Machine Learning of Cell Mechanobiology Experiments to Predict Metastatic Risk

Background: Today 90% of cancer mortality is related to metastasis. Prognosis of metastatic risk directly affects cancer treatment. Current prognosis is done via histopathology, statistics, or genetics, yet these methods inaccurate, slow, and costly. We have developed a rapid (~2hr) approach to quantify the metastatic potential and in vivo risk of tumor samples. Specifically, we have shown that invasive cell-subsets will forcefully indent impenetrable, physiological-stiffness polyacrylamide gels as compared to non-invasive cells (e.g. benign). The mechanical invasiveness capacity or metastatic risk of a sample is then determined by the number of indenting cells and their attained depths. Methods: We utilized the lab’s experimental results to create a support vector classifier (SVC) model to predict the metastatic risk of a new, untested samples. The model was trained in a supervised manner on features of tumor samples and cell lines, e.g. tumor origin, patient age, gender and mechanical invasiveness. The classification was made according to literature and ATCC for cell lines and clinical histopathology-based patient prognoses for tumor samples. Results: classifying metastatic potential to five classes (normal, benign, cancer metastatic–non/low/high) attained an average sensitivity and specificity of 0.71 and 0.91. In addition, when simply classifying to two classes, normal/benign/cancer non-metastatic vs. cancer metastatic as typically done in the literature, we obtained sensitivity and specificity of 0.92 and 0.96, improving on other works in this field. Conclusion: The results support the notion that mechanobiology parameters can be used to improve cost, time and accuracy of cancer invasiveness prognosis.