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

MorphoHELM: A Comprehensive Benchmark for Evaluating Representations for Microscopy-Based Morphology Assays

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

Introduces MorphoHELM, an open benchmark evaluating morphological feature extractors on Cell Painting across MoA enrichment, gene pathway enrichment, and replicate retrieval under increasing batch-effect stringency.

Problem

Evaluation of representations for Cell Painting morphological profiling is fragmented across custom tasks, datasets, and metrics, and rarely tests robustness to batch effects of increasing severity.

Approach

A unified benchmark that aggregates biological-signal retrieval tasks (MoA, gene pathway, replicate retrieval) and evaluates each under four batch-effect stringency levels (NR, NSB, NSS, NSL), with corrected metrics (modified Haldane-Anscombe odds ratio with geometric mean, mAP and NegCon mAP) and a standardized feature normalization pipeline.

Key results

• No existing model outperforms classic CellProfiler features across all settings; CellProfiler is superior on kNN Recall@1 and mAP in most settings.
• Among deep models, DINOv2 (natural image foundation model) outperforms all microscopy foundation models in almost all settings.
• All models perform below a 0.05 false discovery level on fraction significant for MoA enrichment in the Not Same Source setting, and replicate retrieval across plate layouts drops substantially; models show negligible performance over random guessing for cross-laboratory comparisons.

Surprising findings

• Natural image foundation model DINOv2 beats microscopy-specific foundation models (OpenPhenom, SubCell, CLOOME) in most settings.
• Whitening harms performance in most cases, and PCA/scaling disproportionately benefit CellProfiler features, suggesting prior work underestimated CellProfiler.

Limitations

• Only Cell Painting assays are evaluated; transferability to other modalities such as label-free or live-cell imaging is unverified.
• Standardized normalization pipeline may not reflect the bespoke postprocessing used by some prior methods, potentially disadvantaging representations that rely on it.
• Replicate retrieval on individual samples may need more extensive postprocessing than the consensus-based enrichment tasks.

Experimental setup

Datasets: BBBC036 cpg-MoA cpg-CRISPR cpg-target2 cpg-compound CORUM HuMAP StringDB SIGNOR Reactome Drug Repurposing Hub

Benchmarks: MorphoHELM BBBC036 cpg-MoA cpg-CRISPR cpg-target2 cpg-compound

Metrics: geometric mean odds ratio fraction significant kNN Recall@1 mAP NegCon mAP

Baselines: CellProfiler ImageNet ResNet101 Untrained ResNet101 DINOv2 OpenPhenom SubCell CLOOME

Builds on: CellProfiler DINOv2 CLOOME OpenPhenom SubCell

Closest papers by meaning

• Micro-Bench: A Microscopy Benchmark for Vision-Language Understanding (NeurIPS 2024 · similarity 0.89)
• CellCLIP - Learning Perturbation Effects in Cell Painting via Text-Guided Contrastive Learning (NeurIPS 2025 · similarity 0.88)
• ViTally Consistent: Scaling Biological Representation Learning for Cell Microscopy (ICML 2025 · similarity 0.87)
• MultiOrg: A Multi-rater Organoid-detection Dataset (NeurIPS 2024 · similarity 0.86)
• A Cross Modal Knowledge Distillation & Data Augmentation Recipe for Improving Transcriptomics Representations through Morphological Features (ICML 2025 · similarity 0.86)
• CellFlux: Simulating Cellular Morphology Changes via Flow Matching (ICML 2025 · similarity 0.85)