Research map / Reinforcement learning / Visual and representation RL
Optimizing Gene-Based Testing for Antibiotic Resistance Prediction
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 GenoARM, which pairs a transformer AR predictor over PCR gene-presence tokens with a PPO-trained RL policy that selects 5 gene tests, using metadata to improve antibiotic resistance prediction on NCBI pathogen data.
Problem
Choosing a small, cost-effective subset of PCR gene tests that best predicts antibiotic resistance profiles of bacterial samples, when labels are missing and metadata (year, country, source) varies.
Approach
Frames gene-test selection as an episodic MDP where a PPO actor-critic policy adds genes one at a time, rewarded by the unreduced loss of a frozen transformer AR predictor that uses existence embeddings (gene presence/absence tokens) and metadata tokens, with sum-based state encoding and prevalence-based initialization of the policy's output layer.
Key results
- GenoARM with metadata achieves 91.36% average accuracy over five pathogens versus 91.01% for RandEvolve-M and 89.55% for OptStat-M (5 gene tests).
- Without metadata, all methods perform similarly, with OptStat best at 89.90% average accuracy versus 89.62% for RandEvolve and 89.44% for GenoAR.
- On E.coli with 15 optimized genes, performance nearly matches a model trained on all 198 genes.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use GenoARM instead of the strongest competing method | E. coli, K. pneumoniae, and Acinetobacter prediction with 5 gene tests and metadata · E. coli, K. pneumoniae, Acinetobacter | accuracy: 93.44 -> 93.77; 85.70 -> 86.80; 90.15 -> 90.69 | Helped |
| use GenoARM instead of the strongest competing method | Salmonella prediction with 5 gene tests and metadata · Salmonella | accuracy: 94.54 -> 94.50 | No clear effect |
| use GenoARM instead of the strongest competing method | S. aureus prediction with 5 gene tests and metadata · S. aureus | accuracy: 92.19 -> 91.07 | Hurt |
| add metadata to GenoAR (ablation) | E. coli prediction · E. coli | accuracy: 89.57 -> 94.04 | Helped |
| use Year metadata only (ablation) | GenoARM · E. coli | accuracy: 89.57 -> 91.94 | Helped |
| use Country metadata only (ablation) | GenoARM · E. coli | accuracy: 89.57 -> 91.25 | Helped |
| use Source metadata only (ablation) | GenoARM · E. coli | accuracy: 89.57 -> 92.33 | Helped |
| use all metadata types rather than no metadata (ablation) | GenoARM · E. coli | accuracy: 89.57 -> 94.04 | Helped |
| increase the gene-test subset size (ablation) | OptStat · E. coli | accuracy | Hurt |
| increase the gene-test subset size (ablation) | GenoAR and RandEvolve · E. coli | accuracy | Mixed |
Surprising findings
- Metadata helps function-approximator methods substantially but gives OptStat-M smaller gains or even decreases in performance.
- Average gap between GenoARM and RandEvolve-M is small (about 0.35 accuracy points).
Limitations
- Relies on known AR genes, requiring retraining to incorporate new genes.
- Data predominantly from the United States, limiting generalizability.
- Differences between GenoARM and RandEvolve-M are relatively small.
Experimental setup
Datasets: NCBI Pathogen Detection
Benchmarks: NCBI Pathogen Detection (E.coli, K.pneumoniae, Salmonella, S.aureus, Acinetobacter) five-fold cross-validation
Metrics: accuracy major error very major error
Baselines: RandEvolve OptStat GenoAR
Builds on: proximal policy optimization transformer encoder
Closest papers by meaning
- CL-MFAP: A Contrastive Learning-Based Multimodal Foundation Model for Molecular Property Prediction and Antibiotic Screening (ICLR 2025 · similarity 0.81)
- Efficient Biological Data Acquisition through Inference Set Design (ICLR 2025 · similarity 0.81)
- Context-Aware Regularization with Markovian Integration for Attention-Based Nucleotide Analysis (NeurIPS 2025 · similarity 0.80)
- ViroGym: Realistic Large-Scale Benchmarks for Evaluating Viral Proteins (NeurIPS 2026 · similarity 0.80)
- Steering Generative Models with Experimental Data for Protein Fitness Optimization (NeurIPS 2025 · similarity 0.80)
- Active Learning for Efficient Discovery of Optimal Combinatorial Perturbations (ICML 2025 · similarity 0.80)
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.