Qin et al. 2025, Active Learning for Efficient Discovery of Optimal Combinatorial Perturbations (ICML 2025, PMLR 267)

What this source contributed

Records in this release that cite this source. A record can cite several sources.

• 4 methods
• 31 configurations
• 2 protocols
• 4 datasets
• 140 evaluations
• 212 results
• 8 evidence claims

Results

Values as the source printed them. Compare results only under compatible protocols, datasets and metrics.

Claims

Statements this source supports, each with where in the source it was checked.

• adaptation: pheno-20261022-qin2025-method-recover: Adaptation Adapted from drugs to genes by a bilinear projection that combines 128-dimensional gene embeddings; trained to predict the gene equivalent of Bliss scores (measured viability minus the product of single-gene viabilities) and converted back to viability. (Appendix B, 'RECOVER adaptation')
• architecture: pheno-20261022-qin2025-method-naiad: Architecture Y_i+j = phi([Y_i, Y_j] W1) A1^T + f(phi(W2 X_i), phi(W2 X_j)) A2^T: an over-parameterised encoding of the two single-gene effects plus a permutation-invariant (summed) combination of learned gene embeddings; the embedding dimension is increased on a fixed schedule as the average number of times each gene is seen grows. (Section 3.1, Eq. 1; Section 3.2; Table 9)
• code_availability: pheno-20261022-qin2025-method-naiad: Code availability Source code at https://github.com/NeptuneBio/NAIAD. (Abstract; Software and Data)
• Evidence summary: Select perturbations for a defined cellular response: Summary The evidence comes from one prospective challenge and two retrospective replays, which measure different things. In the Cancer Immunotherapy Data Science Challenge, scored prospectively in mouse T cells, the 20 original Challenge 2 submissions had a top-k overlap AUC of 0.452 to 0.566, with 17 of the 20 between 0.50 and 0.57; a random ordering is expected to score 0.5. That score is not a ROC AUC, and the genes it scores were drawn from the leading Challenge 1 nominations. In the follow-up, two nominated targets exceeded the Screen 1 targets on the immune checkpoint blockade objective; the source reports 61 nominations, 57 selected targets, 59 screen-2 genes and 50 perturbations passing quality control without reconciling them, so no success fraction follows. AssayBench-Loop replays 20 genome-wide CRISPR screens published after 2021, with each policy choosing 10 rounds of 100 genes. One random draw found 4.9% of each screen's hits on average (enrichment 1.04 over the random expectation). Language models used directly as the policy, with hit labels, reached enrichment of 1.39 to 4.81, BPMF 4.49, and the authors' handoffs from a language model to their AssayFormer up to 5.67, with the Gemini-3.1-Pro handoff finding 27.7% of hits. Many recovered hits are common-essential genes: 30.1% for that handoff, 18.6% for the random draw and 99.1% for a baseline that picks earlier screens' hits. Test screens were kept only if the best language model already passed a performance floor on them, and the paper does not say whether the models saw the screens' publications in training. The authors ran every row, and no uncertainty is printed. For two-gene combinations, NAIAD's replay of four viability screens counted how many of each screen's 200 strongest pairs the model ranked in its top 200 after four rounds: 143.0, 141.7, 99.7 and 150.0 when it acquired the pairs with the largest predicted effect, against 93.3, 70.3, 38.7 and 81.3 when it acquired pairs uniformly at random (means of three replicates). That count includes pairs the model predicted but never measured, and the uniform arm still uses the model's ranking, so it is not the yield of testing random pairs. The challenge's submissions scored close to a random ordering, while in both replays the leading model-guided policies scored above the random or uniform arm; the two differ in endpoint, candidate set, setting and who ran them, so their results cannot be compared directly. No source compares a model shortlist with random selection prospectively at the same budget. (ucc-research-source-challenge-supp-s2: Supplementary Table S2, media-3.xlsx, sheet 'Challenge 1 & 2', column D rows 2 to 24; ucc-research-source-challenge-paper: Abstract and Results Figure 5; ucpheno-20261022-source-edwards2026: Table 1, rows 'Random', 'GPT-5.6 Sol', 'BPMF [41]', 'Gemini-3.1-Pro → AssayFormer', 'GPT-5.6 Sol → AssayFormer' and 'Prior hit baseline', columns 'EF', 'FH (%)' and '%ess (%)'; Methods 4.2 and Appendix A (test-screen criteria); Supplementary Table 4, Base LLMs group, column 'EF'; pheno-20261022-qin2025-source-pmlr: Table 4, rows 'NAIAD MPE' and 'NAIAD Uniform')
• hardware: pheno-20261022-qin2025-method-naiad: Hardware All model training was done on a single Paperspace A100-80G server with 100 GB of RAM. (Appendix E)
• input_requirement: pheno-20261022-qin2025-method-naiad: Input requirement Assumes the individual effect of each gene on the phenotype is known; the current model cannot predict combinations involving genes whose single-gene effect is unseen. (Section 6, Discussion and Conclusion)
• Relevance of Combination phenotype prediction from small training sets: test RMSE at fixed average gene frequency to "Select perturbations for a defined cellular response": Assessed by Combination phenotype prediction from small training sets: test RMSE at fixed average gene frequency (pheno-20261022-qin2025-source-pmlr: Sections 4.2 and 5.2; Table 1; Table 6)
• Relevance of Retrospective active learning on a fully measured two-gene CRISPR viability screen: top-200 gene pairs found by round 4 to "Select perturbations for a defined cellular response": Assessed by Retrospective active learning on a fully measured two-gene CRISPR viability screen: top-200 gene pairs found by round 4 (pheno-20261022-qin2025-source-pmlr: Sections 4.3 and 5.3; Table 2; Table 3; Table 4; Tables 5, 7 and 8; Appendices C and D)

Models, benchmarks and other records

Evidence

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