Research map / Probabilistic and scientific modeling / Uncertainty and Gaussian processes

PRESCRIBE: Predicting Single-Cell Responses with Bayesian Estimation

NeurIPS 2025 · Probabilistic and scientific modeling › Uncertainty and Gaussian processes · approach: Bayesian decision-making under uncertainty · 0 citations (OpenAlex)

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 PRESCRIBE, a multivariate Natural Posterior Network with a Normal-Inverse-Wishart posterior that predicts single-cell perturbation responses and a pseudo E-distance confidence score for filtering unreliable unseen-gene predictions.

Problem

Single-cell perturbation prediction for unseen genes lacks reliable per-prediction uncertainty quantification, which must capture both epistemic (similarity to training perturbations) and aleatoric (outcome variability) uncertainty.

Approach

Extends Natural Posterior Network to multivariate gene expression using a Normal-Inverse-Wishart conjugate prior, combines posterior evidence (flow-estimated latent density) with predictive entropy into a pseudo E-distance confidence score, and trains with an E-distance ranking loss (ListMLE) plus auxiliary regularizers.

Key results

• On Norman, PRESCRIBE-10% reaches r_pred,truth 64.32% vs 61.48% for scGPT and 45.30% for GEARS, with filtering of the 10% least confident predictions.
• PRESCRIBE achieves the highest Spearman correlation between confidence and accuracy across all three datasets, per the paper's calibration analysis.
• Random filtering of 10% of predictions returns accuracy to roughly the unfiltered level, while the model's confidence-based filtering yields consistent gains.

Measured effects reported in the paper

Surprising findings

• Removing the E-distance ranking loss (w/o L3) slightly increases directional accuracy (63.61 vs 63.24) while substantially reducing uncertainty calibration.
• Models that predict distributions are inherently harder to train, so PRESCRIBE without filtering did not stand out in raw accuracy against scGPT.

Limitations

• Uncertainty scores rely on a hand-designed pseudo E-distance and an auxiliary ranking loss, with ablations showing trade-offs between accuracy and calibration.
• Accuracy gains come from discarding the 5-10% least confident predictions rather than from improving all predictions.
• Hyperparameter tuning is sensitive, with an inverse relationship between calibration and prediction accuracy observed across lambda ratios.

Experimental setup

Datasets: Norman Replogle2022_Rep1 Replogle2022_K562

Benchmarks: Norman Replogle2022_Rep1 Replogle2022_K562

Metrics: Pearson correlation r_pred,truth directional accuracy ACC DEG-restricted Pearson and accuracy Spearman correlation between confidence and performance ECE

Baselines: AverageKnown Linear Linear scGPT CellOracle samsVAE GraphVCI scFoundation scGPT GEARS GEARS-Drop

Builds on: Natural Posterior Network deep evidential regression energy distance GEARS scGPT

Closest papers by meaning

• PerturBench: Benchmarking Machine Learning Models for Cellular Perturbation Analysis (NeurIPS 2025 · similarity 0.87)
• A benchmark for prediction of transcriptomic responses to chemical perturbations across cell types (NeurIPS 2024 · similarity 0.87)
• A Simple and Comprehensive Benchmark for Single-Cell Transcriptomics (AAAI 2025 · similarity 0.87)
• Cradle-VAE: Enhancing Single-Cell Gene Perturbation Modeling with Counterfactual Reasoning-based Artifact Disentanglement (AAAI 2025 · similarity 0.86)
• MapPFN: Learning Causal Perturbation Maps in Context (NeurIPS 2026 · similarity 0.86)
• PertEval-scFM: Benchmarking Single-Cell Foundation Models for Perturbation Effect Prediction (ICML 2025 · similarity 0.86)