Research map / Probabilistic and scientific modeling / Molecules and proteins
Active Learning for Efficient Discovery of Optimal Combinatorial Perturbations
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes NAIAD, an active learning framework with an overparameterized single-gene effect encoder and data-size-adaptive gene embeddings, plus a maximum-predicted-effect acquisition, to find strong gene and drug combinations with fewer experiments.
Problem
Exhaustively testing combinatorial genetic or drug perturbations is infeasible, and models trained on small initial datasets overfit, so efficient discovery of the strongest synergistic pairs is needed.
Approach
A neural surrogate that combines over-parameterized single-gene effect inputs with gene embeddings whose dimensionality grows with training-set size, paired with a maximum-predicted-effect acquisition function inside an AI + lab active learning loop.
Key results
- At about 4 observations per gene, NAIAD achieved RMSE 5.1 on Norman vs 6.2 for linear and 7.1 for RECOVER, and was over 40% better than the second-best model on average across four datasets.
- After four active learning rounds with MPE sampling, NAIAD+MPE found 143 to 150 of the top 200 pairs on three datasets and 1.7x more than RECOVER+MPE on average.
- On drug combination screening, NAIAD consistently outperformed RECOVER in identifying the strongest drug pairs, with the gap widening from round 2.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use NAIAD instead of the linear model | NAIAD · Norman, gene frequency 4 | RMSE (×10−2): 6.2 (1.7) -> 5.1 (1.8) | Helped |
| Use NAIAD instead of the linear model | NAIAD · Simpson, gene frequency 4 | RMSE (×10−2): 3.3 (0.3) -> 2.2 (0.1) | Helped |
| Use NAIAD instead of the linear model | NAIAD · Horlbeck K562, gene frequency 4 | RMSE (×10−2): 6.4 (0.9) -> 6.1 (1.9) | Helped |
| Use NAIAD instead of the linear model | NAIAD · Horlbeck Jurkat, gene frequency 4 | RMSE (×10−2): 3.9 (0.6) -> 3.0 (0.6) | Helped |
| Use NAIAD instead of the strongest baseline | NAIAD · Norman, gene frequency 20 | RMSE (×10−2): 4.7 (0.5) -> 4.7 (0.1) | No clear effect |
| Use NAIAD instead of the strongest baseline | NAIAD · Simpson, gene frequency 20 | RMSE (×10−2): 1.9 (0.4) -> 1.9 (0.2) | No clear effect |
| Use NAIAD instead of RECOVER | NAIAD · Horlbeck K562, gene frequency 20 | RMSE (×10−2): 5.6 (1.0) -> 5.4 (0.6) | Helped |
| Use NAIAD instead of RECOVER | NAIAD · Horlbeck Jurkat, gene frequency 20 | RMSE (×10−2): 3.0 (0.4) -> 2.8 (0.6) | Helped |
| Use MPE rather than uniform sampling | NAIAD active learning · Norman, active learning round 4 | top 200 combinations discovered: 92 -> 144 | Helped |
| Use MPE rather than uniform sampling | NAIAD active learning · Simpson, active learning round 4 | top 200 combinations discovered: 66 -> 144 | Helped |
| Use MPE rather than uniform sampling | NAIAD active learning · Horlbeck K562, active learning round 4 | top 200 combinations discovered: 39 -> 94 | Helped |
| Use MPE rather than uniform sampling | NAIAD active learning · Horlbeck Jurkat, active learning round 4 | top 200 combinations discovered: 79 -> 161 | Helped |
Surprising findings
- MPE sampling gave worse overall test-set MSE than uncertainty-based methods, because it skews training data toward strong-effect pairs.
- Adding an uncertainty term to the MPE acquisition did not improve performance and reduced discovery of strong perturbations when weighted heavily.
Limitations
- Requires known individual effects of each gene; cannot predict combinations involving unseen genes.
- Does not predict gene expression profiles from single-cell perturbation data.
- Evaluated only on bulk cell-viability-style scalar phenotypes and pairwise (2-gene) combinations.
Experimental setup
Datasets: Norman et al. 2019 CRISPR combinatorial screen Simpson et al. 2023 CRISPR screen Horlbeck et al. 2018 K562 and Jurkat screens Zheng et al. 2021 / Bertin et al. 2023 drug combination dataset
Benchmarks: Norman Simpson Horlbeck K562 Horlbeck Jurkat Zheng/Bertin drug combination
Metrics: RMSE log MSE Pearson correlation true positive rate top-N prediction accuracy number of top 200 combinations discovered
Baselines: Linear model MLP GEARS RECOVER Uniform sampling UCB sampling Uncertainty sampling
Builds on: RECOVER GEARS Bayesian optimization deep ensembles
Closest papers by meaning
- Efficient Biological Data Acquisition through Inference Set Design (ICLR 2025 · similarity 0.83)
- Contextualizing biological perturbation experiments through language (ICLR 2025 · similarity 0.82)
- BioDiscoveryAgent: An AI Agent for Designing Genetic Perturbation Experiments (ICLR 2025 · similarity 0.82)
- Learning Cross-Domain Representations for Transferable Drug Perturbations on Single-Cell Transcriptional Responses (AAAI 2025 · similarity 0.81)
- PerturbReason: A Knowledge-Grounded Benchmark and Framework for Cell-State–Conditioned Mechanistic Reasoning of Perturbation Effects (NeurIPS 2026 · similarity 0.81)
- ScreenShot: A Foundation Model for Few-Shot Combination Drug Screening (NeurIPS 2026 · similarity 0.81)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.