Ion channel predictions Machine Learning Case Study

For this case study, we trained a Random Forest classifier to distinguish between ion channels and other transmembrane proteins, and then applied the classifier to predict if any uncharacterized transmembrane proteins might be ion channels. The classifier performed very well in cross-validation with an area under the receiver operating characteristic curve of 0.99, F1-score of 0.922, and Matthew’s correlation coefficient of 0.918. Methods and results for this machine learning case study are described in detail in the Harmonizome publication.

Each row in the results table provides the gene symbol of a transmembrane protein (Gene column), provides the predicted probability that the protein is an ion channel (Probability column), indicates whether the protein is a known ion channel (Known column), indicates whether the protein was used as an example for training the classifier (Training column), provides the predicted false discovery rate if all predictions with better or equivalent probability were tested (FDR column), and lists up to two ion channels that the classifier thinks are most similar to the protein (Similar IC columns).

The table below shows predictions with probabilities greater than or equal to 0.5. To download the full table: