CHAMMI-75: Pre-training multi-channel models with heterogeneous microscopy images

AI summary

Overview

Research area: Computer vision applied to computational biology — self-supervised representation learning for multi-channel microscopy images and image-based cell phenotyping.

Technical level: Intermediate. The paper is a dataset, benchmark, and empirical evaluation study; it assumes familiarity with vision transformers, self-supervised learning (SSL) objectives such as DINO, MAE, and SimCLR, and standard benchmarking metrics, but the biological motivation is explained plainly.

Scope: The paper introduces CHAMMI-75, a large heterogeneous multi-channel microscopy dataset, three new evaluation benchmarks, and MorphEm, a channel-adaptive pre-trained model, and evaluates how dataset diversity affects multi-channel model performance.

What This Paper Is About

Cellular morphology models are usually trained on a single microscopy imaging type with a fixed number of channels, so a model trained in one study cannot be reused in another where the microscope, channels, or resolution differ. The authors argue that the field lacks a large, well-curated, multi-channel dataset to test channel-adaptive models at scale. They build one by combining images from many public biological studies, then use it to pre-train and benchmark models that can accept any microscopy image type.

Key Contributions

1. A large heterogeneous multi-channel microscopy dataset. CHAMMI-75 contains 2,792,462 fields-of-view (FoV) for pre-training sampled from 74 publicly available microscopy studies, plus 2,474,875 FoV held out for testing, with 25 channel types and variation across 16 organisms, 223 cell lines, microscopes, resolutions, and channel counts. Each pre-training FoV has up to seven channels. A total of 76 unique sources is reported (four of the six benchmarks share images with the pre-training set).

2. Three new evaluation benchmarks plus adoption of existing ones. The authors adopt the CHAMMI benchmark, HPAv23 at 256x256, and JUMP-CP1 Compounds, and introduce IDR-0017, CellPHIE (14-channel images), and RBC-MC, which test channel generalization and cross-domain generalization.

3. A systematic evaluation of CHAMMI-75 as a pre-training resource. Six data ablation studies, dataset/model/compute scaling experiments, a weakly supervised learning comparison, and a batch-correction and disentanglement analysis using the JUMP-CP batch correction benchmark.

4. A top-performing released model, MorphEm. A bag-of-channels (BoC) ViT-small model trained with DINO on the full CHAMMI-75 dataset, which the authors describe as achieving the best result among SSL methods in six of seven reported metrics. Data, code, and model weights are released publicly.

Main Findings

• MorphEm leads among SSL methods. The BoC ViT-small model trained on CHAMMI-75 scored 48.75 (CHAMMI), 58.87 (HPAv23 256x256), 76.32 (JUMP-CP1), 06.79 (JUMP-CP2), 75.52 (IDR0017), 80.51 (CellPHIE), and 68.34 (RBC-MC), the best of the SSL models in six of seven metrics.

• A larger weakly supervised suite remains competitive. SubCell, a suite of ViT-base models, scored 53.38, 69.33, 77.60, 07.44, 75.37, 71.23, and 59.10 respectively, and the authors attribute its strength to factors including training with biological objectives, larger models, channel specialization, and manual selection of the best result across its configurations.

• CHAMMI-75 beats a comparable multi-channel dataset. A model trained on IDRCell100k (79 sources vs. 75, but 100k multi-channel images) underperformed in most tasks, which the authors interpret as CHAMMI-75 containing additional informative and higher-quality images.

• Strong generalization to novel channel combinations and modalities. MorphEm outperformed the larger, weakly supervised SubCell by 13% on CellPHIE (novel 14-channel combination) and 15% on RBC-MC (novel modality and cross-domain clinical sites).

• Heterogeneity is the main driver, not any single biological attribute. Models trained only on target data underperformed by a 27% relative difference, while training with CHAMMI-75 improved performance by up to 38%. Restricting to the two dominant modalities (fluorescence and epi-fluorescence) underperformed by 13% relative, whereas the other 12 less-represented modalities improved performance by 15% relative.

• Magnification matters modestly; cell line and channel count barely matter. Microscope magnification showed a relative performance difference of about 3%, while excluding the most common cell lines (U2OS and A549), or limiting training to studies with up to two, three, or four channels, changed performance by about 1%.

• Bag of channels beats multi-channel attention under SSL. The BoC approach yielded up to 19% relative improvement over multi-channel attention at a fraction of the computational cost; MCA required 3X to 5X more GPU hours due to longer sequences.

• DINO was the strongest SSL algorithm. About 15% relative improvement over MAE and about 7% over SimCLR. Scaling from ViT-small to ViT-large gave a 10% relative improvement. Adding weak supervision, even with noisy labels, improved the MCA-SSL setting by 1-19% relative.

• Scaling behavior. As data and model size increase, SSL models approach fully supervised, specialized models. MorphEm was trained on the full 2.8M-image set (5X larger) using 2,352 GPU hours, yielding a 9.8% relative improvement over the best dataset-scaling result.

• Better technical/biological separation. In the JUMP-CP batch correction benchmark, CHAMMI-75 features required less batch correction and maximized biological signal more than features extracted with CellProfiler or learned from IDRCell100K. After correction with the Seurat CCA algorithm, features separated biological clusters from technical batches well. UMAP visualizations show source study (technical domain) is the primary clustering factor.

• Weakly supervised models also benefit. Trained with 74ds-small using the reagent identifier as a target, ChA-MAEViT scored 38.72, 56.92, 74.99, 05.53, 24.99, and 76.12 across the six reported benchmarks, ahead of MCA-SupC and MCA-SSL.

Methodology in Plain English

The authors assembled the dataset in three phases. In data acquisition, they selected 75 source datasets from 18 hosting platforms that publish biological images, including the Image Data Resource, Zenodo, Mendeley Data, and Figshare; 97% of the datasets carry a Creative Commons license (60% CC BY 4.0, 12% CC0 1.0). In metadata integration, they built a table with 22 columns in 6 groups, parsing values from resource descriptor files, image filenames, and source publications, with large language models assisting identification and organization of some information, followed by deterministic programmatic parsing and manual curation. In data curation, they reduced roughly 26M downloaded images to 2.8M by sampling 2D slices from 3D stacks, frames from live microscopy videos, and wells from control conditions, then applied K-means-based filtering to select diverse, high-quality images. They ran Cellpose segmentation to record center-of-mass coordinates for 1.8B single cells so that training crops would contain visible cells.

For evaluation they compared channel-adaptive architectures — BoC models that read one channel at a time and concatenate features (Microsnoop, uniDINO, DINO-BoC), and multi-channel attention models that treat channels as tokens in a sequence (Channel-ViT, CA-MAE, ChA-MAEViT) — training ViT-small models with SSL and extracting features from frozen weights without fine-tuning.

Why This Matters

Impact on research. The paper argues that progress in representation learning has often followed large, rigorous curation efforts such as ImageNet and LAION, and that computational biology has lacked an equivalent heterogeneous resource for multi-channel microscopy. CHAMMI-75 provides a common testbed so that channel-adaptive models can be compared at scale rather than in small proof-of-concept experiments.

Real-world applications (as described or implied by the benchmarks):

• Profiling how cells respond to chemical compounds, including ranking gene-compound combinations likely to have a large effect (IDR-0017).
• Mapping protein localization in human cells for 19 or 31 localization classes (HPAv23).
• Classifying genetic perturbations, as in the Huntington's Disease pooled screen benchmark CellPHIE.
• Assessing red blood cell morphology for blood quality across clinical sites (RBC-MC).

Industry relevance. The motivating use case is cell phenotyping — quantifying morphological differences between healthy and perturbed cells — which underlies drug discovery and high-throughput screening. The authors note the work is intended to further scientific research and understanding of foundation models in microscopy, not commercial purposes, and the work was supported by an unrestricted grant from Meta AI and the AWS Open Data Sponsorship Program.

Future Directions

• Applying the resource to new multi-channel modeling methods. The authors state the work focused on high-quality data curation and explicitly leave investigation of novel multi-channel modeling methods to future research.

• Scaling the model further. The authors note "potential for further gains using a larger ViT in future work," since ViT-small to ViT-large already yielded a 10% relative improvement.

• Improving metadata quality and balance. The collected metadata is described as informative but noisy despite standardization efforts, and the dataset is sparse and does not cover all relevant variables in a balanced way — both a challenge and an opportunity for supervised and similar study types.

• Reducing dependence on batch correction. Since the primary clustering factor in the feature space is the source study, future work would need models that are intrinsically more robust to technical variation rather than relying on post-hoc correction with methods such as Seurat CCA.

Target Audience

Researchers in computer vision and machine learning working on multi-channel or channel-adaptive architectures; computational biologists and bioimage analysts building cell phenotyping models; and groups developing foundation models for microscopy imaging. It is also relevant to practitioners who need models that transfer across labs, microscopes, and channel configurations, and to dataset builders interested in large-scale biological data curation, metadata integration, and quality control.

Authors’ abstract

Quantifying cell morphology using images and machine learning has proven to be a powerful tool to study the response of cells to treatments. However, models used to quantify cellular morphology are typically trained with a single microscopy imaging type. This results in specialized models that cannot be reused across biological studies because the technical specifications do not match (e.g., different number of channels). Here, we present CHAMMI-75, an open access dataset of heterogeneous, multi-channel microscopy images from 75 diverse biological studies. We curated this resource from publicly available sources to investigate cellular morphology models that are channel-adaptive and can process any microscopy image type. Our experiments show that training with CHAMMI-75 can improve performance in multi-channel bioimaging tasks primarily because of its high diversity in microscopy modalities. This work paves the way to create the next generation of cellular morphology models for biological studies.