The probability of primary colorectal tumour patients developing metastasis is currently dependent on their node status, which is not always accurate. Mestastop has integrated the functional properties from primary patient tumour-derived cells into a learning algorithm to predict the metastatic potential of non-metastatic pathological grade patients, blinded of tumour staging or node status. We have duplicated the complete and complex metastasis biology in vitro, distinguishing functional differences, between moving and growing tumour cells in multiple cell lines. Data derived from this platform was used to train a learning algorithm METSCAN™, which identifies cells having different moving and growing phenotypes. Various supervised learning algorithms were used to differentiate the metastatic from non- and pre-metastatic cells, identifying a specific pattern. This was followed by feature selection using multiple methods, which reduced the feature size from twenty-five to eight and yielded an accuracy of ninety percent. Cellular functional data from patient samples were then fed into this algorithm to identify the distribution of cells between met and non-met axis, predicting their metastatic probability. To minimize false negatives and overfitting, the data set was further converted into a binary class label, of metastatic and non-metastatic cells, following which Support-Vector-Machine showed greater than ninety percent accuracy. Among the thirteen patient samples of the current blinded study, follow-up has identified three patients to be metastatic, all of which were correctly predicted by METSCAN™. There were no false negatives, but a few false positives, which can be indicative of higher platform sensitivity of METSCAN™.