Protocol

Virtual CRISPR Simple Benchmark: yes/no prediction of a screen hit

Evidence concern: excluded from comparisons

Appendix A.2 describes the random baseline as predicting hit with probability equal to 'the overall proportion of actual hits in the specific benchmark dataset'. With one pooled rate (485 of 1,175 = 41.3%, Table 3) the expected PPV, sensitivity, F1 and FPR are all about 0.41 and specificity and NPV about 0.59. Table 8 prints PPV 0.47, sensitivity 0.46, F1 0.47, FPR 0.36, specificity 0.64 and NPV 0.63. With each screen's own hit rate (Table 3) the expected values are 0.47, 0.47, 0.47, 0.37, 0.63 and 0.63, which match. One draw at the pooled rate would need sensitivity about 2 standard errors above and FPR about 3 below expectation at once. The baseline therefore probably uses per-screen hit rates, which is information a screen designer would not have before the screen. The expected values were computed by the reviewer from Table 3 counts and are not stored as results. Treat the random row as a per-screen-prevalence classifier, not a matched random-selection comparator.Virtual CRISPR: Can LLMs Predict CRISPR Screen Results? · Appendix A.2; Table 3; Table 1, row "Random Baseline"; Table 8, row "Random N/A N/A"

A system answers whether a stated CRISPR perturbation of a gene in a cell line causally produces a stated phenotype, scored against high-confidence hit labels from four BioGRID-ORCS screens.

28 evaluations · 100 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.

28 recorded evaluations, 100 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

28 evaluations · 100 results. Different protocols are not a single leaderboard.

Protocol: Virtual CRISPR Simple Benchmark: yes/no prediction of a screen hit · Dataset: Virtual CRISPR Simple Benchmark

Sorted by F1 (Simple Benchmark) (higher is better). The best value in each column is highlighted. Decimals are rounded for display; each value links to the printed value and its source.

Tested configuration	F1 (Simple Benchmark)	False positive rate (Simple Benchmark)	Negative predictive value (Simple Benchmark)	Precision (Simple Benchmark)	Recall (Simple Benchmark)	Specificity (Simple Benchmark)
Llama-2-7B, direct prompting, mean over prompting strategies (Song et al. 2025)Evaluation	0.58 ± 0.02 (best in column)	0.97 ± 0.05	Not reported	Not reported	Not reported	Not reported
GPT-4o-mini, direct prompting, mean over prompting strategies (Song et al. 2025)Evaluation	0.55 ± 0.04	0.77 ± 0.16	Not reported	Not reported	Not reported	Not reported
GPT-4o, 0-shot, with chain-of-thought (Song et al. 2025)Evaluation	0.53	0.25	0.68 (best in column)	0.58	0.49	0.75
GPT-4o, 2-shot (-/+), without chain-of-thought (Song et al. 2025)Evaluation	0.53	0.32	0.67	0.54	0.53 (best in column)	0.68
GPT-4o, 2-shot (+/-), without chain-of-thought (Song et al. 2025)Evaluation	0.53	0.24	0.68 (best in column)	0.58	0.48	0.76
Llama-3-70B, direct prompting, mean over prompting strategies (Song et al. 2025)Evaluation	0.53 ± 0.09	0.58 ± 0.24	Not reported	Not reported	Not reported	Not reported
Llama-2-13B, direct prompting, mean over prompting strategies (Song et al. 2025)Evaluation	0.51 ± 0.12	0.80 ± 0.32	Not reported	Not reported	Not reported	Not reported
GPT-4o, 1-shot (+), with chain-of-thought (Song et al. 2025)Evaluation	0.5	0.23	0.66	0.57	0.44	0.77
GPT-4o, 1-shot (+), without chain-of-thought (Song et al. 2025)Evaluation	0.49	0.23	0.66	0.56	0.43	0.77
GPT-4o, 0-shot, without chain-of-thought (Song et al. 2025)Evaluation	0.48	0.28	0.65	0.53	0.44	0.72
Llama-3-8B, direct prompting, mean over prompting strategies (Song et al. 2025)Evaluation	0.48 ± 0.24	0.85 ± 0.38	Not reported	Not reported	Not reported	Not reported
GPT-4o, direct prompting, mean over prompting strategies (Song et al. 2025)Evaluation	0.47 ± 0.06	0.22 ± 0.06	Not reported	Not reported	Not reported	Not reported
Llama-2-70B, direct prompting, mean over prompting strategies (Song et al. 2025)Evaluation	0.47 ± 0.20	0.71 ± 0.36	Not reported	Not reported	Not reported	Not reported
Random baseline, Simple Benchmark hit rate (Song et al. 2025)Evaluation	0.47	0.36	0.63	0.47	0.46	0.64
GPT-4o, 1-shot (-), with chain-of-thought (Song et al. 2025)Evaluation	0.46	0.2	0.65	0.57	0.39	0.8
GPT-4o, 2-shot (-/+), with chain-of-thought (Song et al. 2025)Evaluation	0.45	0.21	0.65	0.56	0.38	0.79
GPT-4, direct prompting, mean over prompting strategies (Song et al. 2025)Evaluation	0.44 ± 0.12	0.38 ± 0.21	Not reported	Not reported	Not reported	Not reported
GPT-4o, 2-shot (+/-), with chain-of-thought (Song et al. 2025)Evaluation	0.44	0.18	0.64	0.58	0.36	0.82
Llama-3.1-70B, direct prompting, mean over prompting strategies (Song et al. 2025)Evaluation	0.44 ± 0.14	0.38 ± 0.23	Not reported	Not reported	Not reported	Not reported
GPT-3.5-turbo, direct prompting, mean over prompting strategies (Song et al. 2025)Evaluation	0.42 ± 0.13	0.39 ± 0.25	Not reported	Not reported	Not reported	Not reported
Llama-3.3-70B, direct prompting, mean over prompting strategies (Song et al. 2025)Evaluation	0.40 ± 0.19	0.40 ± 0.31	Not reported	Not reported	Not reported	Not reported
Llama-3.1-8B, direct prompting, mean over prompting strategies (Song et al. 2025)Evaluation	0.39 ± 0.17	0.39 ± 0.27	Not reported	Not reported	Not reported	Not reported
Llama-3.2-1B, direct prompting, mean over prompting strategies (Song et al. 2025)Evaluation	0.37 ± 0.25	0.57 ± 0.43	Not reported	Not reported	Not reported	Not reported
GPT-4-turbo, direct prompting, mean over prompting strategies (Song et al. 2025)Evaluation	0.32 ± 0.10	0.15 ± 0.08	Not reported	Not reported	Not reported	Not reported
GPT-4o, 1-shot (-), without chain-of-thought (Song et al. 2025)Evaluation	0.31	0.09	0.62	0.62 (best in column)	0.21	0.91 (best in column)
o1-mini, direct prompting, mean over prompting strategies (Song et al. 2025)Evaluation	0.31 ± 0.04	0.10 ± 0.03	Not reported	Not reported	Not reported	Not reported
Llama-3.2-3B, direct prompting, mean over prompting strategies (Song et al. 2025)Evaluation	0.26 ± 0.24	0.28 ± 0.35	Not reported	Not reported	Not reported	Not reported
o1, direct prompting, mean over prompting strategies (Song et al. 2025)Evaluation	0.16 ± 0.02	0.03 ± 0.00 (best in column)	Not reported	Not reported	Not reported	Not reported

All 100 result rows with coverage, uncertainty and sources

Filter evaluations

Tested model or methodProtocolDatasetEvaluation setupMetricEvidence origin

Applied filters: All linked evaluations

Exact evaluated configurations and original reported results
Tested configuration	Protocol and dataset	Finding	Evidence and details
Configuration: GPT-3.5-turbo, direct prompting, mean over prompting strategies (Song et al. 2025)	Protocol: Virtual CRISPR Simple Benchmark: yes/no prediction of a screen hit Dataset: Virtual CRISPR Simple Benchmark	0.42 ± 0.13 F1 (Simple Benchmark) fraction · higher Uncertainty: SD 0.13. Caption: mean and std. dev. across up to 10 combinations of CoT and few-shot prompting; the number of combinations per model is not printed Coverage: Not reported scored / Not reported eligible	Independent external evaluation · Source checked Evidence concern: excluded from comparisons Appendix A.2 describes the random baseline as predicting hit with probability equal to 'the overall proportion of actual hits in the specific benchmark dataset'. With one pooled rate (485 of 1,175 = 41.3%, Table 3) the expected PPV, sensitivity, F1 and FPR are all about 0.41 and specificity and NPV about 0.59. Table 8 prints PPV 0.47, sensitivity 0.46, F1 0.47, FPR 0.36, specificity 0.64 and NPV 0.63. With each screen's own hit rate (Table 3) the expected values are 0.47, 0.47, 0.47, 0.37, 0.63 and 0.63, which match. One draw at the pooled rate would need sensitivity about 2 standard errors above and FPR about 3 below expectation at once. The baseline therefore probably uses per-screen hit rates, which is information a screen designer would not have before the screen. The expected values were computed by the reviewer from Table 3 counts and are not stored as results. Treat the random row as a per-screen-prevalence classifier, not a matched random-selection comparator.Virtual CRISPR: Can LLMs Predict CRISPR Screen Results? · Appendix A.2; Table 3; Table 1, row "Random Baseline"; Table 8, row "Random N/A N/A" Methods, coverage and source GPT-3.5-turbo (direct prompting, mean over strategies) on the Virtual CRISPR Simple Benchmark pheno-20261022-protocol-song2025-simple-hit-classification Aggregation: Not reported Virtual CRISPR: Can LLMs Predict CRISPR Screen Results? · Table 1, row "GPT-3.5-turbo", column "F1 Score"
Configuration: GPT-3.5-turbo, direct prompting, mean over prompting strategies (Song et al. 2025)	Protocol: Virtual CRISPR Simple Benchmark: yes/no prediction of a screen hit Dataset: Virtual CRISPR Simple Benchmark	0.39 ± 0.25 False positive rate (Simple Benchmark) fraction · lower Uncertainty: SD 0.25. Caption: mean and std. dev. across up to 10 combinations of CoT and few-shot prompting; the number of combinations per model is not printed Coverage: Not reported scored / Not reported eligible	Independent external evaluation · Source checked Evidence concern: excluded from comparisons Appendix A.2 describes the random baseline as predicting hit with probability equal to 'the overall proportion of actual hits in the specific benchmark dataset'. With one pooled rate (485 of 1,175 = 41.3%, Table 3) the expected PPV, sensitivity, F1 and FPR are all about 0.41 and specificity and NPV about 0.59. Table 8 prints PPV 0.47, sensitivity 0.46, F1 0.47, FPR 0.36, specificity 0.64 and NPV 0.63. With each screen's own hit rate (Table 3) the expected values are 0.47, 0.47, 0.47, 0.37, 0.63 and 0.63, which match. One draw at the pooled rate would need sensitivity about 2 standard errors above and FPR about 3 below expectation at once. The baseline therefore probably uses per-screen hit rates, which is information a screen designer would not have before the screen. The expected values were computed by the reviewer from Table 3 counts and are not stored as results. Treat the random row as a per-screen-prevalence classifier, not a matched random-selection comparator.Virtual CRISPR: Can LLMs Predict CRISPR Screen Results? · Appendix A.2; Table 3; Table 1, row "Random Baseline"; Table 8, row "Random N/A N/A" Methods, coverage and source GPT-3.5-turbo (direct prompting, mean over strategies) on the Virtual CRISPR Simple Benchmark pheno-20261022-protocol-song2025-simple-hit-classification Aggregation: Not reported Virtual CRISPR: Can LLMs Predict CRISPR Screen Results? · Table 1, row "GPT-3.5-turbo", column "FPR"
Configuration: GPT-4, direct prompting, mean over prompting strategies (Song et al. 2025)	Protocol: Virtual CRISPR Simple Benchmark: yes/no prediction of a screen hit Dataset: Virtual CRISPR Simple Benchmark	0.44 ± 0.12 F1 (Simple Benchmark) fraction · higher Uncertainty: SD 0.12. Caption: mean and std. dev. across up to 10 combinations of CoT and few-shot prompting; the number of combinations per model is not printed Coverage: Not reported scored / Not reported eligible	Independent external evaluation · Source checked Evidence concern: excluded from comparisons Appendix A.2 describes the random baseline as predicting hit with probability equal to 'the overall proportion of actual hits in the specific benchmark dataset'. With one pooled rate (485 of 1,175 = 41.3%, Table 3) the expected PPV, sensitivity, F1 and FPR are all about 0.41 and specificity and NPV about 0.59. Table 8 prints PPV 0.47, sensitivity 0.46, F1 0.47, FPR 0.36, specificity 0.64 and NPV 0.63. With each screen's own hit rate (Table 3) the expected values are 0.47, 0.47, 0.47, 0.37, 0.63 and 0.63, which match. One draw at the pooled rate would need sensitivity about 2 standard errors above and FPR about 3 below expectation at once. The baseline therefore probably uses per-screen hit rates, which is information a screen designer would not have before the screen. The expected values were computed by the reviewer from Table 3 counts and are not stored as results. Treat the random row as a per-screen-prevalence classifier, not a matched random-selection comparator.Virtual CRISPR: Can LLMs Predict CRISPR Screen Results? · Appendix A.2; Table 3; Table 1, row "Random Baseline"; Table 8, row "Random N/A N/A" Methods, coverage and source GPT-4 (direct prompting, mean over strategies) on the Virtual CRISPR Simple Benchmark pheno-20261022-protocol-song2025-simple-hit-classification Aggregation: Not reported Virtual CRISPR: Can LLMs Predict CRISPR Screen Results? · Table 1, row "GPT-4", column "F1 Score"
Configuration: GPT-4, direct prompting, mean over prompting strategies (Song et al. 2025)	Protocol: Virtual CRISPR Simple Benchmark: yes/no prediction of a screen hit Dataset: Virtual CRISPR Simple Benchmark	0.38 ± 0.21 False positive rate (Simple Benchmark) fraction · lower Uncertainty: SD 0.21. Caption: mean and std. dev. across up to 10 combinations of CoT and few-shot prompting; the number of combinations per model is not printed Coverage: Not reported scored / Not reported eligible	Independent external evaluation · Source checked Evidence concern: excluded from comparisons Appendix A.2 describes the random baseline as predicting hit with probability equal to 'the overall proportion of actual hits in the specific benchmark dataset'. With one pooled rate (485 of 1,175 = 41.3%, Table 3) the expected PPV, sensitivity, F1 and FPR are all about 0.41 and specificity and NPV about 0.59. Table 8 prints PPV 0.47, sensitivity 0.46, F1 0.47, FPR 0.36, specificity 0.64 and NPV 0.63. With each screen's own hit rate (Table 3) the expected values are 0.47, 0.47, 0.47, 0.37, 0.63 and 0.63, which match. One draw at the pooled rate would need sensitivity about 2 standard errors above and FPR about 3 below expectation at once. The baseline therefore probably uses per-screen hit rates, which is information a screen designer would not have before the screen. The expected values were computed by the reviewer from Table 3 counts and are not stored as results. Treat the random row as a per-screen-prevalence classifier, not a matched random-selection comparator.Virtual CRISPR: Can LLMs Predict CRISPR Screen Results? · Appendix A.2; Table 3; Table 1, row "Random Baseline"; Table 8, row "Random N/A N/A" Methods, coverage and source GPT-4 (direct prompting, mean over strategies) on the Virtual CRISPR Simple Benchmark pheno-20261022-protocol-song2025-simple-hit-classification Aggregation: Not reported Virtual CRISPR: Can LLMs Predict CRISPR Screen Results? · Table 1, row "GPT-4", column "FPR"
Configuration: GPT-4-turbo, direct prompting, mean over prompting strategies (Song et al. 2025)	Protocol: Virtual CRISPR Simple Benchmark: yes/no prediction of a screen hit Dataset: Virtual CRISPR Simple Benchmark	0.32 ± 0.10 F1 (Simple Benchmark) fraction · higher Uncertainty: SD 0.10. Caption: mean and std. dev. across up to 10 combinations of CoT and few-shot prompting; the number of combinations per model is not printed Coverage: Not reported scored / Not reported eligible	Independent external evaluation · Source checked Evidence concern: excluded from comparisons Appendix A.2 describes the random baseline as predicting hit with probability equal to 'the overall proportion of actual hits in the specific benchmark dataset'. With one pooled rate (485 of 1,175 = 41.3%, Table 3) the expected PPV, sensitivity, F1 and FPR are all about 0.41 and specificity and NPV about 0.59. Table 8 prints PPV 0.47, sensitivity 0.46, F1 0.47, FPR 0.36, specificity 0.64 and NPV 0.63. With each screen's own hit rate (Table 3) the expected values are 0.47, 0.47, 0.47, 0.37, 0.63 and 0.63, which match. One draw at the pooled rate would need sensitivity about 2 standard errors above and FPR about 3 below expectation at once. The baseline therefore probably uses per-screen hit rates, which is information a screen designer would not have before the screen. The expected values were computed by the reviewer from Table 3 counts and are not stored as results. Treat the random row as a per-screen-prevalence classifier, not a matched random-selection comparator.Virtual CRISPR: Can LLMs Predict CRISPR Screen Results? · Appendix A.2; Table 3; Table 1, row "Random Baseline"; Table 8, row "Random N/A N/A" Methods, coverage and source GPT-4-turbo (direct prompting, mean over strategies) on the Virtual CRISPR Simple Benchmark pheno-20261022-protocol-song2025-simple-hit-classification Aggregation: Not reported Virtual CRISPR: Can LLMs Predict CRISPR Screen Results? · Table 1, row "GPT-4-turbo", column "F1 Score"
Configuration: GPT-4-turbo, direct prompting, mean over prompting strategies (Song et al. 2025)	Protocol: Virtual CRISPR Simple Benchmark: yes/no prediction of a screen hit Dataset: Virtual CRISPR Simple Benchmark	0.15 ± 0.08 False positive rate (Simple Benchmark) fraction · lower Uncertainty: SD 0.08. Caption: mean and std. dev. across up to 10 combinations of CoT and few-shot prompting; the number of combinations per model is not printed Coverage: Not reported scored / Not reported eligible	Independent external evaluation · Source checked Evidence concern: excluded from comparisons Appendix A.2 describes the random baseline as predicting hit with probability equal to 'the overall proportion of actual hits in the specific benchmark dataset'. With one pooled rate (485 of 1,175 = 41.3%, Table 3) the expected PPV, sensitivity, F1 and FPR are all about 0.41 and specificity and NPV about 0.59. Table 8 prints PPV 0.47, sensitivity 0.46, F1 0.47, FPR 0.36, specificity 0.64 and NPV 0.63. With each screen's own hit rate (Table 3) the expected values are 0.47, 0.47, 0.47, 0.37, 0.63 and 0.63, which match. One draw at the pooled rate would need sensitivity about 2 standard errors above and FPR about 3 below expectation at once. The baseline therefore probably uses per-screen hit rates, which is information a screen designer would not have before the screen. The expected values were computed by the reviewer from Table 3 counts and are not stored as results. Treat the random row as a per-screen-prevalence classifier, not a matched random-selection comparator.Virtual CRISPR: Can LLMs Predict CRISPR Screen Results? · Appendix A.2; Table 3; Table 1, row "Random Baseline"; Table 8, row "Random N/A N/A" Methods, coverage and source GPT-4-turbo (direct prompting, mean over strategies) on the Virtual CRISPR Simple Benchmark pheno-20261022-protocol-song2025-simple-hit-classification Aggregation: Not reported Virtual CRISPR: Can LLMs Predict CRISPR Screen Results? · Table 1, row "GPT-4-turbo", column "FPR"
Configuration: GPT-4o, 0-shot, without chain-of-thought (Song et al. 2025)	Protocol: Virtual CRISPR Simple Benchmark: yes/no prediction of a screen hit Dataset: Virtual CRISPR Simple Benchmark	0.48 F1 (Simple Benchmark) fraction · higher Uncertainty: Not reported by the source Coverage: Not reported scored / Not reported eligible	Independent external evaluation · Source checked Evidence concern: excluded from comparisons Appendix A.2 describes the random baseline as predicting hit with probability equal to 'the overall proportion of actual hits in the specific benchmark dataset'. With one pooled rate (485 of 1,175 = 41.3%, Table 3) the expected PPV, sensitivity, F1 and FPR are all about 0.41 and specificity and NPV about 0.59. Table 8 prints PPV 0.47, sensitivity 0.46, F1 0.47, FPR 0.36, specificity 0.64 and NPV 0.63. With each screen's own hit rate (Table 3) the expected values are 0.47, 0.47, 0.47, 0.37, 0.63 and 0.63, which match. One draw at the pooled rate would need sensitivity about 2 standard errors above and FPR about 3 below expectation at once. The baseline therefore probably uses per-screen hit rates, which is information a screen designer would not have before the screen. The expected values were computed by the reviewer from Table 3 counts and are not stored as results. Treat the random row as a per-screen-prevalence classifier, not a matched random-selection comparator.Virtual CRISPR: Can LLMs Predict CRISPR Screen Results? · Appendix A.2; Table 3; Table 1, row "Random Baseline"; Table 8, row "Random N/A N/A" Methods, coverage and source GPT-4o, 0-shot, CoT N, on the Virtual CRISPR Simple Benchmark pheno-20261022-protocol-song2025-simple-hit-classification Aggregation: Not reported Virtual CRISPR: Can LLMs Predict CRISPR Screen Results? · Table 8, row "GPT-4o 0-shot N", column "F1"
Configuration: GPT-4o, 0-shot, without chain-of-thought (Song et al. 2025)	Protocol: Virtual CRISPR Simple Benchmark: yes/no prediction of a screen hit Dataset: Virtual CRISPR Simple Benchmark	0.28 False positive rate (Simple Benchmark) fraction · lower Uncertainty: Not reported by the source Coverage: Not reported scored / Not reported eligible	Independent external evaluation · Source checked Evidence concern: excluded from comparisons Appendix A.2 describes the random baseline as predicting hit with probability equal to 'the overall proportion of actual hits in the specific benchmark dataset'. With one pooled rate (485 of 1,175 = 41.3%, Table 3) the expected PPV, sensitivity, F1 and FPR are all about 0.41 and specificity and NPV about 0.59. Table 8 prints PPV 0.47, sensitivity 0.46, F1 0.47, FPR 0.36, specificity 0.64 and NPV 0.63. With each screen's own hit rate (Table 3) the expected values are 0.47, 0.47, 0.47, 0.37, 0.63 and 0.63, which match. One draw at the pooled rate would need sensitivity about 2 standard errors above and FPR about 3 below expectation at once. The baseline therefore probably uses per-screen hit rates, which is information a screen designer would not have before the screen. The expected values were computed by the reviewer from Table 3 counts and are not stored as results. Treat the random row as a per-screen-prevalence classifier, not a matched random-selection comparator.Virtual CRISPR: Can LLMs Predict CRISPR Screen Results? · Appendix A.2; Table 3; Table 1, row "Random Baseline"; Table 8, row "Random N/A N/A" Methods, coverage and source GPT-4o, 0-shot, CoT N, on the Virtual CRISPR Simple Benchmark pheno-20261022-protocol-song2025-simple-hit-classification Aggregation: Not reported Virtual CRISPR: Can LLMs Predict CRISPR Screen Results? · Table 8, row "GPT-4o 0-shot N", column "FPR"
Configuration: GPT-4o, 0-shot, without chain-of-thought (Song et al. 2025)	Protocol: Virtual CRISPR Simple Benchmark: yes/no prediction of a screen hit Dataset: Virtual CRISPR Simple Benchmark	0.65 Negative predictive value (Simple Benchmark) fraction · higher Uncertainty: Not reported by the source Coverage: Not reported scored / Not reported eligible	Independent external evaluation · Source checked Evidence concern: excluded from comparisons Appendix A.2 describes the random baseline as predicting hit with probability equal to 'the overall proportion of actual hits in the specific benchmark dataset'. With one pooled rate (485 of 1,175 = 41.3%, Table 3) the expected PPV, sensitivity, F1 and FPR are all about 0.41 and specificity and NPV about 0.59. Table 8 prints PPV 0.47, sensitivity 0.46, F1 0.47, FPR 0.36, specificity 0.64 and NPV 0.63. With each screen's own hit rate (Table 3) the expected values are 0.47, 0.47, 0.47, 0.37, 0.63 and 0.63, which match. One draw at the pooled rate would need sensitivity about 2 standard errors above and FPR about 3 below expectation at once. The baseline therefore probably uses per-screen hit rates, which is information a screen designer would not have before the screen. The expected values were computed by the reviewer from Table 3 counts and are not stored as results. Treat the random row as a per-screen-prevalence classifier, not a matched random-selection comparator.Virtual CRISPR: Can LLMs Predict CRISPR Screen Results? · Appendix A.2; Table 3; Table 1, row "Random Baseline"; Table 8, row "Random N/A N/A" Methods, coverage and source GPT-4o, 0-shot, CoT N, on the Virtual CRISPR Simple Benchmark pheno-20261022-protocol-song2025-simple-hit-classification Aggregation: Not reported Virtual CRISPR: Can LLMs Predict CRISPR Screen Results? · Table 8, row "GPT-4o 0-shot N", column "NPV"
Configuration: GPT-4o, 0-shot, without chain-of-thought (Song et al. 2025)	Protocol: Virtual CRISPR Simple Benchmark: yes/no prediction of a screen hit Dataset: Virtual CRISPR Simple Benchmark	0.53 Precision (Simple Benchmark) fraction · higher Uncertainty: Not reported by the source Coverage: Not reported scored / Not reported eligible	Independent external evaluation · Source checked Evidence concern: excluded from comparisons Appendix A.2 describes the random baseline as predicting hit with probability equal to 'the overall proportion of actual hits in the specific benchmark dataset'. With one pooled rate (485 of 1,175 = 41.3%, Table 3) the expected PPV, sensitivity, F1 and FPR are all about 0.41 and specificity and NPV about 0.59. Table 8 prints PPV 0.47, sensitivity 0.46, F1 0.47, FPR 0.36, specificity 0.64 and NPV 0.63. With each screen's own hit rate (Table 3) the expected values are 0.47, 0.47, 0.47, 0.37, 0.63 and 0.63, which match. One draw at the pooled rate would need sensitivity about 2 standard errors above and FPR about 3 below expectation at once. The baseline therefore probably uses per-screen hit rates, which is information a screen designer would not have before the screen. The expected values were computed by the reviewer from Table 3 counts and are not stored as results. Treat the random row as a per-screen-prevalence classifier, not a matched random-selection comparator.Virtual CRISPR: Can LLMs Predict CRISPR Screen Results? · Appendix A.2; Table 3; Table 1, row "Random Baseline"; Table 8, row "Random N/A N/A" Methods, coverage and source GPT-4o, 0-shot, CoT N, on the Virtual CRISPR Simple Benchmark pheno-20261022-protocol-song2025-simple-hit-classification Aggregation: Not reported Virtual CRISPR: Can LLMs Predict CRISPR Screen Results? · Table 8, row "GPT-4o 0-shot N", column "PPV"
Configuration: GPT-4o, 0-shot, without chain-of-thought (Song et al. 2025)	Protocol: Virtual CRISPR Simple Benchmark: yes/no prediction of a screen hit Dataset: Virtual CRISPR Simple Benchmark	0.44 Recall (Simple Benchmark) fraction · higher Uncertainty: Not reported by the source Coverage: Not reported scored / Not reported eligible	Independent external evaluation · Source checked Evidence concern: excluded from comparisons Appendix A.2 describes the random baseline as predicting hit with probability equal to 'the overall proportion of actual hits in the specific benchmark dataset'. With one pooled rate (485 of 1,175 = 41.3%, Table 3) the expected PPV, sensitivity, F1 and FPR are all about 0.41 and specificity and NPV about 0.59. Table 8 prints PPV 0.47, sensitivity 0.46, F1 0.47, FPR 0.36, specificity 0.64 and NPV 0.63. With each screen's own hit rate (Table 3) the expected values are 0.47, 0.47, 0.47, 0.37, 0.63 and 0.63, which match. One draw at the pooled rate would need sensitivity about 2 standard errors above and FPR about 3 below expectation at once. The baseline therefore probably uses per-screen hit rates, which is information a screen designer would not have before the screen. The expected values were computed by the reviewer from Table 3 counts and are not stored as results. Treat the random row as a per-screen-prevalence classifier, not a matched random-selection comparator.Virtual CRISPR: Can LLMs Predict CRISPR Screen Results? · Appendix A.2; Table 3; Table 1, row "Random Baseline"; Table 8, row "Random N/A N/A" Methods, coverage and source GPT-4o, 0-shot, CoT N, on the Virtual CRISPR Simple Benchmark pheno-20261022-protocol-song2025-simple-hit-classification Aggregation: Not reported Virtual CRISPR: Can LLMs Predict CRISPR Screen Results? · Table 8, row "GPT-4o 0-shot N", column "Sensitivity"
Configuration: GPT-4o, 0-shot, without chain-of-thought (Song et al. 2025)	Protocol: Virtual CRISPR Simple Benchmark: yes/no prediction of a screen hit Dataset: Virtual CRISPR Simple Benchmark	0.72 Specificity (Simple Benchmark) fraction · higher Uncertainty: Not reported by the source Coverage: Not reported scored / Not reported eligible	Independent external evaluation · Source checked Evidence concern: excluded from comparisons Appendix A.2 describes the random baseline as predicting hit with probability equal to 'the overall proportion of actual hits in the specific benchmark dataset'. With one pooled rate (485 of 1,175 = 41.3%, Table 3) the expected PPV, sensitivity, F1 and FPR are all about 0.41 and specificity and NPV about 0.59. Table 8 prints PPV 0.47, sensitivity 0.46, F1 0.47, FPR 0.36, specificity 0.64 and NPV 0.63. With each screen's own hit rate (Table 3) the expected values are 0.47, 0.47, 0.47, 0.37, 0.63 and 0.63, which match. One draw at the pooled rate would need sensitivity about 2 standard errors above and FPR about 3 below expectation at once. The baseline therefore probably uses per-screen hit rates, which is information a screen designer would not have before the screen. The expected values were computed by the reviewer from Table 3 counts and are not stored as results. Treat the random row as a per-screen-prevalence classifier, not a matched random-selection comparator.Virtual CRISPR: Can LLMs Predict CRISPR Screen Results? · Appendix A.2; Table 3; Table 1, row "Random Baseline"; Table 8, row "Random N/A N/A" Methods, coverage and source GPT-4o, 0-shot, CoT N, on the Virtual CRISPR Simple Benchmark pheno-20261022-protocol-song2025-simple-hit-classification Aggregation: Not reported Virtual CRISPR: Can LLMs Predict CRISPR Screen Results? · Table 8, row "GPT-4o 0-shot N", column "Specificity"
Configuration: GPT-4o, 0-shot, with chain-of-thought (Song et al. 2025)	Protocol: Virtual CRISPR Simple Benchmark: yes/no prediction of a screen hit Dataset: Virtual CRISPR Simple Benchmark	0.53 F1 (Simple Benchmark) fraction · higher Uncertainty: Not reported by the source Coverage: Not reported scored / Not reported eligible	Independent external evaluation · Source checked Evidence concern: excluded from comparisons Appendix A.2 describes the random baseline as predicting hit with probability equal to 'the overall proportion of actual hits in the specific benchmark dataset'. With one pooled rate (485 of 1,175 = 41.3%, Table 3) the expected PPV, sensitivity, F1 and FPR are all about 0.41 and specificity and NPV about 0.59. Table 8 prints PPV 0.47, sensitivity 0.46, F1 0.47, FPR 0.36, specificity 0.64 and NPV 0.63. With each screen's own hit rate (Table 3) the expected values are 0.47, 0.47, 0.47, 0.37, 0.63 and 0.63, which match. One draw at the pooled rate would need sensitivity about 2 standard errors above and FPR about 3 below expectation at once. The baseline therefore probably uses per-screen hit rates, which is information a screen designer would not have before the screen. The expected values were computed by the reviewer from Table 3 counts and are not stored as results. Treat the random row as a per-screen-prevalence classifier, not a matched random-selection comparator.Virtual CRISPR: Can LLMs Predict CRISPR Screen Results? · Appendix A.2; Table 3; Table 1, row "Random Baseline"; Table 8, row "Random N/A N/A" Methods, coverage and source GPT-4o, 0-shot, CoT Y, on the Virtual CRISPR Simple Benchmark pheno-20261022-protocol-song2025-simple-hit-classification Aggregation: Not reported Virtual CRISPR: Can LLMs Predict CRISPR Screen Results? · Table 8, row "GPT-4o 0-shot Y", column "F1"
Configuration: GPT-4o, 0-shot, with chain-of-thought (Song et al. 2025)	Protocol: Virtual CRISPR Simple Benchmark: yes/no prediction of a screen hit Dataset: Virtual CRISPR Simple Benchmark	0.25 False positive rate (Simple Benchmark) fraction · lower Uncertainty: Not reported by the source Coverage: Not reported scored / Not reported eligible	Independent external evaluation · Source checked Evidence concern: excluded from comparisons Appendix A.2 describes the random baseline as predicting hit with probability equal to 'the overall proportion of actual hits in the specific benchmark dataset'. With one pooled rate (485 of 1,175 = 41.3%, Table 3) the expected PPV, sensitivity, F1 and FPR are all about 0.41 and specificity and NPV about 0.59. Table 8 prints PPV 0.47, sensitivity 0.46, F1 0.47, FPR 0.36, specificity 0.64 and NPV 0.63. With each screen's own hit rate (Table 3) the expected values are 0.47, 0.47, 0.47, 0.37, 0.63 and 0.63, which match. One draw at the pooled rate would need sensitivity about 2 standard errors above and FPR about 3 below expectation at once. The baseline therefore probably uses per-screen hit rates, which is information a screen designer would not have before the screen. The expected values were computed by the reviewer from Table 3 counts and are not stored as results. Treat the random row as a per-screen-prevalence classifier, not a matched random-selection comparator.Virtual CRISPR: Can LLMs Predict CRISPR Screen Results? · Appendix A.2; Table 3; Table 1, row "Random Baseline"; Table 8, row "Random N/A N/A" Methods, coverage and source GPT-4o, 0-shot, CoT Y, on the Virtual CRISPR Simple Benchmark pheno-20261022-protocol-song2025-simple-hit-classification Aggregation: Not reported Virtual CRISPR: Can LLMs Predict CRISPR Screen Results? · Table 8, row "GPT-4o 0-shot Y", column "FPR"
Configuration: GPT-4o, 0-shot, with chain-of-thought (Song et al. 2025)	Protocol: Virtual CRISPR Simple Benchmark: yes/no prediction of a screen hit Dataset: Virtual CRISPR Simple Benchmark	0.68 Negative predictive value (Simple Benchmark) fraction · higher Uncertainty: Not reported by the source Coverage: Not reported scored / Not reported eligible	Independent external evaluation · Source checked Evidence concern: excluded from comparisons Appendix A.2 describes the random baseline as predicting hit with probability equal to 'the overall proportion of actual hits in the specific benchmark dataset'. With one pooled rate (485 of 1,175 = 41.3%, Table 3) the expected PPV, sensitivity, F1 and FPR are all about 0.41 and specificity and NPV about 0.59. Table 8 prints PPV 0.47, sensitivity 0.46, F1 0.47, FPR 0.36, specificity 0.64 and NPV 0.63. With each screen's own hit rate (Table 3) the expected values are 0.47, 0.47, 0.47, 0.37, 0.63 and 0.63, which match. One draw at the pooled rate would need sensitivity about 2 standard errors above and FPR about 3 below expectation at once. The baseline therefore probably uses per-screen hit rates, which is information a screen designer would not have before the screen. The expected values were computed by the reviewer from Table 3 counts and are not stored as results. Treat the random row as a per-screen-prevalence classifier, not a matched random-selection comparator.Virtual CRISPR: Can LLMs Predict CRISPR Screen Results? · Appendix A.2; Table 3; Table 1, row "Random Baseline"; Table 8, row "Random N/A N/A" Methods, coverage and source GPT-4o, 0-shot, CoT Y, on the Virtual CRISPR Simple Benchmark pheno-20261022-protocol-song2025-simple-hit-classification Aggregation: Not reported Virtual CRISPR: Can LLMs Predict CRISPR Screen Results? · Table 8, row "GPT-4o 0-shot Y", column "NPV"
Configuration: GPT-4o, 0-shot, with chain-of-thought (Song et al. 2025)	Protocol: Virtual CRISPR Simple Benchmark: yes/no prediction of a screen hit Dataset: Virtual CRISPR Simple Benchmark	0.58 Precision (Simple Benchmark) fraction · higher Uncertainty: Not reported by the source Coverage: Not reported scored / Not reported eligible	Independent external evaluation · Source checked Evidence concern: excluded from comparisons Appendix A.2 describes the random baseline as predicting hit with probability equal to 'the overall proportion of actual hits in the specific benchmark dataset'. With one pooled rate (485 of 1,175 = 41.3%, Table 3) the expected PPV, sensitivity, F1 and FPR are all about 0.41 and specificity and NPV about 0.59. Table 8 prints PPV 0.47, sensitivity 0.46, F1 0.47, FPR 0.36, specificity 0.64 and NPV 0.63. With each screen's own hit rate (Table 3) the expected values are 0.47, 0.47, 0.47, 0.37, 0.63 and 0.63, which match. One draw at the pooled rate would need sensitivity about 2 standard errors above and FPR about 3 below expectation at once. The baseline therefore probably uses per-screen hit rates, which is information a screen designer would not have before the screen. The expected values were computed by the reviewer from Table 3 counts and are not stored as results. Treat the random row as a per-screen-prevalence classifier, not a matched random-selection comparator.Virtual CRISPR: Can LLMs Predict CRISPR Screen Results? · Appendix A.2; Table 3; Table 1, row "Random Baseline"; Table 8, row "Random N/A N/A" Methods, coverage and source GPT-4o, 0-shot, CoT Y, on the Virtual CRISPR Simple Benchmark pheno-20261022-protocol-song2025-simple-hit-classification Aggregation: Not reported Virtual CRISPR: Can LLMs Predict CRISPR Screen Results? · Table 8, row "GPT-4o 0-shot Y", column "PPV"
Configuration: GPT-4o, 0-shot, with chain-of-thought (Song et al. 2025)	Protocol: Virtual CRISPR Simple Benchmark: yes/no prediction of a screen hit Dataset: Virtual CRISPR Simple Benchmark	0.49 Recall (Simple Benchmark) fraction · higher Uncertainty: Not reported by the source Coverage: Not reported scored / Not reported eligible	Independent external evaluation · Source checked Evidence concern: excluded from comparisons Appendix A.2 describes the random baseline as predicting hit with probability equal to 'the overall proportion of actual hits in the specific benchmark dataset'. With one pooled rate (485 of 1,175 = 41.3%, Table 3) the expected PPV, sensitivity, F1 and FPR are all about 0.41 and specificity and NPV about 0.59. Table 8 prints PPV 0.47, sensitivity 0.46, F1 0.47, FPR 0.36, specificity 0.64 and NPV 0.63. With each screen's own hit rate (Table 3) the expected values are 0.47, 0.47, 0.47, 0.37, 0.63 and 0.63, which match. One draw at the pooled rate would need sensitivity about 2 standard errors above and FPR about 3 below expectation at once. The baseline therefore probably uses per-screen hit rates, which is information a screen designer would not have before the screen. The expected values were computed by the reviewer from Table 3 counts and are not stored as results. Treat the random row as a per-screen-prevalence classifier, not a matched random-selection comparator.Virtual CRISPR: Can LLMs Predict CRISPR Screen Results? · Appendix A.2; Table 3; Table 1, row "Random Baseline"; Table 8, row "Random N/A N/A" Methods, coverage and source GPT-4o, 0-shot, CoT Y, on the Virtual CRISPR Simple Benchmark pheno-20261022-protocol-song2025-simple-hit-classification Aggregation: Not reported Virtual CRISPR: Can LLMs Predict CRISPR Screen Results? · Table 8, row "GPT-4o 0-shot Y", column "Sensitivity"
Configuration: GPT-4o, 0-shot, with chain-of-thought (Song et al. 2025)	Protocol: Virtual CRISPR Simple Benchmark: yes/no prediction of a screen hit Dataset: Virtual CRISPR Simple Benchmark	0.75 Specificity (Simple Benchmark) fraction · higher Uncertainty: Not reported by the source Coverage: Not reported scored / Not reported eligible	Independent external evaluation · Source checked Evidence concern: excluded from comparisons Appendix A.2 describes the random baseline as predicting hit with probability equal to 'the overall proportion of actual hits in the specific benchmark dataset'. With one pooled rate (485 of 1,175 = 41.3%, Table 3) the expected PPV, sensitivity, F1 and FPR are all about 0.41 and specificity and NPV about 0.59. Table 8 prints PPV 0.47, sensitivity 0.46, F1 0.47, FPR 0.36, specificity 0.64 and NPV 0.63. With each screen's own hit rate (Table 3) the expected values are 0.47, 0.47, 0.47, 0.37, 0.63 and 0.63, which match. One draw at the pooled rate would need sensitivity about 2 standard errors above and FPR about 3 below expectation at once. The baseline therefore probably uses per-screen hit rates, which is information a screen designer would not have before the screen. The expected values were computed by the reviewer from Table 3 counts and are not stored as results. Treat the random row as a per-screen-prevalence classifier, not a matched random-selection comparator.Virtual CRISPR: Can LLMs Predict CRISPR Screen Results? · Appendix A.2; Table 3; Table 1, row "Random Baseline"; Table 8, row "Random N/A N/A" Methods, coverage and source GPT-4o, 0-shot, CoT Y, on the Virtual CRISPR Simple Benchmark pheno-20261022-protocol-song2025-simple-hit-classification Aggregation: Not reported Virtual CRISPR: Can LLMs Predict CRISPR Screen Results? · Table 8, row "GPT-4o 0-shot Y", column "Specificity"
Configuration: GPT-4o, 1-shot (-), without chain-of-thought (Song et al. 2025)	Protocol: Virtual CRISPR Simple Benchmark: yes/no prediction of a screen hit Dataset: Virtual CRISPR Simple Benchmark	0.31 F1 (Simple Benchmark) fraction · higher Uncertainty: Not reported by the source Coverage: Not reported scored / Not reported eligible	Independent external evaluation · Source checked Evidence concern: excluded from comparisons Appendix A.2 describes the random baseline as predicting hit with probability equal to 'the overall proportion of actual hits in the specific benchmark dataset'. With one pooled rate (485 of 1,175 = 41.3%, Table 3) the expected PPV, sensitivity, F1 and FPR are all about 0.41 and specificity and NPV about 0.59. Table 8 prints PPV 0.47, sensitivity 0.46, F1 0.47, FPR 0.36, specificity 0.64 and NPV 0.63. With each screen's own hit rate (Table 3) the expected values are 0.47, 0.47, 0.47, 0.37, 0.63 and 0.63, which match. One draw at the pooled rate would need sensitivity about 2 standard errors above and FPR about 3 below expectation at once. The baseline therefore probably uses per-screen hit rates, which is information a screen designer would not have before the screen. The expected values were computed by the reviewer from Table 3 counts and are not stored as results. Treat the random row as a per-screen-prevalence classifier, not a matched random-selection comparator.Virtual CRISPR: Can LLMs Predict CRISPR Screen Results? · Appendix A.2; Table 3; Table 1, row "Random Baseline"; Table 8, row "Random N/A N/A" Methods, coverage and source GPT-4o, 1-shot (-), CoT N, on the Virtual CRISPR Simple Benchmark pheno-20261022-protocol-song2025-simple-hit-classification Aggregation: Not reported Virtual CRISPR: Can LLMs Predict CRISPR Screen Results? · Table 8, row "GPT-4o 1-shot (-) N", column "F1"
Configuration: GPT-4o, 1-shot (-), without chain-of-thought (Song et al. 2025)	Protocol: Virtual CRISPR Simple Benchmark: yes/no prediction of a screen hit Dataset: Virtual CRISPR Simple Benchmark	0.09 False positive rate (Simple Benchmark) fraction · lower Uncertainty: Not reported by the source Coverage: Not reported scored / Not reported eligible	Independent external evaluation · Source checked Evidence concern: excluded from comparisons Appendix A.2 describes the random baseline as predicting hit with probability equal to 'the overall proportion of actual hits in the specific benchmark dataset'. With one pooled rate (485 of 1,175 = 41.3%, Table 3) the expected PPV, sensitivity, F1 and FPR are all about 0.41 and specificity and NPV about 0.59. Table 8 prints PPV 0.47, sensitivity 0.46, F1 0.47, FPR 0.36, specificity 0.64 and NPV 0.63. With each screen's own hit rate (Table 3) the expected values are 0.47, 0.47, 0.47, 0.37, 0.63 and 0.63, which match. One draw at the pooled rate would need sensitivity about 2 standard errors above and FPR about 3 below expectation at once. The baseline therefore probably uses per-screen hit rates, which is information a screen designer would not have before the screen. The expected values were computed by the reviewer from Table 3 counts and are not stored as results. Treat the random row as a per-screen-prevalence classifier, not a matched random-selection comparator.Virtual CRISPR: Can LLMs Predict CRISPR Screen Results? · Appendix A.2; Table 3; Table 1, row "Random Baseline"; Table 8, row "Random N/A N/A" Methods, coverage and source GPT-4o, 1-shot (-), CoT N, on the Virtual CRISPR Simple Benchmark pheno-20261022-protocol-song2025-simple-hit-classification Aggregation: Not reported Virtual CRISPR: Can LLMs Predict CRISPR Screen Results? · Table 8, row "GPT-4o 1-shot (-) N", column "FPR"
Configuration: GPT-4o, 1-shot (-), without chain-of-thought (Song et al. 2025)	Protocol: Virtual CRISPR Simple Benchmark: yes/no prediction of a screen hit Dataset: Virtual CRISPR Simple Benchmark	0.62 Negative predictive value (Simple Benchmark) fraction · higher Uncertainty: Not reported by the source Coverage: Not reported scored / Not reported eligible	Independent external evaluation · Source checked Evidence concern: excluded from comparisons Appendix A.2 describes the random baseline as predicting hit with probability equal to 'the overall proportion of actual hits in the specific benchmark dataset'. With one pooled rate (485 of 1,175 = 41.3%, Table 3) the expected PPV, sensitivity, F1 and FPR are all about 0.41 and specificity and NPV about 0.59. Table 8 prints PPV 0.47, sensitivity 0.46, F1 0.47, FPR 0.36, specificity 0.64 and NPV 0.63. With each screen's own hit rate (Table 3) the expected values are 0.47, 0.47, 0.47, 0.37, 0.63 and 0.63, which match. One draw at the pooled rate would need sensitivity about 2 standard errors above and FPR about 3 below expectation at once. The baseline therefore probably uses per-screen hit rates, which is information a screen designer would not have before the screen. The expected values were computed by the reviewer from Table 3 counts and are not stored as results. Treat the random row as a per-screen-prevalence classifier, not a matched random-selection comparator.Virtual CRISPR: Can LLMs Predict CRISPR Screen Results? · Appendix A.2; Table 3; Table 1, row "Random Baseline"; Table 8, row "Random N/A N/A" Methods, coverage and source GPT-4o, 1-shot (-), CoT N, on the Virtual CRISPR Simple Benchmark pheno-20261022-protocol-song2025-simple-hit-classification Aggregation: Not reported Virtual CRISPR: Can LLMs Predict CRISPR Screen Results? · Table 8, row "GPT-4o 1-shot (-) N", column "NPV"
Configuration: GPT-4o, 1-shot (-), without chain-of-thought (Song et al. 2025)	Protocol: Virtual CRISPR Simple Benchmark: yes/no prediction of a screen hit Dataset: Virtual CRISPR Simple Benchmark	0.62 Precision (Simple Benchmark) fraction · higher Uncertainty: Not reported by the source Coverage: Not reported scored / Not reported eligible	Independent external evaluation · Source checked Evidence concern: excluded from comparisons Appendix A.2 describes the random baseline as predicting hit with probability equal to 'the overall proportion of actual hits in the specific benchmark dataset'. With one pooled rate (485 of 1,175 = 41.3%, Table 3) the expected PPV, sensitivity, F1 and FPR are all about 0.41 and specificity and NPV about 0.59. Table 8 prints PPV 0.47, sensitivity 0.46, F1 0.47, FPR 0.36, specificity 0.64 and NPV 0.63. With each screen's own hit rate (Table 3) the expected values are 0.47, 0.47, 0.47, 0.37, 0.63 and 0.63, which match. One draw at the pooled rate would need sensitivity about 2 standard errors above and FPR about 3 below expectation at once. The baseline therefore probably uses per-screen hit rates, which is information a screen designer would not have before the screen. The expected values were computed by the reviewer from Table 3 counts and are not stored as results. Treat the random row as a per-screen-prevalence classifier, not a matched random-selection comparator.Virtual CRISPR: Can LLMs Predict CRISPR Screen Results? · Appendix A.2; Table 3; Table 1, row "Random Baseline"; Table 8, row "Random N/A N/A" Methods, coverage and source GPT-4o, 1-shot (-), CoT N, on the Virtual CRISPR Simple Benchmark pheno-20261022-protocol-song2025-simple-hit-classification Aggregation: Not reported Virtual CRISPR: Can LLMs Predict CRISPR Screen Results? · Table 8, row "GPT-4o 1-shot (-) N", column "PPV"
Configuration: GPT-4o, 1-shot (-), without chain-of-thought (Song et al. 2025)	Protocol: Virtual CRISPR Simple Benchmark: yes/no prediction of a screen hit Dataset: Virtual CRISPR Simple Benchmark	0.21 Recall (Simple Benchmark) fraction · higher Uncertainty: Not reported by the source Coverage: Not reported scored / Not reported eligible	Independent external evaluation · Source checked Evidence concern: excluded from comparisons Appendix A.2 describes the random baseline as predicting hit with probability equal to 'the overall proportion of actual hits in the specific benchmark dataset'. With one pooled rate (485 of 1,175 = 41.3%, Table 3) the expected PPV, sensitivity, F1 and FPR are all about 0.41 and specificity and NPV about 0.59. Table 8 prints PPV 0.47, sensitivity 0.46, F1 0.47, FPR 0.36, specificity 0.64 and NPV 0.63. With each screen's own hit rate (Table 3) the expected values are 0.47, 0.47, 0.47, 0.37, 0.63 and 0.63, which match. One draw at the pooled rate would need sensitivity about 2 standard errors above and FPR about 3 below expectation at once. The baseline therefore probably uses per-screen hit rates, which is information a screen designer would not have before the screen. The expected values were computed by the reviewer from Table 3 counts and are not stored as results. Treat the random row as a per-screen-prevalence classifier, not a matched random-selection comparator.Virtual CRISPR: Can LLMs Predict CRISPR Screen Results? · Appendix A.2; Table 3; Table 1, row "Random Baseline"; Table 8, row "Random N/A N/A" Methods, coverage and source GPT-4o, 1-shot (-), CoT N, on the Virtual CRISPR Simple Benchmark pheno-20261022-protocol-song2025-simple-hit-classification Aggregation: Not reported Virtual CRISPR: Can LLMs Predict CRISPR Screen Results? · Table 8, row "GPT-4o 1-shot (-) N", column "Sensitivity"
Configuration: GPT-4o, 1-shot (-), without chain-of-thought (Song et al. 2025)	Protocol: Virtual CRISPR Simple Benchmark: yes/no prediction of a screen hit Dataset: Virtual CRISPR Simple Benchmark	0.91 Specificity (Simple Benchmark) fraction · higher Uncertainty: Not reported by the source Coverage: Not reported scored / Not reported eligible	Independent external evaluation · Source checked Evidence concern: excluded from comparisons Appendix A.2 describes the random baseline as predicting hit with probability equal to 'the overall proportion of actual hits in the specific benchmark dataset'. With one pooled rate (485 of 1,175 = 41.3%, Table 3) the expected PPV, sensitivity, F1 and FPR are all about 0.41 and specificity and NPV about 0.59. Table 8 prints PPV 0.47, sensitivity 0.46, F1 0.47, FPR 0.36, specificity 0.64 and NPV 0.63. With each screen's own hit rate (Table 3) the expected values are 0.47, 0.47, 0.47, 0.37, 0.63 and 0.63, which match. One draw at the pooled rate would need sensitivity about 2 standard errors above and FPR about 3 below expectation at once. The baseline therefore probably uses per-screen hit rates, which is information a screen designer would not have before the screen. The expected values were computed by the reviewer from Table 3 counts and are not stored as results. Treat the random row as a per-screen-prevalence classifier, not a matched random-selection comparator.Virtual CRISPR: Can LLMs Predict CRISPR Screen Results? · Appendix A.2; Table 3; Table 1, row "Random Baseline"; Table 8, row "Random N/A N/A" Methods, coverage and source GPT-4o, 1-shot (-), CoT N, on the Virtual CRISPR Simple Benchmark pheno-20261022-protocol-song2025-simple-hit-classification Aggregation: Not reported Virtual CRISPR: Can LLMs Predict CRISPR Screen Results? · Table 8, row "GPT-4o 1-shot (-) N", column "Specificity"
Configuration: GPT-4o, 1-shot (-), with chain-of-thought (Song et al. 2025)	Protocol: Virtual CRISPR Simple Benchmark: yes/no prediction of a screen hit Dataset: Virtual CRISPR Simple Benchmark	0.46 F1 (Simple Benchmark) fraction · higher Uncertainty: Not reported by the source Coverage: Not reported scored / Not reported eligible	Independent external evaluation · Source checked Evidence concern: excluded from comparisons Appendix A.2 describes the random baseline as predicting hit with probability equal to 'the overall proportion of actual hits in the specific benchmark dataset'. With one pooled rate (485 of 1,175 = 41.3%, Table 3) the expected PPV, sensitivity, F1 and FPR are all about 0.41 and specificity and NPV about 0.59. Table 8 prints PPV 0.47, sensitivity 0.46, F1 0.47, FPR 0.36, specificity 0.64 and NPV 0.63. With each screen's own hit rate (Table 3) the expected values are 0.47, 0.47, 0.47, 0.37, 0.63 and 0.63, which match. One draw at the pooled rate would need sensitivity about 2 standard errors above and FPR about 3 below expectation at once. The baseline therefore probably uses per-screen hit rates, which is information a screen designer would not have before the screen. The expected values were computed by the reviewer from Table 3 counts and are not stored as results. Treat the random row as a per-screen-prevalence classifier, not a matched random-selection comparator.Virtual CRISPR: Can LLMs Predict CRISPR Screen Results? · Appendix A.2; Table 3; Table 1, row "Random Baseline"; Table 8, row "Random N/A N/A" Methods, coverage and source GPT-4o, 1-shot (-), CoT Y, on the Virtual CRISPR Simple Benchmark pheno-20261022-protocol-song2025-simple-hit-classification Aggregation: Not reported Virtual CRISPR: Can LLMs Predict CRISPR Screen Results? · Table 8, row "GPT-4o 1-shot (-) Y", column "F1"

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: 1
External evaluations: 27

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

Null control

Proposed control: requires review

No-change prediction under matched control conditions

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

Training-only mean-effect or linear prediction

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.

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

Virtual CRISPR: Can LLMs Predict CRISPR Screen Results? · Original source · Proceedings of the 24th Workshop on Biomedical Language Processing (BioNLP 2025), pages 354-364, 1 August 2025; ACL Anthology 2025.bionlp-1.30

Download this release (gzip)

Technical metadata and extraction receipts

Stable ID: pheno-20261022-protocol-song2025-simple-hit-classification

areas: cells-tissues
contexts: research
protocol: Each query is posed as 'Does {perturbation method} of {gene} in {cell line} causally result in {phenotype}?' and the system answers Yes (hit) or No (no-hit). Prompted language models are run with zero-shot, one- or two-shot and chain-of-thought prompts at temperature 0 with a fixed seed; the binary answer is extracted from the generated text. Metrics: F1 and false positive rate, with PPV, NPV, sensitivity and specificity for GPT-4o and the random baseline in Appendix Table 8.
task: Binary hit/no-hit prediction for gene-phenotype queries from CRISPR screens
version: Song et al., BioNLP 2025, Sections 2.2, 2.3, 4.1 and Appendix A.1-A.2
metric: f1-score
metric direction: higher
unit: fraction
metric definition: F1 is the harmonic mean of precision and recall for the hit class; FPR is the proportion of actual negatives predicted positive. Under the high-confidence definition a false positive asserts an effect opposite to the observed significant effect.
control: Random baseline predicting hit with probability equal to the benchmark's overall hit proportion (Appendix A.2)
limitations: Binary classification of pre-selected gene-phenotype queries; there is no candidate pool, ranking step or test budget, so it measures hit prediction rather than which perturbations to test.; Only genes with a statistically significant effect in either direction are queried; genes with no effect, which make up most of a screen, are not negatives.; Retrospective: labels come from published screens and no new experiment was run.; The four screens were selected by the authors from BioGRID-ORCS on domain knowledge, and are public, so prompted models may have seen them or their source papers in pre-training.; Table 1 values are means across up to 10 prompting-strategy combinations; the o1 family was run without explicit chain-of-thought prompts, so its n is smaller and not printed.
source locator: Sections 2.2, 2.3 and 4.1; Table 1; Appendix A.1, A.2, Tables 3 and 8

Related records

Suggest a correction