CellPaintSSL_IN_MOA: Preprocessed Cell Painting MOA plate (images and dataset.csv) for CellPaintSSL zero-shot inference

Summary

CellPaintSSL_IN_MOA is a preprocessed Cell Painting dataset built from a single MOA plate and prepared specifically for running the CellPaintSSL model in a strictly zero‐shot regime (no fine‐tuning). The dataset contains five‐channel Cell Painting images transformed into the input format required by the official CellPaintSSL implementation, together with a manifest file (`dataset.csv`) that links each image to well, site, and perturbation metadata.

This dataset is derived from the same experimental plate used in our MOAProfiler_Feature_Extractor workflow, but it uses a different preprocessing and manifest format tailored to CellPaintSSL rather than MOAProfiler.

Contents

`images/` – preprocessed five‐channel Cell Painting images in the format expected by CellPaintSSL.

`dataset.csv` – manifest linking each image to `plate`, `well`, `site`, `perturbation_id`, etc.

All metadata and file naming conventions are compatible with the official CellPaintSSL codebase from Bayer (https://github.com/Bayer-Group/CellPaintSSL).

How the data were generated

Raw Cell Painting images from the MOA plate were stored as separate 16-bit TIFF files per channel with filenames of the form BaseName-chX*.tiff. For each field of view (identified by BaseName), we collected the required five channels and built a 5×H×W uint16 stack in the channel order expected by our CellPaintSSL pipeline, corresponding to (mito, agp, rna, er, dna). For each channel, intensities were clipped to the 0.1–99.9th percentiles and then linearly rescaled to the full 16-bit range (0–65535). The five rescaled channels were stacked along a new axis and saved as a single multi-channel TIFF (BaseName.tiff). In addition, a global Otsu threshold was computed on the DNA channel (normalized to [0, 1]) and stored in the TIFF metadata as the field otsuth. The resulting CellPaintSSL_IN_MOA dataset therefore consists of 5-channel, per-field TIFF stacks (mito, agp, rna, er, dna) with standardized 16-bit intensity scaling and an accompanying per-image DNA Otsu threshold.

概述
CellPaintSSL_IN_MOA是一款从单个MOA板预处理得到的细胞绘画（Cell Painting）数据集，专为严格零样本（zero-shot）设置下运行CellPaintSSL模型打造（无需微调）。本数据集包含转换为官方CellPaintSSL实现所需输入格式的五通道Cell Painting图像，以及一份关联每张图像与孔位、视野位点、扰动元数据的清单文件（`dataset.csv`）。

本数据集源自我们MOAProfiler_Feature_Extractor工作流所使用的同一实验板，但采用专为CellPaintSSL定制的预处理与清单格式，而非适配MOAProfiler的格式。

数据集内容
`images/`：符合CellPaintSSL输入格式的预处理五通道Cell Painting图像。
`dataset.csv`：关联每张图像与`板号（plate）`、`孔位（well）`、`视野位点（site）`、`扰动ID（perturbation_id）`等信息的清单文件。

所有元数据与文件命名规范均兼容拜耳（Bayer）官方发布的CellPaintSSL代码库（https://github.com/Bayer-Group/CellPaintSSL）。

数据生成方式
MOA板的原始Cell Painting图像以单通道16位TIFF文件形式存储，文件名格式为`BaseName-chX*.tiff`。针对每个以`BaseName`标识的视野位点，我们收集所需的五个通道，并按照CellPaintSSL流水线要求的通道顺序构建5×H×W的uint16格式图像堆叠数组，对应通道依次为线粒体（mito）、质膜/肌动蛋白（agp）、RNA（rna）、内质网（er）、DNA（dna）。对每个通道的像素强度进行0.1至99.9百分位截断，随后线性缩放至完整16位范围（0至65535）。将五个经过缩放的通道沿新轴堆叠，保存为单张多通道TIFF文件（`BaseName.tiff`）。此外，在归一化至[0,1]的DNA通道上计算全局大津（Otsu）阈值，并以`otsuth`字段名存入TIFF元数据中。最终生成的CellPaintSSL_IN_MOA数据集包含五通道、单视野TIFF堆叠文件（对应线粒体（mito）、质膜/肌动蛋白（agp）、RNA（rna）、内质网（er）、DNA（dna）），采用标准化16位强度缩放方案，并附带每张图像对应的DNA通道全局大津阈值。