Research map / Label-efficient and robust learning / Recognition and long-tail

Chem-PerturBridge: a harmonized compendium of small molecule perturbation transcriptomic effects

NeurIPS 2026 · Label-efficient and robust learning › Recognition and long-tail · approach: Dense features and foundation priors · 1 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.

Introduces Chem-PerturBridge, a harmonized compendium of ~1.25M small-molecule transcriptomic samples across eight assay types, and uses it to benchmark cross-dataset logFC agreement and to pretrain LPM compound embeddings evaluated on held-out OP3 compounds and molecule-holdout splits.

Problem

Small-molecule transcriptomic perturbation data are fragmented across assays, metadata conventions, controls, doses, and preprocessing, so it is unclear whether matched same-compound signatures agree across datasets and whether pooled data helps perturbation models.

Approach

A unified AnnData schema plus a shared limma-based differential-effect layer across count and normalized assays, with matched-condition agreement scores calibrated against same-context different-compound baselines, and pretraining LPM compound embeddings on the pooled resource.

Key results

• On held-out OP3 compounds, LPM embeddings pretrained on Chem-PerturBridge reach Spearman 0.2745 vs 0.2638 for Morgan fingerprints and 0.2533 for descriptor-free (10 seeds).
• Cross-dataset DEG-restricted logFC Spearman agreement is weak and often falls below same-context different-compound baselines for most dataset pairs.
• LogFC direction agreement is more stable and usually exceeds same-context different-compound baselines.
• Across 11 molecule-holdout datasets, a Chem-PerturBridge-trained model gives the lowest or tied-lowest mean test RMSE in 10 of 11 datasets.

Surprising findings

• Within-dataset DEG-restricted logFC Spearman replicate agreement can fall below the same-context different-compound baseline, even for technical replicates.
• At the dataset-pair summary level, the same-context different-compound baseline ranks above the true same-compound partner in retrieval for every evaluated dataset pair.

Limitations

• Cross-dataset agreement only covers compounds, contexts, doses and genes that match after harmonization, a non-random subset.
• OP3 evaluation covers four immune cell contexts and a narrow compound panel.
• No single training strategy dominates across datasets; target-only training and Morgan fingerprints remain competitive.

Experimental setup

Datasets: L1000 Phase I L1000 Phase II Tahoe-100M CIGS-MCE Novartis/DRUG-seq U2OS VCPI-0001 CIGS-TCM VCPI-0002 GDPx2 sci-Plex DILImap OP3

Benchmarks: OP3 compound-held-out split molecule-holdout across 11 datasets

Metrics: Spearman cosine MRRMSE MAE RMSE logFC Spearman direction agreement retrieval normalized rank

Baselines: Descriptor-free OP3 baseline Morgan fingerprints LPM on L1000

Builds on: LPM limma edgeR scPerturb LINCS L1000

Closest papers by meaning

• A benchmark for prediction of transcriptomic responses to chemical perturbations across cell types (NeurIPS 2024 · similarity 0.91)
• Contextualizing biological perturbation experiments through language (ICLR 2025 · similarity 0.88)
• PerturBench: Benchmarking Machine Learning Models for Cellular Perturbation Analysis (NeurIPS 2025 · similarity 0.88)
• PerturbReason: A Knowledge-Grounded Benchmark and Framework for Cell-State–Conditioned Mechanistic Reasoning of Perturbation Effects (NeurIPS 2026 · similarity 0.86)
• Learning Cross-Domain Representations for Transferable Drug Perturbations on Single-Cell Transcriptional Responses (AAAI 2025 · similarity 0.85)
• PertEval-scFM: Benchmarking Single-Cell Foundation Models for Perturbation Effect Prediction (ICML 2025 · similarity 0.85)