I built this prediction model because it connects my interests in biology and computer science. Programming can be more than a way to solve logic problems; it can also become a scientific tool for modeling biological patterns.
Cancer research now produces large sequencing datasets, and these datasets make it possible to train programs that search for disease-related signals. This project explores how machine learning can classify COAD tumor and normal samples from gene expression data.
My first model combined TCGA data with UCSC Xena data, and the result showed 100% accuracy. Although this looked impressive, it made the model less trustworthy because it might have learned dataset-source differences instead of real cancer biology.
After that, I chose to use only TCGA COAD data. This made the comparison cleaner, but it also created a new limitation: the TCGA-only normal group has only 41 samples, compared with 471 tumor samples.
These numbers belong to the Model page context. They should not be mixed with the separate COAD Data page RNA expression report.
Metrics, confusion matrices, and important features are shown together so the three machine learning models can be compared on one page.
Perfect or near-perfect scores are not automatic proof that a model is ready for diagnosis. High scores can happen when the test set is small, the task is too easy, or hidden data-source differences exist.