Deep phenotyping

Deep Phenotyping - Clustering Approach

Calculating Clinical Similarity on Patient-level

Plain Binary comparison

Calculate the clinical similarity / distance based on the boolean vectors (doesn't require the structure of the HPO DAG). Example metrics:
Hamming
Dice
Jaccard

Semantic Similarity

Calculate clinical similarity while taking semantic similarity into account (requires the HPO DAG). Example metrics:
HRSS

Calculating Clinical Similarity on Group-level

Occurrence ratio

Calculate the clinical similarity by calculating the occurrence ratio (Hanneke et al) between predefined groups. In contrast to patient-level, this is a supervised approach.

Prioritisation

Phenotypes were ranked on phenotypic specificity. (Inspired by Hanneke et al)

Pathophysiology

The annotated HPO file (HPOA) was used to infer the genes associated with the phenotypes. These genes were supplied to PHRANK. This tool outperforms Phevor, PhenIX and Phenomizer at ranking the causative genes & diseases.

Training

We first explored the potential of clustering for deepphenotyping on the ACTG1 ACTB case study collection.

Validation

The Deep Phenotyping clustering algorithm is validated on the COG-CDG case study collection of Hanneke et al. Where we hope to acquire the same pathophysiological insights.