We’re proud to announce CellusAI, our artificial intelligence powered segmentation, phenotyping, and prediction of molecular bioactivity and toxicity from cell painting images.
CellusAI is our machine learning platform for analysis of cell painting images, providing a complete solution for image segmentation, profile extraction and downstream processing, including predictions of bioactivity, mechanism of action, and toxicity. CellusAI provides increased performance in comparison to legacy analytical pipelines, such as CellProfiler:
- CellusAI mechanism of action prediction features perform almost twice as good as CellProfiler features.
- Segmentation is fully automated and does not require any user input.
- CellusAI features are more informative than legacy tools features.
- The analyses are faster due to support for GPU acceleration.
Figure 1: Comparison of CellusAI platform and CellProfiler. The AI modules of the CellusAI platform are shown in the purple section at the top. The Bioactivity prediction module is already implemented as prototypes. Other modules are being developed or are on the roadmap. The bottom section shows comparable alternatives in CellProfiler (if available).
Our Bioactivity prediction module for can predict biochemical and cell-based assay measurements from CellusAI features. Currently, we can predict 32 different assays from either CellusAI features or chemical structures. This is being further expanded with better models and more assays.
The Toxicity prediction and Mechanism of Action prediction modules are currently under development. These modules will predict Toxicity assay measurements and Mechanism of Action of molecules either from CellusAI features or chemical structures.
The New Molecule generative AI module is on our roadmap. This module will generate chemical structures with the help of AI, which would induce a similar cell painting profile. Such an approach would generate new, not yet tested or patented molecules, and is a great tool for generating new assets. The module will also assess different molecular properties, such as whether molecules can be synthesized, how they pass the blood-brain-barrier and estimate the toxicity.
Performance estimation of CellusAI features
The CellusAI feature extraction module is based on a weakly supervised learning model. The model is trained to predict the molecular structure of the profile and then the embeddings are used as features. To estimate the performance of this model we followed the approach described in Moschkov 2022, where the embeddings are used in a non-related downstream task. We also used a similar mechanism of action matching. This is currently the best described benchmarking methodology and therefore was the best choice for estimating the performance of our own models.
We used the benchmarking dataset BBBC022 from Broad Institute, and performed the segmentation and feature extraction with CelllusAI. CellProfiler features were downloaded from the Broad Institute repository. Both sets of features were processed as described in Moschkov 2022.
Figure 2: Performance metrics of CellusAI. Horizontal axis shows mean average precision and vertical axis shows folds of enrichment in matching the mechanism of action based on CellusAI or CellProfiler features. Blue dashed lines show the CellProfiler baselines.
In figure 2 we show two measurements: folds of enrichment and mean average precision (higher is better for both, see Moschkov 2022 for detailed description). The fold of enrichment for Cellus AI was 48.2, while for CellProfiler it was 23.7 (Figure 2). Similarly, CellusAI mean average precision was 0.090, compared to 0.068 in CellProfiler. Both measurements indicate that CellusAI features are more informative and can be used to make better predictions in downstream analyses.
Roadmap
The vision for CellusAI is to provide a comprehensive set of AI tools for cell painting, which will speed up and reduce the cost of drug discovery campaigns. The platform will reduce the amount of laboratory tests needed due to better predictions, enabling researchers to shortlist the molecules with a greater potential to have the desired properties for a drug candidate.
Cell painting images are rich with information and we aim to leverage them as much as possible to predict the properties of the molecules. Currently, we have a working prototype of segmentation and feature extraction modules, as well as an early prototype of the assay prediction module. The underlying models were trained on publicly available data. However, the real value is in many non-public datasets that our customers hold and will be used to improve the models by orders of magnitude.
CellusAI will therefore be used in multiple different ways: as an AI tool for CROs providing HCS services, in close collaboration with pharma and biotech companies who would like to boost their drug discovery, or as an internal tool for our own drug discovery projects.
Currently our development focuses on improving existing models and developing new predictors. The initial version of the platform will be released early next year.
If you want to learn more about how CellusAI can upgrade your cell painting analysis process and help you achieve better and faster results, contact our team. We’ll be happy to answer any questions you might have.