• Title: Learning Reveals Cancer Metastasis and Therapeutic Antibody Targeting in the Entire Body

• Address: https: //www.cell.com/cell/fulltext/S0092-8674(19)31269-3

• Project homepage: https: //discotechnologies.org/DeepMACT/

• GitHub repository: https: //discotechnologies.org/DeepMACT/

📝Introduction | Deep Learning Reveals Whole Body Cancer Metastasis and Therapeutic Antibody Targeting

Background.

Research Question & Motivation:

Objective & Contribution:

1. In order to address the above problems, the research constructed a fully automated analysis process based on deep learning and vDISCO imaging - DeepMACT.(What did it do).

2. Specifically DeepMACT is a collection of deep learning models. vDISCO technology can amplify the fluorescence signal by more than 100 times, effectively improving the imaging sensitivity. Combined with CNN's image analysis algorithm,, it is used to automatically quantify metastasis and improve the accuracy and efficiency of automatically identifying metastatic foci. (How well does the method work and what problem does it solve)

3. Quantification was performed by DeepMACT in five different metastatic cancer models including breast, lung and pancreatic cancers, which have different organ tropisms. And successfully evaluated the metastatic foci targeting efficiency of anti-CA XII antibody. (What are the results and do the findings matter)

🗺️ Methods

Mouse model

• Spontaneous breast cancer metastasis model

• Estrogen positive breast cancer model and brain metastatic lung cancer model

• Pancreatic cancer model

Data Annotation

Deep learning-based metastasis detection

DeepMACT Model Architecture

Custom 3D U-Net implementation

Training set, validation set and test set (division)

Training process

• Training process:
• Initial training: train multiple models by k-fold cross-validation (k=5), testing different hyperparameters for 10 epochs.
• Optimization of hyperparameters: select the best performing hyperparameters and continue training for the remaining epochs until 40 epochs.

• Data enhancement: use random vertical and horizontal flips to enhance data diversity.

• Input Normalization: Normalization based on local sub-volume peaks works best.

• Loss function: weighted binary cross entropy is used, with foreground (FG) pixels weighted more heavily to cope with category imbalance.

• Training data optimization: only 90% of the training data containing foreground is used to improve validation set performance.

• Optimizer and learning rate:
• Use Adam optimizer with an initial learning rate of 10^-4.
• The learning rate is reduced by a factor of 10 every 2 epochs when the loss is not decreasing, down to 10^-7.

• Training Time: 40 epochs training time is about 20-30 minutes on a workstation with NVIDIA Titan XP GPUs.

Test and process inference model

• The study used k-fold cross-validation with k=5

• The trained algorithm was used to generate probabilistic masks from the three projected views (PXY, PYZ, and PZX), with each pixel value indicating the network's confidence that the pixel is a transfer or not. The final segmentation mask is obtained by combining the three 2D probability masks into a 3D space. The default confidence threshold was set to 50%

• The connected component analysis was used to find out and segment the transferred region.

Performance Evaluation

• Performance is evaluated using the F1 score, which combines precision (the proportion of correct predictions that are positive) and recall (the proportion of correct predictions that are actually positive), and is equivalent to the Sørensen-Dice coefficient, which is commonly used in image segmentation tasks.

• Performance is evaluated using the full test set and the variability of the results is estimated by a self-help method (n=1000 resamples). The accuracy of 3D segmentation of metastatic cancers was verified by comparing it with 3D contours drawn by a manual annotator, with an overlap accuracy of 90% for the worst segmentation and 97.5% or more for 90% of metastatic cancers.

Metastatic foci analysis

Organ Alignment | Manual segmentation of region of interest

• Manually segment 3D contours of organs of interest (lungs, brain, kidneys, heart, liver) using Fiji software.

• Custom Python scripts were used to detect whether metastatic centers of mass were located in the 3D segmentation of these organs, and unattributed metastases were labeled as "rest of torso".

• Lung volumes were calculated and tumor density was assessed as the proportion of metastases in the lungs.

Drug targeting analysis

• The targeting effect of the 6A10 antibody is evaluated by analyzing the distribution of fluorescent signals within the metastases and in the surrounding area.

• The ratio of signal intensity inside and outside the metastases (antibody signal ratio) was calculated; if the ratio was greater than 1, the targeting effect was significant.

• The difference in signal intensity between the metastases and the surrounding area was evaluated using the Welch t-test.

Fluorescence signal characterization

• Fluorescence signal profiles were studied at different excitation wavelengths (470 nm, 561 nm, 647 nm).

• Image acquisition was performed after removal of lung metastases using uDISCO and vDISCO, and the signal-to-background signal ratio (SBR) was analyzed.

• Normalized signal contours were compared and representative graphs were drawn.

Metastasis diameter and vessel distance

• The diameter of the metastases was manually verified.

• Ten points at the border of the metastasis were randomly selected, the distance to the nearest vessel was measured, and the average value was calculated to generate a scatter plot of the distance between the metastasis and the vessel.

🍎 Results

vDISCO imaging of cancer metastases in optically transparent mice | Imaging Results

Deep Learning for Detecting and Quantifying Metastasis | Segmentation Performance Comparison

• DeepMACT uses a U-net-like CNN architecture to extract cancer features and perform pixel-level segmentation through encoding and decoding units

• By comparing DeepMACT to traditional filter detection methods and human annotators, DeepMACT outperforms metastasis detection with an F1 score of 80%, close to the performance of human experts, and significantly improves detection speed.

• DeepMACT is able to accomplish in a few hours what would normally take months, greatly improving productivity. The method supports adjusting the balance of precision and recall to adapt to different needs, improving the flexibility and utility of detection.

DeepMACT reliably detects micrometastases in different tumor models | Performance in different tissues, time points and multiple tumor models

• The researchers used the method to analyze whole mice and found 520 micrometastases, 306 of which were in the lungs. Most of the micrometastases were located in deeper tissues, about 1 cm from the surface, making them difficult to detect by traditional methods.

• The micrometastases in the lungs were uniformly distributed in all lobes of the lungs and were size-independent, suggesting that these micrometastases were formed by multiple independent metastatic events. In addition, the size of the micrometastases varied considerably, and although most of them were in close proximity to each other (<1 mm), there were also isolated lesions with a maximum distance of 9.3 mm.

• Further analysis revealed that many of the micrometastases had cell counts of a few hundred or less and diameters of less than 50-100 μm, making these micrometastases very difficult to detect by conventional methods. At the same time, the load of lung metastases is much larger than that of the trunk, and lung micrometastases are, on average, 30% larger than that of the trunk, with more than two times the cell count of the trunk.

• To validate the accuracy of DeepMACT, the researchers first performed negative control experiments and found that the hyalinization process does not result in false positives until the tumor cells have formed metastases.

• DeepMACT is effective in detecting metastases in a variety of mouse breeds and tumor types, including immunodeficient and immunocompetent mice. Three different tumor models were tested:
• MCF-7 model: metastases in lung, liver, kidney, bone and brain.
• R254 model: mainly in lung, liver and peritoneum
• H2030-BrM3 model: brain was the primary site of metastasis (31 foci)

• The researchers further investigated the evolution of the metastatic process at different time points, using MDA-MB-231 breast cancer cells injected into the heart chambers of mice and analyzed at 2, 6, and 14 days, showing that the number of metastatic foci continued to increase over time, and that the lungs were consistently the most metastasis-concentrated organ.

DeepMACT Reveals Therapeutic Antibody Targeting at the Level of Individual Micrometastases | Antibody Distribution Whole Body Quantification Results

• The study was conducted with monoclonal antibody 6A10 targeting human carbonic anhydrase XII (CA12), known to inhibit tumor growth and enhance chemosensitivity. In the experiment, 20 μg of Alexa-568-labeled 6A10 was administered intravenously 9 weeks after transplantation of MDA-MB-231 breast cancer cells, and its distribution was analyzed 2 days later.

• Antibody 6A10 efficiently aggregated to the primary tumor and some metastases (e.g., axillary lymph nodes and lungs).

• However, not all micrometastases were targeted by the antibody and there was heterogeneity (e.g., some were successfully targeted in the lung and some were not). In the lungs, 85% of the micrometastases were successfully targeted, compared with 66%in other sites; the antibody signal was more concentrated and evenly distributed in the lung foci.

• Size correlation: large micrometastases were more likely to be targeted by the antibody (88%); the targeting rate was 67% for small-sized foci.

• Nonspecific binding: Although the antibody did not bind to mouse CA12, some off-target binding was observed, presumably due to nonspecific binding.

Potential Mechanisms for Failure of Antibody Drugs to Completely Target Micrometastases

• Antibody drugs fail to target up to 23% of micrometastases in some cases. The goal of the study was to understand possible mechanisms for this "missed target" phenomenon.

• The researchers considered whether some micrometastases do not have neighboring blood vessels, which prevents the drug from being delivered. They labeled lung tissue using lectin staining and found that all of the micrometastases were close to blood vessels at distances of 1-6 μm, which is much smaller than the diameter of a cell (~10 μm). Thus, inadequate drug delivery pathways were not the main problem.

• The researchers turned to analyzing whether the tumor microenvironment affects the targeting ability of antibody drugs. A deep learning method (DeepMACT) was utilized to study the spatial distribution relationship between targeted and untargeted foci. The results showed that successfully targeted foci were more likely to cluster together (average distance ~0.8 mm), while untargeted foci were more dispersed in distribution (distance ~1.7-2.0 mm). This non-random distribution suggests that certain local microenvironments may influence antibody recognition or penetration.

🐿️ Conclusions

🤗 Insights.

1. It is not the case that the segmentation is done, but the analysis actually accounts for the majority of the workload.

2. The analysis results of the mouse whole body is what we should pay most attention to, but the limitation of storage and computational costs, can only first umbrella the big picture, do it on the small picture, and finally splicing, these are time costs.

3. It is also necessary to pay attention to the comparison of the processing time overhead of different methods, do not be lazy.

4. The result part should be at least 4~5 small paragraphs, which can be the expected result when writing.

Citation