2026-09-26
Five-agent AutoScreen ranks CRISPR targets from 26 databases. Top-100 recovery on 320 screens: 6.4% vs 5.4% for the strongest agent; three NK-resistance genes confirmed wet-lab.
Functional genomics can already dump hundreds of candidate genes on a bench. CRISPR screens, RNA-seq, and GWAS all nominate hits. The bottleneck sits one step later: which genes deserve follow-up, and which few hundred to test when a genome-wide library is too expensive.
Current tools cover pieces. MAGeCK and DESeq2 produce statistical ranks with little biological context. Pathway databases and literature mining supply context, but the evidence lives in silos that biologists still reconcile by hand. LLM agents such as BioDiscovery Agent can propose perturbation panels from a text description, yet they tend to recycle textbook hub genes, scale poorly to genome-wide libraries, and leave almost no audit trail for why a target was chosen.
AutoScreen, from Le Cong's lab at Stanford and Mengdi Wang's group at Princeton, splits the job across five specialist agents. The default reasoning backend is GPT-5.4.
Two modes. Pre-screen Design takes only a study description and builds a library from scratch. Post-screen Analysis also ingests a MAGeCK or DESeq2 hit table and re-ranks measured genes by mixing the screening signal with external evidence.
From BioGRID ORCS the team kept 320 genome-wide CRISPR screens with complete metadata and a condition-specific contrast. Ground truth is the original analysis top 100. Every method saw only a text research goal and returned at most 500 genes.
| Method | top-100 recovery | top-500 recovery | MAP@100 vs random |
| AutoScreen | 6.4% | 13.4% | 275.43× |
| BioDiscovery Agent | 5.4% | 12.2% | 216.22× |
| GPT-5.4 | 3.8% | 9.1% | 126.30× |
| Random | 0.2% | n.a. | 1× |
Relative to the strongest agent, validated-hit recovery at top 100 is about 19% higher. Matching AutoScreen's true-positive count at the top-500 reference point takes a 1.2-fold larger library for BioDiscovery, 2.4-fold for GPT-5.4, and 50-fold for random. Hits unique to AutoScreen make up 3.32% of all ground-truth genes.
Two pathway cases show what the anti-hub step does. In autophagy screen 1593, BioDiscovery ranked non-hit hubs MTOR and AKT1 at 1 and 16; AutoScreen put WIPI2, ATG2A, and ATG9A in the top 15. In the K562 diphtheria-toxin screen, BioDiscovery stayed on the ganglioside receptor pathway; AutoScreen ranked DPH1 through DPH7 at positions 3 to 19.
Wet-lab work used immune evasion. After a K562 surfaceome CRISPRa screen called 317 significant genes, Post-screen moved MUC1 from rank 118 to 5, PDPN from 81 to 44, and LRRC15 from 1384 to 659. Three gRNAs per gene, three replicates: activating each gene raised resistance to primary human NK killing. On an A375 melanoma screen, CEACAM5 moved from 143 to 38 and CEACAM6 from 155 to 26.
The harder test is prospective design. The system got two text prompts with no prior screen data: B16F0-PMEL melanoma versus cytotoxic T cells, and MC38-OT1 colon cancer versus OT-I. Library generation took 35.54 minutes on average versus about 3 days for an expert. Each method nominated up to 500 genes from a shared 3,340-gene focused library. Ground truth is the intersection of MAGeCK and Broad GPP calls at FDR<0.1 (53 and 17 genes). AutoScreen recovered 75.5% and 82.4% (hit-weighted mean 77.1%) versus about 71.4% for the strongest general-purpose LLM, an 8% relative gain.
A Pre-screen pass over 500-plus public CRISPR, RNA-seq, and UK Biobank GWAS datasets produced the AutoScreen Resource Hub: 538 evidence-backed target reports covering about 21,000 papers and 98,000 knowledge-base queries.
For a CRISPR lab this is an auditable library-design and hit-rerank pipeline. Each rank traces back to which anti-hub list, which fusion step, and which knowledge-base record. The reasoning backend is swappable; across seven models on 207 shared screens, the ensemble setting almost always beat any single hub length.
The useful setting is narrow and concrete. When a genome-wide screen is too costly, primary cells are scarce, or combinatorial space explodes, shrinking the library from a text description has numbers behind it. Post-screen mode fits teams that already have a hit list full of statistically ranked genes they cannot interpret.
The relative gain is real. Absolute recovery is still low: 6.4% of genome-wide true hits in the top 100. This shrinks libraries. It does not replace a genome-wide screen.
The paper is candid. Knowledge bases favor well-studied genes, so poorly annotated biology is under-counted. Ground truth is the original screen's top 100, which rewards matching prior calls; about 25% of AutoScreen-only hits have not been validated at scale. Wet-lab tests cover three genes and remain small.
A few points get less weight in the text. The 19% relative lift from 5.4% to 6.4% is roughly one extra top-100 true hit per screen; statistical significance is not a large lab-level gap. LRRC15 at rank 659 is still deep in the list, and how the three validation genes were chosen is not pre-registered. BioDiscovery Agent was built for small panels, so a 500-gene bake-off is a mismatch. AutoScreen itself defaults to GPT-5.4, so the GPT-5.4 baseline comparison tests the framework, not a stronger model. Stanford and Princeton have filed a patent on this work.