 training screens. Policies may ignore any of these (the label-ablation arms omit the hit labels).
heldout test: 20 genome-wide screens published after 2021 (temporal split)
heldout validation: 20 screens from 2021; used to select the handoff round k = 3
selection: Test screens filtered for genome-wide coverage, at least 50 hits, hit rate below 15%, and a best-LLM AnDCG@100 of at least 0.05, then chosen for phenotype spread and description diversity.
uncertainty reporting: printed: Table 1 and Supplementary Table 4 print means without spread.; runs: Not stated, except that Random is one draw.; note: Supplementary Figure 6D shows ±1 SEM across 20 screens for five methods as a figure only.
limitations: Retrospective: labels are historical screen calls, and the authors state the policies remain to be validated prospectively in a live experimental loop.; The held-out unit is the screen; there are 20 test screens and no per-screen values in the extracted tables.; Test screens were selected with a performance floor for the best LLM on AssayBench, which favours screens whose hits an LLM can name.; Recovered hits are often DepMap common-essential genes (the %ess column), so a high EF can partly reflect generic fitness genes.; Random is one draw of the random policy, not its expectation (Table 1 caption).; The AssayLoop, AssayFormer and AssayLLM rows are the authors' own methods; all other rows were also run by the same authors.
source locator: Section 2.1; Figure 2A and D captions; Methods 4.1-4.3; Appendix H; Table 1 caption

Related records

Suggest a correction

Exact evaluated configurations and original reported results
Tested configuration	Protocol and dataset	Finding	Evidence and details
Configuration: AssayFormer, BPMF embeddings, supervised only (AssayBench-Loop)	Protocol: AssayBench-Loop adaptive hit discovery: 10 rounds of 100 genes on 20 held-out CRISPR screens Dataset subset: AssayBench-Loop test split (20 genome-wide CRISPR screens published after 2021)	3.83 Hit enrichment factor (hits found divided by hits expected under random selection, after 10 rounds of 100 genes; hallucinated genes and unfilled slots count in the denominator, valid genes outside the screen library do not; mean over the 20 AssayBench-Loop test screens) unitless · higher Uncertainty: Not reported by the source: Table 1 and Supplementary Table 4 print means without spread. Coverage: 20 test screens scored	Author-reported evaluation · Source checked Methods, coverage and source AssayFormer, BPMF embeddings, supervised only on AssayBench-Loop test (Edwards et al. 2026) ucpheno-20261022-protocol-edwards2026-assaybench-loop-test-10x100 Aggregation: Not reported Biology-in-the-loop: Amortized Adaptive Hit Discovery in CRISPR Screens · Table 1, row "+ BPMF Embeddings" (group 'AssayFormer (Ours)'), column "EF (↑)"
Configuration: AssayFormer, BPMF embeddings, supervised only (AssayBench-Loop)	Protocol: AssayBench-Loop adaptive hit discovery: 10 rounds of 100 genes on 20 held-out CRISPR screens Dataset subset: AssayBench-Loop test split (20 genome-wide CRISPR screens published after 2021)	18.3 Effective number of groups (EP-B: Hill number of order 1 at batch scope, rarefied to 30 annotated genes per batch, one random group per gene per draw, 400 Monte Carlo draws, averaged within each screen then across the 20 test screens; not comparable with EP-S or EP-D) unitless · unknown Uncertainty: Not reported by the source: Table 1 and Supplementary Table 4 print means without spread. Coverage: 20 test screens scored (Appendix H excludes batches or screens with fewer annotated genes than the rarefaction size; how many were excluded is not reported.)	Author-reported evaluation · Source checked Methods, coverage and source AssayFormer, BPMF embeddings, supervised only on AssayBench-Loop test (Edwards et al. 2026) ucpheno-20261022-protocol-edwards2026-assaybench-loop-test-10x100 Aggregation: Not reported Biology-in-the-loop: Amortized Adaptive Hit Discovery in CRISPR Screens · Table 1, row "+ BPMF Embeddings" (group 'AssayFormer (Ours)'), column "EP-B"
Configuration: AssayFormer, BPMF embeddings, supervised only (AssayBench-Loop)	Protocol: AssayBench-Loop adaptive hit discovery: 10 rounds of 100 genes on 20 held-out CRISPR screens Dataset subset: AssayBench-Loop test split (20 genome-wide CRISPR screens published after 2021)	56.9 Effective number of groups (EP-D: Hill number of order 1 at dataset scope, rarefied to 6,000 annotated genes across the complete test set, one random group per gene per draw, 300 Monte Carlo draws; not comparable with EP-B or EP-S) unitless · unknown Uncertainty: Not reported by the source: Table 1 and Supplementary Table 4 print means without spread. Coverage: 20 test screens scored (One value pooled over the 20 test screens (Appendix H), not a per-screen mean as the Table 1 caption says of all metrics. Appendix H also excludes sample sets with fewer annotated genes than the rarefaction size; how many is not reported.)	Author-reported evaluation · Source checked Methods, coverage and source AssayFormer, BPMF embeddings, supervised only on AssayBench-Loop test (Edwards et al. 2026) ucpheno-20261022-protocol-edwards2026-assaybench-loop-test-10x100 Aggregation: Not reported Biology-in-the-loop: Amortized Adaptive Hit Discovery in CRISPR Screens · Table 1, row "+ BPMF Embeddings" (group 'AssayFormer (Ours)'), column "EP-D"
Configuration: AssayFormer, BPMF embeddings, supervised only (AssayBench-Loop)	Protocol: AssayBench-Loop adaptive hit discovery: 10 rounds of 100 genes on 20 held-out CRISPR screens Dataset subset: AssayBench-Loop test split (20 genome-wide CRISPR screens published after 2021)	42.7 Effective number of groups (EP-S: Hill number of order 1 at screen scope, rarefied to 200 annotated genes per screen, one random group per gene per draw, 400 Monte Carlo draws, averaged across the 20 test screens; not comparable with EP-B or EP-D) unitless · unknown Uncertainty: Not reported by the source: Table 1 and Supplementary Table 4 print means without spread. Coverage: 20 test screens scored (Appendix H excludes batches or screens with fewer annotated genes than the rarefaction size; how many were excluded is not reported.)	Author-reported evaluation · Source checked Methods, coverage and source AssayFormer, BPMF embeddings, supervised only on AssayBench-Loop test (Edwards et al. 2026) ucpheno-20261022-protocol-edwards2026-assaybench-loop-test-10x100 Aggregation: Not reported Biology-in-the-loop: Amortized Adaptive Hit Discovery in CRISPR Screens · Table 1, row "+ BPMF Embeddings" (group 'AssayFormer (Ours)'), column "EP-S"
Configuration: AssayFormer, BPMF embeddings, supervised only (AssayBench-Loop)	Protocol: AssayBench-Loop adaptive hit discovery: 10 rounds of 100 genes on 20 held-out CRISPR screens Dataset subset: AssayBench-Loop test split (20 genome-wide CRISPR screens published after 2021)	37.5% Proportion (share of acquired hits that are DepMap common-essential genes (1,827 genes); mean over the 20 AssayBench-Loop test screens) percent · unknown Uncertainty: Not reported by the source: Table 1 and Supplementary Table 4 print means without spread. Coverage: 20 test screens scored	Author-reported evaluation · Source checked Methods, coverage and source AssayFormer, BPMF embeddings, supervised only on AssayBench-Loop test (Edwards et al. 2026) ucpheno-20261022-protocol-edwards2026-assaybench-loop-test-10x100 Aggregation: Not reported Biology-in-the-loop: Amortized Adaptive Hit Discovery in CRISPR Screens · Table 1, row "+ BPMF Embeddings" (group 'AssayFormer (Ours)'), column "%ess (%)"
Configuration: AssayFormer, BPMF embeddings, supervised only (AssayBench-Loop)	Protocol: AssayBench-Loop adaptive hit discovery: 10 rounds of 100 genes on 20 held-out CRISPR screens Dataset subset: AssayBench-Loop test split (20 genome-wide CRISPR screens published after 2021)	18.3% Recall (fraction of the screen's hits acquired within 10 rounds of 100 genes (FH); mean over the 20 AssayBench-Loop test screens) percent · higher Uncertainty: Not reported by the source: Table 1 and Supplementary Table 4 print means without spread. Coverage: 20 test screens scored	Author-reported evaluation · Source checked Methods, coverage and source AssayFormer, BPMF embeddings, supervised only on AssayBench-Loop test (Edwards et al. 2026) ucpheno-20261022-protocol-edwards2026-assaybench-loop-test-10x100 Aggregation: Not reported Biology-in-the-loop: Amortized Adaptive Hit Discovery in CRISPR Screens · Table 1, row "+ BPMF Embeddings" (group 'AssayFormer (Ours)'), column "FH (%)(↑)"
Configuration: AssayFormer, BPMF embeddings, supervised only (AssayBench-Loop)	Protocol: AssayBench-Loop adaptive hit discovery: 10 rounds of 100 genes on 20 held-out CRISPR screens Dataset subset: AssayBench-Loop test split (20 genome-wide CRISPR screens published after 2021)	13.2 Normalized cumulative hits AUC (area under the fraction-of-hits against fraction-of-library-queried curve over 10 rounds, divided by an oracle's area over the same effective budget; mean over the 20 AssayBench-Loop test screens) unitless · higher Uncertainty: Not reported by the source: Table 1 and Supplementary Table 4 print means without spread. Coverage: 20 test screens scored	Author-reported evaluation · Source checked Methods, coverage and source AssayFormer, BPMF embeddings, supervised only on AssayBench-Loop test (Edwards et al. 2026) ucpheno-20261022-protocol-edwards2026-assaybench-loop-test-10x100 Aggregation: Not reported Biology-in-the-loop: Amortized Adaptive Hit Discovery in CRISPR Screens · Table 1, row "+ BPMF Embeddings" (group 'AssayFormer (Ours)'), column "nAUC (%)(↑)"
Configuration: AssayFormer, BPMF embeddings, supervised only, no screen description (AssayBench-Loop)	Protocol: AssayBench-Loop adaptive hit discovery: 10 rounds of 100 genes on 20 held-out CRISPR screens Dataset subset: AssayBench-Loop test split (20 genome-wide CRISPR screens published after 2021)	3.82 Hit enrichment factor (hits found divided by hits expected under random selection, after 10 rounds of 100 genes; hallucinated genes and unfilled slots count in the denominator, valid genes outside the screen library do not; mean over the 20 AssayBench-Loop test screens) unitless · higher Uncertainty: Not reported by the source: Table 1 and Supplementary Table 4 print means without spread. Coverage: 20 test screens scored	Author-reported evaluation · Source checked Methods, coverage and source AssayFormer, BPMF embeddings, supervised only, no screen description on AssayBench-Loop test (Edwards et al. 2026) ucpheno-20261022-protocol-edwards2026-assaybench-loop-test-10x100 Aggregation: Not reported Biology-in-the-loop: Amortized Adaptive Hit Discovery in CRISPR Screens · Supplementary Table 4, row "- screen desc." (group 'Ablations'), column "EF"
Configuration: AssayFormer, BPMF embeddings, supervised only, no screen description (AssayBench-Loop)	Protocol: AssayBench-Loop adaptive hit discovery: 10 rounds of 100 genes on 20 held-out CRISPR screens Dataset subset: AssayBench-Loop test split (20 genome-wide CRISPR screens published after 2021)	18.8 Effective number of groups (EP-B: Hill number of order 1 at batch scope, rarefied to 30 annotated genes per batch, one random group per gene per draw, 400 Monte Carlo draws, averaged within each screen then across the 20 test screens; not comparable with EP-S or EP-D) unitless · unknown Uncertainty: Not reported by the source: Table 1 and Supplementary Table 4 print means without spread. Coverage: 20 test screens scored (Appendix H excludes batches or screens with fewer annotated genes than the rarefaction size; how many were excluded is not reported.)	Author-reported evaluation · Source checked Methods, coverage and source AssayFormer, BPMF embeddings, supervised only, no screen description on AssayBench-Loop test (Edwards et al. 2026) ucpheno-20261022-protocol-edwards2026-assaybench-loop-test-10x100 Aggregation: Not reported Biology-in-the-loop: Amortized Adaptive Hit Discovery in CRISPR Screens · Supplementary Table 4, row "- screen desc." (group 'Ablations'), column "EP-B"
Configuration: AssayFormer, BPMF embeddings, supervised only, no screen description (AssayBench-Loop)	Protocol: AssayBench-Loop adaptive hit discovery: 10 rounds of 100 genes on 20 held-out CRISPR screens Dataset subset: AssayBench-Loop test split (20 genome-wide CRISPR screens published after 2021)	56.5 Effective number of groups (EP-D: Hill number of order 1 at dataset scope, rarefied to 6,000 annotated genes across the complete test set, one random group per gene per draw, 300 Monte Carlo draws; not comparable with EP-B or EP-S) unitless · unknown Uncertainty: Not reported by the source: Table 1 and Supplementary Table 4 print means without spread. Coverage: 20 test screens scored (One value pooled over the 20 test screens (Appendix H), not a per-screen mean as the Table 1 caption says of all metrics. Appendix H also excludes sample sets with fewer annotated genes than the rarefaction size; how many is not reported.)	Author-reported evaluation · Source checked Methods, coverage and source AssayFormer, BPMF embeddings, supervised only, no screen description on AssayBench-Loop test (Edwards et al. 2026) ucpheno-20261022-protocol-edwards2026-assaybench-loop-test-10x100 Aggregation: Not reported Biology-in-the-loop: Amortized Adaptive Hit Discovery in CRISPR Screens · Supplementary Table 4, row "- screen desc." (group 'Ablations'), column "EP-D"
Configuration: AssayFormer, BPMF embeddings, supervised only, no screen description (AssayBench-Loop)	Protocol: AssayBench-Loop adaptive hit discovery: 10 rounds of 100 genes on 20 held-out CRISPR screens Dataset subset: AssayBench-Loop test split (20 genome-wide CRISPR screens published after 2021)	43.2 Effective number of groups (EP-S: Hill number of order 1 at screen scope, rarefied to 200 annotated genes per screen, one random group per gene per draw, 400 Monte Carlo draws, averaged across the 20 test screens; not comparable with EP-B or EP-D) unitless · unknown Uncertainty: Not reported by the source: Table 1 and Supplementary Table 4 print means without spread. Coverage: 20 test screens scored (Appendix H excludes batches or screens with fewer annotated genes than the rarefaction size; how many were excluded is not reported.)	Author-reported evaluation · Source checked Methods, coverage and source AssayFormer, BPMF embeddings, supervised only, no screen description on AssayBench-Loop test (Edwards et al. 2026) ucpheno-20261022-protocol-edwards2026-assaybench-loop-test-10x100 Aggregation: Not reported Biology-in-the-loop: Amortized Adaptive Hit Discovery in CRISPR Screens · Supplementary Table 4, row "- screen desc." (group 'Ablations'), column "EP-S"
Configuration: AssayFormer, BPMF embeddings, supervised only, no screen description (AssayBench-Loop)	Protocol: AssayBench-Loop adaptive hit discovery: 10 rounds of 100 genes on 20 held-out CRISPR screens Dataset subset: AssayBench-Loop test split (20 genome-wide CRISPR screens published after 2021)	62.6% Proportion (share of acquired hits that are DepMap common-essential genes (1,827 genes); mean over the 20 AssayBench-Loop test screens) percent · unknown Uncertainty: Not reported by the source: Table 1 and Supplementary Table 4 print means without spread. Coverage: 20 test screens scored	Author-reported evaluation · Source checked Methods, coverage and source AssayFormer, BPMF embeddings, supervised only, no screen description on AssayBench-Loop test (Edwards et al. 2026) ucpheno-20261022-protocol-edwards2026-assaybench-loop-test-10x100 Aggregation: Not reported Biology-in-the-loop: Amortized Adaptive Hit Discovery in CRISPR Screens · Supplementary Table 4, row "- screen desc." (group 'Ablations'), column "Ess. (%)"
Configuration: AssayFormer, BPMF embeddings, supervised only, no screen description (AssayBench-Loop)	Protocol: AssayBench-Loop adaptive hit discovery: 10 rounds of 100 genes on 20 held-out CRISPR screens Dataset subset: AssayBench-Loop test split (20 genome-wide CRISPR screens published after 2021)	19.1% Recall (fraction of the screen's hits acquired within 10 rounds of 100 genes (FH); mean over the 20 AssayBench-Loop test screens) percent · higher Uncertainty: Not reported by the source: Table 1 and Supplementary Table 4 print means without spread. Coverage: 20 test screens scored	Author-reported evaluation · Source checked Methods, coverage and source AssayFormer, BPMF embeddings, supervised only, no screen description on AssayBench-Loop test (Edwards et al. 2026) ucpheno-20261022-protocol-edwards2026-assaybench-loop-test-10x100 Aggregation: Not reported Biology-in-the-loop: Amortized Adaptive Hit Discovery in CRISPR Screens · Supplementary Table 4, row "- screen desc." (group 'Ablations'), column "Frac. hits (%)"
Configuration: AssayFormer, BPMF embeddings, supervised only, no screen description (AssayBench-Loop)	Protocol: AssayBench-Loop adaptive hit discovery: 10 rounds of 100 genes on 20 held-out CRISPR screens Dataset subset: AssayBench-Loop test split (20 genome-wide CRISPR screens published after 2021)	13.1 Normalized cumulative hits AUC (area under the fraction-of-hits against fraction-of-library-queried curve over 10 rounds, divided by an oracle's area over the same effective budget; mean over the 20 AssayBench-Loop test screens) unitless · higher Uncertainty: Not reported by the source: Table 1 and Supplementary Table 4 print means without spread. Coverage: 20 test screens scored	Author-reported evaluation · Source checked Methods, coverage and source AssayFormer, BPMF embeddings, supervised only, no screen description on AssayBench-Loop test (Edwards et al. 2026) ucpheno-20261022-protocol-edwards2026-assaybench-loop-test-10x100 Aggregation: Not reported Biology-in-the-loop: Amortized Adaptive Hit Discovery in CRISPR Screens · Supplementary Table 4, row "- screen desc." (group 'Ablations'), column "nAUC (%)"
Configuration: AssayFormer, BPMF embeddings, supervised only, no screen description (AssayBench-Loop)	Protocol: AssayBench-Loop adaptive hit discovery: 10 rounds of 100 genes on 20 held-out CRISPR screens Dataset subset: AssayBench-Loop test split (20 genome-wide CRISPR screens published after 2021)	2.68 Annotation overlap ratio (mean pairwise Jaccard overlap of GO biological-process annotations within a batch, divided by that of an equal-size random batch from the screen library, averaged over batches; mean over the 20 AssayBench-Loop test screens) unitless · unknown Uncertainty: Not reported by the source: Table 1 and Supplementary Table 4 print means without spread. Coverage: 20 test screens scored	Author-reported evaluation · Source checked Methods, coverage and source AssayFormer, BPMF embeddings, supervised only, no screen description on AssayBench-Loop test (Edwards et al. 2026) ucpheno-20261022-protocol-edwards2026-assaybench-loop-test-10x100 Aggregation: Not reported Biology-in-the-loop: Amortized Adaptive Hit Discovery in CRISPR Screens · Supplementary Table 4, row "- screen desc." (group 'Ablations'), column "Path. Ov."
Configuration: AssayFormer, BPMF embeddings, supervised only, no screen description (AssayBench-Loop)	Protocol: AssayBench-Loop adaptive hit discovery: 10 rounds of 100 genes on 20 held-out CRISPR screens Dataset subset: AssayBench-Loop test split (20 genome-wide CRISPR screens published after 2021)	5.1% Proportion (shortfall: share of acquired genes outside the screen library, including hallucinated gene names; mean over the 20 AssayBench-Loop test screens) percent · lower Uncertainty: Not reported by the source: Table 1 and Supplementary Table 4 print means without spread. Coverage: 20 test screens scored	Author-reported evaluation · Source checked Methods, coverage and source AssayFormer, BPMF embeddings, supervised only, no screen description on AssayBench-Loop test (Edwards et al. 2026) ucpheno-20261022-protocol-edwards2026-assaybench-loop-test-10x100 Aggregation: Not reported Biology-in-the-loop: Amortized Adaptive Hit Discovery in CRISPR Screens · Supplementary Table 4, row "- screen desc." (group 'Ablations'), column "Shortfall (%)"
Configuration: AssayFormer, BPMF embeddings, supervised only, no screen description (AssayBench-Loop)	Protocol: AssayBench-Loop adaptive hit discovery: 10 rounds of 100 genes on 20 held-out CRISPR screens Dataset subset: AssayBench-Loop test split (20 genome-wide CRISPR screens published after 2021)	56.7 Vendi ratio (Vendi score of each batch under a cosine kernel on GenePT embeddings, divided by batch size and averaged over batches; mean over the 20 AssayBench-Loop test screens) unitless · unknown Uncertainty: Not reported by the source: Table 1 and Supplementary Table 4 print means without spread. Coverage: 20 test screens scored	Author-reported evaluation · Source checked Methods, coverage and source AssayFormer, BPMF embeddings, supervised only, no screen description on AssayBench-Loop test (Edwards et al. 2026) ucpheno-20261022-protocol-edwards2026-assaybench-loop-test-10x100 Aggregation: Not reported Biology-in-the-loop: Amortized Adaptive Hit Discovery in CRISPR Screens · Supplementary Table 4, row "- screen desc." (group 'Ablations'), column "Vendi (%)"
Configuration: AssayFormer, BPMF embeddings, supervised only (AssayBench-Loop)	Protocol: AssayBench-Loop adaptive hit discovery: 10 rounds of 100 genes on 20 held-out CRISPR screens Dataset subset: AssayBench-Loop test split (20 genome-wide CRISPR screens published after 2021)	2.54 Annotation overlap ratio (mean pairwise Jaccard overlap of GO biological-process annotations within a batch, divided by that of an equal-size random batch from the screen library, averaged over batches; mean over the 20 AssayBench-Loop test screens) unitless · unknown Uncertainty: Not reported by the source: Table 1 and Supplementary Table 4 print means without spread. Coverage: 20 test screens scored	Author-reported evaluation · Source checked Methods, coverage and source AssayFormer, BPMF embeddings, supervised only on AssayBench-Loop test (Edwards et al. 2026) ucpheno-20261022-protocol-edwards2026-assaybench-loop-test-10x100 Aggregation: Not reported Biology-in-the-loop: Amortized Adaptive Hit Discovery in CRISPR Screens · Supplementary Table 4, row "+ BPMF Embeddings" (group 'AssayFormer (Ours)'), column "Path. Ov."
Configuration: AssayFormer, BPMF embeddings, supervised only (AssayBench-Loop)	Protocol: AssayBench-Loop adaptive hit discovery: 10 rounds of 100 genes on 20 held-out CRISPR screens Dataset subset: AssayBench-Loop test split (20 genome-wide CRISPR screens published after 2021)	10.2% Proportion (shortfall: share of acquired genes outside the screen library, including hallucinated gene names; mean over the 20 AssayBench-Loop test screens) percent · lower Uncertainty: Not reported by the source: Table 1 and Supplementary Table 4 print means without spread. Coverage: 20 test screens scored	Author-reported evaluation · Source checked Methods, coverage and source AssayFormer, BPMF embeddings, supervised only on AssayBench-Loop test (Edwards et al. 2026) ucpheno-20261022-protocol-edwards2026-assaybench-loop-test-10x100 Aggregation: Not reported Biology-in-the-loop: Amortized Adaptive Hit Discovery in CRISPR Screens · Table 1, row "+ BPMF Embeddings" (group 'AssayFormer (Ours)'), column "SF (%)(↓)"
Configuration: AssayFormer, BPMF embeddings, supervised only (AssayBench-Loop)	Protocol: AssayBench-Loop adaptive hit discovery: 10 rounds of 100 genes on 20 held-out CRISPR screens Dataset subset: AssayBench-Loop test split (20 genome-wide CRISPR screens published after 2021)	52.5 Vendi ratio (Vendi score of each batch under a cosine kernel on GenePT embeddings, divided by batch size and averaged over batches; mean over the 20 AssayBench-Loop test screens) unitless · unknown Uncertainty: Not reported by the source: Table 1 and Supplementary Table 4 print means without spread. Coverage: 20 test screens scored	Author-reported evaluation · Source checked Methods, coverage and source AssayFormer, BPMF embeddings, supervised only on AssayBench-Loop test (Edwards et al. 2026) ucpheno-20261022-protocol-edwards2026-assaybench-loop-test-10x100 Aggregation: Not reported Biology-in-the-loop: Amortized Adaptive Hit Discovery in CRISPR Screens · Supplementary Table 4, row "+ BPMF Embeddings" (group 'AssayFormer (Ours)'), column "Vendi (%)"
Configuration: AssayFormer (BPMF embeddings, supervised then GRPO) (AssayBench-Loop)	Protocol: AssayBench-Loop adaptive hit discovery: 10 rounds of 100 genes on 20 held-out CRISPR screens Dataset subset: AssayBench-Loop test split (20 genome-wide CRISPR screens published after 2021)	4.83 Hit enrichment factor (hits found divided by hits expected under random selection, after 10 rounds of 100 genes; hallucinated genes and unfilled slots count in the denominator, valid genes outside the screen library do not; mean over the 20 AssayBench-Loop test screens) unitless · higher Uncertainty: Not reported by the source: Table 1 and Supplementary Table 4 print means without spread. Coverage: 20 test screens scored	Author-reported evaluation · Source checked Methods, coverage and source AssayFormer (BPMF embeddings, supervised then GRPO) on AssayBench-Loop test (Edwards et al. 2026) ucpheno-20261022-protocol-edwards2026-assaybench-loop-test-10x100 Aggregation: Not reported Biology-in-the-loop: Amortized Adaptive Hit Discovery in CRISPR Screens · Table 1, row "+ GRPO (= AssayFormer)" (group 'AssayFormer (Ours)'), column "EF (↑)"
Configuration: AssayFormer, BPMF embeddings, GRPO with terminal-EF reward (AssayBench-Loop)	Protocol: AssayBench-Loop adaptive hit discovery: 10 rounds of 100 genes on 20 held-out CRISPR screens Dataset subset: AssayBench-Loop test split (20 genome-wide CRISPR screens published after 2021)	4.29 Hit enrichment factor (hits found divided by hits expected under random selection, after 10 rounds of 100 genes; hallucinated genes and unfilled slots count in the denominator, valid genes outside the screen library do not; mean over the 20 AssayBench-Loop test screens) unitless · higher Uncertainty: Not reported by the source: Table 1 and Supplementary Table 4 print means without spread. Coverage: 20 test screens scored	Author-reported evaluation · Source checked Methods, coverage and source AssayFormer, BPMF embeddings, GRPO with terminal-EF reward on AssayBench-Loop test (Edwards et al. 2026) ucpheno-20261022-protocol-edwards2026-assaybench-loop-test-10x100 Aggregation: Not reported Biology-in-the-loop: Amortized Adaptive Hit Discovery in CRISPR Screens · Supplementary Table 4, row "BPMF + GRPO (EFterminal )" (group 'Ablations'), column "EF"
Configuration: AssayFormer, BPMF embeddings, GRPO with terminal-EF reward (AssayBench-Loop)	Protocol: AssayBench-Loop adaptive hit discovery: 10 rounds of 100 genes on 20 held-out CRISPR screens Dataset subset: AssayBench-Loop test split (20 genome-wide CRISPR screens published after 2021)	19 Effective number of groups (EP-B: Hill number of order 1 at batch scope, rarefied to 30 annotated genes per batch, one random group per gene per draw, 400 Monte Carlo draws, averaged within each screen then across the 20 test screens; not comparable with EP-S or EP-D) unitless · unknown Uncertainty: Not reported by the source: Table 1 and Supplementary Table 4 print means without spread. Coverage: 20 test screens scored (Appendix H excludes batches or screens with fewer annotated genes than the rarefaction size; how many were excluded is not reported.)	Author-reported evaluation · Source checked Methods, coverage and source AssayFormer, BPMF embeddings, GRPO with terminal-EF reward on AssayBench-Loop test (Edwards et al. 2026) ucpheno-20261022-protocol-edwards2026-assaybench-loop-test-10x100 Aggregation: Not reported Biology-in-the-loop: Amortized Adaptive Hit Discovery in CRISPR Screens · Supplementary Table 4, row "BPMF + GRPO (EFterminal )" (group 'Ablations'), column "EP-B"
Configuration: AssayFormer, BPMF embeddings, GRPO with terminal-EF reward (AssayBench-Loop)	Protocol: AssayBench-Loop adaptive hit discovery: 10 rounds of 100 genes on 20 held-out CRISPR screens Dataset subset: AssayBench-Loop test split (20 genome-wide CRISPR screens published after 2021)	62.7 Effective number of groups (EP-D: Hill number of order 1 at dataset scope, rarefied to 6,000 annotated genes across the complete test set, one random group per gene per draw, 300 Monte Carlo draws; not comparable with EP-B or EP-S) unitless · unknown Uncertainty: Not reported by the source: Table 1 and Supplementary Table 4 print means without spread. Coverage: 20 test screens scored (One value pooled over the 20 test screens (Appendix H), not a per-screen mean as the Table 1 caption says of all metrics. Appendix H also excludes sample sets with fewer annotated genes than the rarefaction size; how many is not reported.)	Author-reported evaluation · Source checked Methods, coverage and source AssayFormer, BPMF embeddings, GRPO with terminal-EF reward on AssayBench-Loop test (Edwards et al. 2026) ucpheno-20261022-protocol-edwards2026-assaybench-loop-test-10x100 Aggregation: Not reported Biology-in-the-loop: Amortized Adaptive Hit Discovery in CRISPR Screens · Supplementary Table 4, row "BPMF + GRPO (EFterminal )" (group 'Ablations'), column "EP-D"
Configuration: AssayFormer, BPMF embeddings, GRPO with terminal-EF reward (AssayBench-Loop)	Protocol: AssayBench-Loop adaptive hit discovery: 10 rounds of 100 genes on 20 held-out CRISPR screens Dataset subset: AssayBench-Loop test split (20 genome-wide CRISPR screens published after 2021)	45 Effective number of groups (EP-S: Hill number of order 1 at screen scope, rarefied to 200 annotated genes per screen, one random group per gene per draw, 400 Monte Carlo draws, averaged across the 20 test screens; not comparable with EP-B or EP-D) unitless · unknown Uncertainty: Not reported by the source: Table 1 and Supplementary Table 4 print means without spread. Coverage: 20 test screens scored (Appendix H excludes batches or screens with fewer annotated genes than the rarefaction size; how many were excluded is not reported.)	Author-reported evaluation · Source checked Methods, coverage and source AssayFormer, BPMF embeddings, GRPO with terminal-EF reward on AssayBench-Loop test (Edwards et al. 2026) ucpheno-20261022-protocol-edwards2026-assaybench-loop-test-10x100 Aggregation: Not reported Biology-in-the-loop: Amortized Adaptive Hit Discovery in CRISPR Screens · Supplementary Table 4, row "BPMF + GRPO (EFterminal )" (group 'Ablations'), column "EP-S"