Using machine learning and/or neural networks to validate stem cells and their derivatives (2-D cells and 3-D tissues) for use in cell therapy and tissue engineered products
Inventors
Bharti, Kapil • Hotaling, Nathan A. • SCHAUB, Nicholas J. • SIMON, Carl G.
Assignees
US Department of Health and Human Services
Publication Number
US-11531844-B2
Publication Date
2022-12-20
Expiration Date
2039-03-15
Interested in licensing this patent?
MTEC can help explore whether this patent might be available for licensing for your application.
Abstract
A method is provided for non-invasively predicting characteristics of one or more cells and cell derivatives. The method includes training a machine learning model using at least one of a plurality of training cell images representing a plurality of cells and data identifying characteristics for the plurality of cells. The method further includes receiving at least one test cell image representing at least one test cell being evaluated, the at least one test cell image being acquired non-invasively and based on absorbance as an absolute measure of light, and providing the at least one test cell image to the trained machine learning model. Using machine learning based on the trained machine learning model, characteristics of the at least one test cell are predicted. The method further includes generating, by the trained machine learning model, release criteria for clinical preparations of cells based on the predicted characteristics of the at least one test cell.
Core Innovation
The disclosed invention provides a method for non-invasively predicting characteristics of one or more cells and their derivatives by using machine learning models trained on imaging data. The method includes training a machine learning model using a plurality of training cell images representing multiple cells along with data identifying their characteristics, receiving test cell images acquired non-invasively using quantitative bright-field absorbance microscopy (QBAM), and then predicting cellular characteristics by providing these test images to the trained machine learning model. Furthermore, the method includes generating release criteria for clinical preparations of cells based on the predicted characteristics.
The invention addresses significant limitations present in current cell health and function assays used for cell therapies, which are often non-quantitative, highly variable, destructive, expensive or require disturbing the cells, thereby impeding longitudinal non-invasive assessment. Current methods frequently increase contamination risk and preclude further use of cells, lack compatibility with high throughput screening, and do not provide comprehensive information on cell health and function. Despite the ability of biologists to visually assess certain cell types non-invasively under bright-field microscopy, this manual analysis suffers from sampling bias, challenges in relating visual data to function, and difficulty in identifying causation.
Accordingly, the invention introduces an automated, high-throughput, non-invasive machine learning-based analysis framework that overcomes the limitations of conventional assays. It combines QBAM for quantitative, reproducible imaging of cells based on absorbance, which accounts for absolute light attenuation, with advanced machine learning and deep neural network models to analyze cellular images. This integrated approach enables prediction of various cell characteristics including identity, function, maturity, disease state, drug effects, and donor origin without disturbing the cells, thereby supporting cell therapy product validation, drug discovery, and toxicity testing efficiently and accurately.
Claims Coverage
The patent includes two independent claims, one directed to a method for non-invasive prediction of cell characteristics using machine learning and QBAM images, and the other directed to a computing system configured to perform the method.
Non-invasive cellular characteristic prediction using machine learning and QBAM images
The method involves training a machine learning model using training cell images and associated characteristics, receiving test cell images acquired by quantitative bright-field absorbance microscopy, providing test images to the trained model, predicting cell characteristics using machine learning, and generating release criteria for clinical cell preparations based on these predictions.
Use of deep neural networks for image segmentation and classification
The method includes employing a deep neural network to segment images into individual cells, extract visual features such as cell boundaries, shapes, and texture metrics (including sub-cellular features), and classify cells based on predicted characteristics like identity, function, drug effects, disease state, or similarity to replicates.
Conversion and processing of raw microscopy images to absorbance data
The method comprises receiving raw microscopy images, converting pixel intensities to absorbance values as an absolute measure of light attenuation, calculating absorbance confidence, establishing microscope equilibrium through benchmarking, and applying color filters to improve image quality and reproducibility.
Handling of large images through tile processing
The method includes dividing large test images into multiple tiles, processing each tile individually through the trained machine learning model, combining the outputs, and providing a unified prediction corresponding to the entire large image.
Computing system configured for the method
A computing system with a processor and memory configured to train the machine learning model with training cell images and characteristics, receive QBAM test cell images, utilize the trained model to predict characteristics, and generate release criteria for clinical preparations based on predictions.
The independent claims cover a comprehensive, non-invasive method and computing system using machine learning models trained on QBAM images for predicting cell characteristics, segmenting and classifying cell images, converting microscopy data to absorbance for robust analysis, handling large images via tile processing, and generating clinical release criteria based on predictions.
Stated Advantages
Eliminates the requirement for a trained user and reduces human bias and error.
Enables non-invasive, high throughput and automated analysis of cells and tissues.
Reduces cost and time needed to make clinically relevant cell functional predictions.
Provides reproducible, statistically robust imaging that correlates to physiological and biochemical cell functions.
Allows longitudinal assessment of cells without disturbing or destroying them.
Supports automated identification of cell identity, function, donor origin, disease state, and effects of drugs or toxins.
Improves scalability and applicability for large-scale biomanufacturing and drug discovery.
Documented Applications
Validation of stem cells and their derivatives for use in cell therapy.
Drug discovery including determining drug efficacy and toxicity on cells.
Drug toxicity testing and screening for toxic effects of compounds.
Non-invasive release criteria generation for clinical preparations of cells.
High throughput screening in biomanufacturing processes.
Longitudinal monitoring of cell maturation, health, and function over time.
Classification and validation of cell donor identity and detection of developmental outliers.
Interested in licensing this patent?