Self-supervision advances morphological profiling by unlocking powerful image representations

Abstract

Morphological profiling is a powerful technology that enables unbiased characterization of cellular states through image-based screening. Inspired by recent progress in self-supervised learning (SSL), we sought to explore the potential benefits of using SSL in this domain and conducted a comprehensive benchmark study of recent SSL methods for learning representations from Cell Painting images without segmentation. We trained DINO, MAE, and SimCLR on subsets of the JUMP-CP consortium data, one of the largest publicly available Cell Painting image sets, and observed improved model performance with larger and more heterogeneous training sets. Our best model (DINO) surpassed the widely used profiling tool CellProfiler by 29% in mean average precision (mAP) on classifying chemical perturbations and significantly accelerated feature extraction by 50x, at a lower cost. Moreover, DINO outperformed CellProfiler in clustering gene families on an independent gene overexpression dataset. Our findings indicate that SSL methods can improve the efficiency and performance of morphological profiling, offering the potential to expedite drug discovery and reduce compute costs.

Evaluating batch correction methods for image-based cell profiling

“…To make interpretation easier, every metric here was normalized between 0 and 1, with 0 being the worst performance and 1 being the best. These quantitative metrics (Sup Figure A) guided the selection of the preprocessing steps to be used as our Baseline approach, in agreement with established protocols in the field 34,36 . All remaining results in this paper apply this four-step preprocessing.…”

“…Methods evaluated in this benchmark process tabular data extracted with standard image processing algorithms 43 . An alternative approach is to learn representations from images 29,36 . The design of neural network architectures and representation learning algorithms for Cell Painting is still an active research area, as is studying the interaction of these methods with those for batch correction; thus, our benchmark serves to establish a baseline for future studies comparing the performance of batch correction methods for learning-based representations.…”

“…However, this approach also requires the tuning of additional classification pipelines during evaluation, making the comparison of batch correction methods somewhat less straightforward. Lastly, Kim et al's 44 evaluation provides another direct approach, using silhouette score, graph connectivity, and local inverse Simpson's index (LISI) to quantify batch effects across different representation learning methods. Their evaluation is similar to our Scenarios 1 and 2, described below.…”

“…Again, we observed that applying both preprocessing steps was preferable according to most metrics. We therefore selected MAD + Sphering as our Baseline approach, in agreement with established protocols in the field 44,45 . All remaining results in this paper apply this two-step preprocessing.…”

Cell-Vision Fusion: A Swin Transformer-based Approach to Predicting Kinase Inhibitor Mechanism of Action from Cell Painting Data

“…Kim et al 70 performed a systematic analysis of common image augmentations for multi-channel microscopy images, finding that a combination of random brightness shifts and intensity changes applied independently to each image channel had the greatest positive impact on model performance. We have extended this approach by adding a probability weighting for each channel, hereby referring to this method as Channel-wise Augmentation (CWA).…”