RSF determines tumor drug responsiveness and provides a response score for patient treatment with defective DNA damage repair mechanisms. The patented method is clinically significant, as malfunctions in DNA damage repair dramatically increase genome instability, a hallmark of most cancers.
Consistently, the RSF method predicts treatment responsiveness in multiple cancer types for multiple DNA damaging agents, such as PARP and CDK4/6 inhibitors. RSF is a classification algorithm that uses whole-genome sequencing (WGS) data to detect complex genomic biomarkers based on a distance-based projection.
The process of training an RSF is as follows:
-variant* type selection (e.g., copy number alterations, deletions or insertions, structural variants: duplications or translocations),
-histogram-based features extraction (calculation of frequencies and densities of variant types),
-Random Similarity Forest classifier training (which performs a distance projection of variant distributions).
Inside, RSF combines the properties of other AI algorithms, namely Random Forests and Similarity Forests. As a result, RSF is capable of learning from simple numeric data (e.g., gene amplifications) and complex data (e.g., structural variant distributions) at the same time. RSF’s data flexibility allows it to process different data types and, therefore, all the modalities WGS data can offer.