System and methods for in vitro structural toxicity testing
Inventors
Maddah, Seyyedeh Mahnaz • Loewke, Kevin Edward
Assignees
Interested in licensing this patent?
MTEC can help explore whether this patent might be available for licensing for your application.
Publication Number
US-11808760-B1
Publication Date
2023-11-07
Expiration Date
Abstract
A system and process use artificial intelligence to evaluate the toxicity of drugs on cells. In some embodiments, a convolutional neural network is trained to identify features in cells and thereafter identify when structural changes in cells are signs of damage from exposure to a drug. Some embodiments use a 2-class deep neural network, comparing drug-treated cells to controls, to learn which images may show signs of toxicity as a result of the drug. In some applications, the system may capture images from a time-lapse experiment to determine from the cell cultures how a drug affects a cell type over time.
Core Innovation
The invention determines a response of a cell to a drug by using images and a convolutional neural network. Images of treated cells exposed to the drug at an initial point in time are received and images of control cells not exposed to the drug are received. The treated and control images are forwarded to a convolutional neural network to determine whether the images of the treated cells are different from the images of the control cells.
A metric is calculated based on the convolutional neural network determination to quantify an amount of change caused by the drug on the treated cells. The document further describes automated response determination where the change quantification is derived from model performance and is reported using quantitative structural-change metrics such as Z-factor and/or Tjur’s pseudo R2.
The disclosed system supports AI-based in vitro structural toxicity testing using microscopy images of drug-treated cells and vehicle/control cells. Separate or repeated CNN training is used to compare each drug condition to controls, with inputs optionally including image patches and image channels/stains. The approach is described as a phenotype-learning method to detect dose- and time-dependent structural damage across multiple drug doses and/or multiple time points, and across example drugs including aspirin, tamoxifen, acetylcysteine, and doxorubicin.
Claims Coverage
The document includes two independent claims, a computer program product and a method, that share a common inventive framework: comparing treated-cell images versus control-cell images using a convolutional neural network and calculating a drug-change quantification metric. Across the dependents, the claim set specifies the metric choice, the CNN classification basis, and additional experimental dimensions such as dosage, time points, multiple drug types, and multiple image channel types.
Receiving treated and control cell images at an initial point in time
Receive images of treated cells, wherein the treated cells are exposed to the drug, at an initial point in time; receive images of control cells not exposed to the drug
Forwarding images to a convolutional neural network for difference determination
Forward the images of the treated cells and the images of the control cells to a convolutional neural network; determine by the convolutional neural network whether the images of the treated cells are different from the images of the control cells
Calculating a metric quantifying an amount of change caused by the drug
Calculate a metric based on the determination by the convolutional neural network quantifying an amount of change caused by the drug on the treated cells
Using a z-factor measure as the calculated metric
The computer program product is configured so that the metric it calculates is a z-factor measure
Using Tjur’s pseudo R2 as the calculated metric
The computer program product is configured to calculate the model performance metric Tjur’s pseudo R2
Performing 2-class classification with the convolutional neural network
The computer program product is configured so that the convolutional neural network performs 2-class classification
Quantifying changes per multiple dosages using a series of convolutional neural networks
A computer program product adds code that treats cells with multiple dosages of a drug and determines and quantifies the changes for each dosage using a series of convolutional neural networks
Quantifying changes shown in multiple image channel types using a series of convolutional neural networks
The claim covers a computer program product that receives treated-cell and control-cell images across multiple image channel types and uses a series of convolutional neural networks to determine and quantify changes in each channel type
Automated response determination method using treated vs control image comparison
A method for determining an automated response of a cell to a drug, comprising: receiving images of treated cells, wherein treated cells are exposed to the drug, at an initial point in time; receiving images of control cells not exposed to the drug; forwarding the images of the treated cells and the images of the control cells to a convolutional neural network; determining by the convolutional neural network whether the images of the treated cells are different from the images of the control cells; and calculating a metric based on the determination by the convolutional neural network quantifying an amount of change caused by the drug on the treated cells
Quantifying changes across multiple points in time using a series of convolutional neural networks
The determination is performed using a series of convolutional neural networks to quantify changes at multiple points in time after the initial point in time
Quantifying changes across multiple drug types using a series of convolutional neural networks
The method further exposes treated cells to multiple types of drugs and quantifies changes associated with each drug type using a series of convolutional neural networks
Quantifying changes shown in each of multiple image channel types using a series of convolutional neural networks
The method receives treated and control cell images on multiple image channel types and quantifies changes shown in each channel type using a series of convolutional neural networks
Using a z-factor measure as the calculated metric in the method
The calculated metric is a z-factor measure
Overall, the claim coverage centers on image-based comparison of treated cells versus control cells using a convolutional neural network, followed by calculating a metric that quantifies drug-caused change. Dependent claim refinements specify named metrics, the CNN classification basis, and expanded analytical dimensions including multiple dosages, multiple drug types, multiple points in time, and multiple image channel types.
Stated Advantages
Improved sensitivity for detecting dose- and time-dependent structural damage using microscopy images and AI-based analysis.
An unbiased, phenotype-learning approach for structural toxicity detection.
Documented Applications
AI-based in vitro structural toxicity testing using microscopy images of drug-treated cells and vehicle/control cells, including detection of dose- and time-dependent structural damage.
Example applications demonstrated with aspirin, tamoxifen, acetylcysteine, and doxorubicin.
Interested in licensing this patent?