Fig. 4: InpherNet improves on existing indirect evidence gene prioritization methods and provides the strongest gene neighbor contributor to the ranking.

(a) We took 255 real patients with diverse prediagnosed conditions and first used two patient evidence-based methods to highlight cases where either patient evidence is missing or differs significantly from the current patient. In 115 cases the causative genes were not among the top 10 PhenIX-ranked genes, and in ~40% fewer cases (70), the causative gene was not in the Phrank_HPOA top-ranked 10 genes. Each case is then reviewed by 8 inference tools after removing the highest ranked 10 genes that were already determined as not causative. (b) InpherNet offers a large improvement on all tools for the tougher (preferred, smaller) Phrank > 10 set, with over three times as many cases where the causative gene ranks 1, and almost twice as many cases where it ranks 1–5 as all other tools. (c) InpherNet still outperforms all other tools over the larger (weaker) PhenIX > 10 set, ranking 10 or more causative genes among top 1–2, 1–3, and 1–5 as all other tools. (d,e) InpherNet also outputs a list of gene neighbors ranked by their relevance to the patient phenotype. In both test sets we see that all 4 types of indirect evidence (i.e., orthology, paralogy, pathways, and interactions) contribute the most to the correct gene being ranked on top for different candidates.