TL;DR: A preprint posted to bioRxiv on Sept. 10, 2026 describes AutoScreen, an AI co-scientist system that designs CRISPR perturbation libraries from free-text descriptions and re-ranks screen hits with biological context, improving validated-hit recovery by about 19% in the top 100 across 320 genome-scale CRISPR screens.

AutoScreen Brings AI Co-Scientist Workflow to CRISPR Target Discovery

The newest verified fact is the preprint itself: researchers led by Stanford University posted “AutoScreen: AI Co-Scientist System for Target Discovery in Functional Genomics” to bioRxiv on Sept. 10, 2026, according to the bioRxiv record summarized by Exa, with an abstract imported by Lifescience.net on Sept. 11, 2026. The report introduces a system for gene-level hypothesis generation in functional genomics, a process the authors describe as still largely manual and time-consuming and lacking systematic tools for efficient and reproducible work.

AutoScreen addresses that gap with two linked functions, according to the bioRxiv abstract. The first function is Pre-screen Design, which constructs perturbation libraries de novo from free-text research descriptions. The second function is Post-screen Analysis, which re-ranks experimental screen hits by integrating statistical scores with biological context.

The distinction matters because the two stages are often handled separately. According to the bioRxiv abstract, Pre-screen Design starts before any experiment from a written research goal, while Post-screen Analysis starts after an experiment from statistical hit scores. AutoScreen places both stages in one AI co-scientist workflow for transparent, explainable gene prioritization with pipeline provenance.

What Pre-screen Design Does Before a Screen Runs

It starts from free-text research descriptions

Pre-screen Design constructs perturbation libraries de novo from free-text research descriptions, according to the bioRxiv abstract. That means a researcher describes a biological question in ordinary written language, and the system proposes a perturbation library for that question.

The abstract states that target discovery lacks systematic tools for efficient and reproducible gene-level hypothesis generation. Pre-screen Design is presented as the systematic entry point that produces a testable gene set before screening begins.

According to the bioRxiv record, the benchmark for this design task involved 320 genome-scale CRISPR screens described as expert-curated benchmarks. The system’s top 100 predictions and top-500 reference point are reported against those benchmarks.

It searches broadly before it proposes genes

AutoScreen leverages the complementary strengths of multiple specialist agents, according to the bioRxiv abstract. Those agents perform deep research into multi-modal evidence, information restructuring, parallel searches across more than 26 biomedical databases, evidence synthesis, and target review.

The parallel search across more than 26 biomedical databases is central to Pre-screen Design, according to the same abstract. Instead of relying on a single source, the agents query many biomedical databases in parallel and then restructure that information for library construction.

The abstract also states that the agents perform evidence synthesis and target review to provide transparent, explainable gene prioritization with pipeline provenance. Provenance here means the system retains the pipeline trail behind each prioritized gene, so a proposed perturbation library can be audited back to its supporting evidence.

What Post-screen Analysis Does After a Screen Runs

It does not accept raw statistical ranks as final

Post-screen Analysis re-ranks experimental screen hits by integrating statistical scores with biological context, according to the bioRxiv abstract. Raw statistical rankings remain the starting point, but biological context can move candidates up or down.

That re-ranking step is illustrated by the leukemia-screen case study reported in the abstract. According to the bioRxiv abstract, AutoScreen recovered natural killer resistance genes that were initially lower-ranked in a leukemia screen, moving MUC1, PDPN, and LRRC15 from raw ranks of 118, 81, and 1384 to ranks of 5, 44, and 659, respectively.

The change from 118 to 5 for MUC1, from 81 to 44 for PDPN, and from 1384 to 659 for LRRC15 shows the scale of re-ranking, according to the same abstract. A gene ranked 1,384th on statistics alone would ordinarily fall outside routine follow-up, while rank 659 remains low but is substantially higher, and ranks 5 and 44 place genes in a directly actionable range.

It connects re-ranked hits to follow-up perturbation

In follow-up tumor killing assay with primary human NK cells, individual perturbation validated all three hits successfully, according to the bioRxiv abstract. The three hits were MUC1, PDPN, and LRRC15 from the leukemia-screen re-ranking.

The use of primary human NK cells and individual perturbation is specified in the abstract. According to that abstract, each of the three genes was tested by individual perturbation in a tumor killing assay, and all three validated successfully.

The cancer immune-evasion case studies focused on natural killer and T-cell therapeutics, according to the bioRxiv abstract. The NK-cell work therefore represents one branch of a broader immune-evasion validation effort that also included cytotoxic T-cell-killing screens.

How the System Performed Across 320 Genome-Scale CRISPR Screens

Top-100 recovery rose about 19% over the strongest agent baseline

Across 320 genome-scale CRISPR screens as expert-curated benchmarks, AutoScreen achieved an approximately 19% increase in validated-hit recovery among its top 100 predictions, relative to the strongest agent baseline, according to the bioRxiv abstract. The comparison is to the strongest agent baseline, not to random selection and not to every possible method.

The top-100 window is the reported evaluation point for that gain, according to the same abstract. Validated-hit recovery among the top 100 predictions means the share of confirmed hits captured inside the system’s first 100 prioritized genes.

Lifescience.net, which imported the authors list and abstract from bioRxiv on Sept. 11, 2026, carries the same 19% figure and the same description of 320 genome-scale CRISPR screens as expert-curated benchmarks. That listing confirms the preprint server as bioRxiv and repeats the Pre-screen Design and Post-screen Analysis definitions.

A smaller library recovered the same number of hits at the top-500 point

AutoScreen required a 1.2-fold smaller library to recover the same number of hits at the top-500 reference point, according to the bioRxiv abstract. The 1.2-fold figure describes library efficiency at that top-500 point.

In practical terms reported by the authors, the same number of hits could be recovered with a smaller starting perturbation set. According to the abstract, that efficiency claim is tied specifically to the top-500 reference point in the 320-screen benchmark.

Precision was more than 100 times the random baseline

AutoScreen reached mean average precision more than two orders of magnitude above random baseline, according to the bioRxiv abstract. Two orders of magnitude corresponds to more than a 100-fold difference over random selection.

That random-baseline comparison is separate from the agent-baseline comparisons, according to the same abstract. The approximately 19% increase and the 8% improvement described below are measured against agent baselines, while the mean average precision claim is measured against random baseline.

Why the NK-Cell Leukemia Case Drew Attention

Three lower-ranked genes moved into contention

AutoScreen recovered NK-resistance genes that were initially lower-ranked in a leukemia screen, according to the bioRxiv abstract. The three named genes were MUC1, PDPN, and LRRC15.

The abstract reports raw ranks of 118, 81, and 1384 for MUC1, PDPN, and LRRC15, respectively. After AutoScreen re-ranking, those ranks became 5, 44, and 659, respectively, according to the same abstract.

The analysis integrated statistical scores with biological context rather than replacing statistics, according to the abstract’s definition of Post-screen Analysis. The resulting ranks therefore reflect combined statistical and biological evidence with pipeline provenance.

All three validated in primary human NK-cell tumor killing assays

In follow-up tumor killing assay with primary human NK cells, individual perturbation validated all three hits successfully, according to the bioRxiv abstract. That sentence establishes both the assay type and the validation outcome for MUC1, PDPN, and LRRC15.

The case study is described as part of cancer immune-evasion work focusing on natural killer and T-cell therapeutics, according to the same abstract. The NK-cell result therefore sits inside a therapeutic context aimed at understanding resistance to immune-cell killing.

No additional clinical claim is made in the abstract beyond successful validation by individual perturbation in that assay. According to the bioRxiv abstract, the broader conclusion is that the agentic approach makes target discovery more auditable, scalable, extensible, and reproducible.

How the System Fared in Prospective T-Cell Screens

It recovered 77.1% of consensus ground-truth hits

In prospective benchmarking across two cytotoxic T-cell-killing screens, AutoScreen recovered 77.1% of ground-truth hits called by the consensus of two gold-standard analysis pipelines with false discovery rate below 0.1, according to the bioRxiv abstract. The false discovery rate threshold is stated as FDR less than 0.1.

The ground truth in that test was defined by agreement between two pipelines, according to the same abstract. Only hits called by the consensus of those two gold-standard analysis pipelines counted toward the 77.1% figure.

The two screens were cytotoxic T-cell-killing screens, according to the abstract. That places the prospective test in the same cancer immune-evasion and T-cell therapeutics area as the NK-cell case study.

The gain was 8% over the strongest general-purpose LLM agent baseline

The 77.1% recovery achieved an 8% improvement over the strongest general-purpose large language model agent baseline, according to the bioRxiv abstract. That baseline is described as a general-purpose LLM agent baseline, distinct from the strongest agent baseline cited for the 320-screen benchmark.

Lifescience.net repeats the 77.1% recovery figure, the consensus of two gold-standard analysis pipelines at FDR below 0.1, and the 8% improvement over the strongest general-purpose LLM agent baseline. That mirrored abstract confirms the two-screen prospective design.

The prospective label indicates the test was framed as forward-looking benchmarking across those two screens, according to the bioRxiv abstract. The reported outcome is recovery of consensus hits, not a claim of a new therapy.

What the AutoScreen Resource Hub Contains

More than 500 public datasets were re-analyzed

The team re-analyzed more than 500 public datasets to construct the AutoScreen Resource Hub, according to the bioRxiv abstract. The hub is described as a growing knowledge base of pre-computed, annotated reports.

The re-analysis spans three data types, according to the same abstract. Those types are CRISPR screens, RNA-seq differential expression, and gene-level UK Biobank genome-wide association studies.

According to the abstract, the reports in the hub are pre-computed and annotated. That means users consult existing annotated reports for those CRISPR, RNA-seq, and UK Biobank GWAS analyses rather than starting each query from raw data.

The hub is meant to support continuous benchmark construction

This agentic AI approach enables real-time genomics benchmark construction that is continuously updated and expanded for evaluating frontier AI co-scientists, according to the bioRxiv abstract. The benchmarks are therefore described as living resources rather than fixed test sets.

The continuous updating and expansion apply to genomics benchmarks for AI co-scientist evaluation, according to the same abstract. The more than 500 re-analyzed datasets provide the starting material for that growing evaluation resource.

The overall claim for AutoScreen is that it enables AI-powered target discovery to be more auditable, scalable, extensible, and reproducible, according to the bioRxiv abstract. Auditability connects to pipeline provenance and target review, while scalability and extensibility connect to multi-agent searches and the growing Resource Hub.

Who Reported the Work and Where It Stands

A Stanford-led team posted the work as a bioRxiv preprint

The author list begins with Yuanhao Qu, Xufeng Liu, Xiaotong Wang, Meitong Chen, Xiao Luo, and Luna Lyu, and the corresponding author is Le Cong of Stanford University, according to the bioRxiv record. Other listed authors include Russ Biagio Altman and Jure Leskovec of Stanford University, Mengdi Wang of Princeton University, Ming Yin of Princeton University, Hanchen Wang of Genentech, Aviv Regev of Genentech, Kexin Huang of Stanford University, and collaborators from the University of North Carolina at Chapel Hill, University of California San Diego, University of Washington, Massachusetts Institute of Technology, Harvard University, and Acelegen Technology Inc. in Menlo Park, California.

The work appeared as a bioRxiv preprint with publication date Sept. 10, 2026, according to the Exa summary of the bioRxiv record, and the DOI is 10.64898/2026.09.06.749678 as listed by KnowThePaper. Lifescience.net states that its authors list and abstract were imported from bioRxiv on Sept. 11, 2026 and identifies the preprint server as bioRxiv.

As a preprint, the report had zero citations at the time of the Exa-indexed bioRxiv record, according to that record. That status is normal for a newly posted preprint and distinguishes the findings from peer-reviewed publication.

Related adaptive-screen work underscores the screening bottleneck

A separate Sept. 10, 2026 arXiv report on adaptive hit discovery in CRISPR screens describes exhaustive perturbation testing as often infeasible and candidate perturbations as needing prioritization over multiple experimental rounds, according to arXiv. That arXiv paper introduces AssayBench-Loop with 1,389 CRISPR screens across five phenotype categories and an AssayLoop framework that combines a transformer-based acquisition policy with LLM-derived biological priors.

The arXiv report states that AssayLoop achieved 5.67-fold enrichment over random selection and recovered 27.7% of hits after assaying approximately 5% of the candidate library on temporally held-out screens, according to arXiv. Those figures belong to the AssayLoop system, not to AutoScreen, and are cited here only to show parallel research on prioritization under constrained budgets.

Together, the two September reports describe different strategies for the same practical constraint, according to their respective abstracts. According to arXiv, AssayLoop learns acquisition policies across historical experiments, while according to the bioRxiv abstract, AutoScreen uses specialist agents, parallel database searches, evidence synthesis, and provenance-tracked re-ranking.

Why This Matters for Functional Genomics in the United States

For laboratories running genome-scale CRISPR screens, the reported 1.2-fold smaller library for the same hit recovery at the top-500 point points directly to reagent, sequencing, and cell-culture workload, because the bioRxiv abstract ties that efficiency to library size. A smaller library that still recovers the same number of hits would reduce the scale of each screen built from free-text descriptions.

For analysts reviewing screen output, Post-screen Analysis offers a second ranking that integrates statistical scores with biological context instead of leaving lower-ranked genes permanently buried, according to the bioRxiv abstract. The MUC1 move from 118 to 5 and the PDPN move from 81 to 44 illustrate how biological context changed the order, while the LRRC15 move from 1384 to 659 shows a large rank improvement that still left the gene outside the top tier before individual perturbation validated it.

For teams evaluating AI assistants, the Resource Hub’s pre-computed, annotated reports for CRISPR screens, RNA-seq differential expression, and gene-level UK Biobank genome-wide association studies provide a common starting point, according to the same abstract. Because the hub draws on more than 500 re-analyzed public datasets and is described as growing, it can supply fresh benchmarks that are continuously updated and expanded.

The work does not by itself establish a treatment, and the bioRxiv abstract makes no drug or clinical-outcome claim. What it documents is a reproducible workflow from research description to perturbation library to re-ranked hits to primary-cell assay validation, with pipeline provenance and multi-database evidence synthesis at each prioritization step.

Sources referenced