Research map / Probabilistic and scientific modeling / Molecules and proteins
Efficient Biological Data Acquisition through Inference Set Design
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 inference set design: a least-confidence active learning hybrid screen with a KL-based lower-bound stopping criterion that reduces the experimental cost of labeling compound libraries while meeting a target system accuracy.
Problem
Hybrid screens in drug discovery require labeling only a subset of a large finite compound library and predicting the rest; the goal is to minimize acquired labels while guaranteeing a target system accuracy.
Approach
Active selection of the least-confident (hardest) examples for labeling prunes difficult examples out of the finite inference set, so the remaining inference set is easier; a weak-calibration lemma lets the acquired batch accuracy serve as a conservative lower bound, enabling a KL-based stopping criterion.
Key results
- On the proprietary phenomics dataset, the least-confidence agent reaches the 98% target system accuracy at about 80% of the acquisition procedure, outperforming baselines.
- On QM9, LC and BALD agents reach near 100% inference-set and system accuracy after acquiring about 30% of the data, with saturating test accuracy across all agents.
- On Molecules3D regression, QBC achieves roughly five times lower MSE than random or heuristic orderings and stops after about 80% of data acquired.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Acquire low-confidence examples before easier examples | Active learning for hybrid screening · MNIST variations | system accuracy | Helped |
| Acquire low-confidence examples before easier examples | Active learning for hybrid screening · MNIST with shuffled labels (80% of data) | inference-set accuracy: 100% | Helped |
| Acquire low-confidence examples before easier examples | Active learning for hybrid screening · QM9 | system accuracy: near 100% after acquiring only 30% of data | Helped |
| Acquire compounds by least confidence | Active learning for hybrid screening · Proprietary phenomics dataset | system accuracy: 98% at around 80% of the acquisition procedure | Helped |
| Acquire using query-by-committee rather than random or heuristic ordering | Molecules3D regression · Molecules3D HOMO-LUMO gap and total energy | MSE: approximately five times lower MSE | Helped |
| Acquire compounds by least confidence or BALD rather than randomly | Active learning for hybrid screening · RxRx3 | accuracy | No clear effect |
| Acquire molecules from large to small | Molecules3D regression · Molecules3D total energy | inference-set prediction performance | Hurt |
| Use least-confidence batch accuracy as a conservative proxy for inference-set accuracy | Least-confidence stopping criterion · Classification datasets | system accuracy at stopping | Helped |
| Stop acquisition using a lower-bound accuracy threshold | Hybrid-screen acquisition · Classification datasets | system accuracy at stopping: 1%, 6%, and 0% | Helped |
Surprising findings
- Test-set generalization saturates for all agents while active agents gain largely on the inference set, so generalization metrics understate the benefit of inference set design.
- Heuristic orderings such as sorting molecules by size can harm both inference and system accuracy by introducing acquisition biases (e.g., missing chemical elements).
Limitations
- Assumes observed labels are deterministic and noise-free; noisy biological measurements are not modeled.
- On the RxRx3 dataset with low predictive power, active agents give only minor improvements over random sampling.
Experimental setup
Datasets: MNIST QM9 Molecules3D RxRx3 proprietary phenomics dataset
Benchmarks: MNIST QM9 Molecules3D RxRx3 proprietary phenomics dataset
Metrics: system accuracy inference set accuracy test accuracy MSE stopping time / fraction acquired
Baselines: random sampling heuristic orderings BALD query-by-committee diversity sampling (Tanimoto)
Builds on: pool-based active learning BALD core set selection Chernoff/KL concentration bound
Closest papers by meaning
- Enhancing Neural Subset Selection: Integrating Background Information into Set Representations (ICLR 2024 · similarity 0.86)
- ALINE: Joint Amortization for Bayesian Inference and Active Data Acquisition (NeurIPS 2025 · similarity 0.86)
- HORSE: Hierarchical Representation for Large-Scale Neural Subset Selection (NeurIPS 2024 · similarity 0.85)
- Monroe: A Molecular Foundation Model for In-context Probabilistic Inference (NeurIPS 2026 · similarity 0.85)
- Training Greedy Policy for Proposal Batch Selection in Expensive Multi-Objective Combinatorial Optimization (ICML 2024 · similarity 0.85)
- Multi-Objective Molecular Design Through Learning Latent Pareto Set (AAAI 2025 · similarity 0.84)
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.