In the ever-evolving landscape of data science, embeddings have become a game-changer, particularly for image analysis. At Spring Science, we leverage this powerful tool within our Spring Engine platform to enhance image analysis and decision making. This blog post aims to explain what embeddings are, how they are used in image analysis, how they perform compared to traditional approaches, and why we have integrated them into our platform to deliver superior user experience and scientific outcomes.

What Are Embeddings?

Embeddings are a way to represent complex data in a lower-dimensional space while preserving critical information. In practical terms, embeddings are the product of transforming high-dimensional data like images into vectors of numbers, through the use of specialized deep learning models. Embeddings can compactly represent the information within an image in a small vector of numbers. These vectors capture the essential features and relationships within the data in an unbiased manner, enabling more efficient and effective analysis.

The Role of Embeddings in Image Analysis

In image analysis, deep learning models  are used to convert images into a numerical format that machines can easily interpret and analyze. Historically, image analysis was slow as images were essentially analyzed on a pixel-by-pixel basis, whereas converting images into embeddings allows the visual data within an image to be turned into a vector of numbers which are much easier for computers to process. Here's a step-by-step look at how this works:

1. Image Input: An image is processed through a neural network model.

2. Feature Extraction: The model identifies and extracts key features from the image, such as edges, textures, and shapes, as well as other higher-level patterns that may be indiscernible to humans.

3. Vector Representation: The image is then represented as a vector, capturing its essential features.

4. Analyze: These vectors can be used to analyze the contents of the images and compare them to one another. This enables scientists to measure similarity between images, classify images  into specific categories or cluster them in an unbiased manner. 

Embeddings in the Spring Engine

Phenotypic profiling involves studying the observable characteristics (phenotypes) of cells to understand underlying biological processes, and how cells are being impacted by various experimental conditions. In this context, image analysis is crucial, especially with the large volumes of data generated in high content screening (HCS). Here's how embeddings are integral to our Spring Engine platform:

• Automated Feature Extraction: Embeddings remove bias from and automate the feature extraction process, ensuring consistency and accuracy. Using embeddings also eliminates the need for manual intervention and allows scientists to have an a priori approach to data analysis.

• Handling High-Dimensional Data: Biological images are complex and high-dimensional. Embeddings simplify this complexity into lower-dimensional representations without losing vital information.

• Precise Image Similarity: Converting images into vectors allows for accurate measurement of similarities between phenotypes, essential for identifying patterns and anomalies.

• Machine Learning Integration: Embeddings enable seamless integration of image data with machine learning models, facilitating advanced analysis like predictive modeling and clustering.

Benchmarking the Performance of Our Embeddings

At Spring Science, we take pride in the superior performance of our AI embeddings. To ensure our embeddings are at the forefront of image analysis technology, we are constantly assessing their performance for prediction tasks. In one benchmarking experiment, we evaluated their ability to accurately predict the mechanism of action (MOA) for a given compound. Utilizing the Broad Institute’s BBBC021 dataset, we compared the prediction accuracy of our AI embeddings to the prediction accuracy of CellProfiler-derived traditional features. This dataset is widely recognized for its utility in evaluating image-based phenotypic profiling methods.

Our benchmarking results are impressive:

• Spring's AI Embeddings: Achieved a 92.8% accuracy in predicting the MOAs from the dataset.

• Traditional CellProfiler Features: Managed a 72.8% accuracy in predicting the same MOAs

These results highlight the significant advantage of using our embeddings over traditional image analysis and feature extraction methods. By capturing more relevant and detailed features from the images, our embeddings provide a more accurate and reliable analysis, making them an invaluable tool in phenotypic profiling.