Deep phenotyping to aid identification of coding & non-coding rare disease variants

Editor's Notes

• #4: There is a lot we don’t know about the genome As of March 2017, OMIM number: 3398 unknown 4,964 known ClinVar number: 121,000 at least with the addition that these are variants that researchers have found suspicious, due to rarity in the population or something else, contextually 160k variants in the entire genome is not much
• #8: If clinvar + omim 20  80%
• #12: This was the novel case we solved. The UDP patient had a number of signs and symptoms including various platelet abnormalities. The same heterozygous, missense mutation was seen in 2 patients and ranked top by Exomiser. It had never been seen in any of the SNP databases and was predicted maximally pathogenic. Finally a mouse curated by MGI involving a heterozygous, missense point mutation introduced by chemical mutagenesis exhibited strikingly similar platelet abnormalities.
• #13: Genomiser scores variants either through existing methods such as CADD or a bespoke machine learning method and combines these with allele frequency, regulatory sequences, chromosomal topological domains, and phenotypic relevance to discover variants associated to specific Mendelian disorders.
• #14: We included only those variations and publications judged to provide plausible evidence of pathogenicity. First, the phenotypic abnormalities of the individual carrying the variant were assessed and a variant was included only if the disease association was regarded as plausible on the basis of evidence such as familial cosegregation or experimental validation, using techniques such as luciferase reporter assays, electrophoretic mobility assay, or telomerase activity assay. In some cases pathogenicity was assigned based on curator judgment or computational predictions; for instance, mutations in RNA genes that affected RNA secondary structure elements such as stem loops were included. https://pixabay.com/en/night-forest-trees-moonlight-1245875/ https://pixabay.com/en/taiwan-macaques-shoushan-macaque-1345438/
• #15: https://pixabay.com/en/night-forest-trees-moonlight-1245875/ https://pixabay.com/en/taiwan-macaques-shoushan-macaque-1345438/
• #16: We included only those variations and publications judged to provide plausible evidence of pathogenicity. First, the phenotypic abnormalities of the individual carrying the variant were assessed and a variant was included only if the disease association was regarded as plausible on the basis of evidence such as familial cosegregation or experimental validation, using techniques such as luciferase reporter assays, electrophoretic mobility assay, or telomerase activity assay. In some cases pathogenicity was assigned based on curator judgment or computational predictions; for instance, mutations in RNA genes that affected RNA secondary structure elements such as stem loops were included.
• #17: Thus, approximately 36,000 negative examples are available for every positive one. In such extremely unbalanced conditions, classical computational and machine learning methods tend to perform poorly. This is because they learn overwhelmingly from negative examples, which leads to a sensitivity and precision close to zero on new (test) data.50
• #18: ANIMATION NOTE, CLICK ONCE FOR EACH STEP In order to train the ReMM model, we first divided the majority class (probably non-deleterious variant sites) randomly into n = 100 partitions and then we added all the minority instances (non-coding Mendelian mutations) to every partition. We chose 100 partitions because no substantial performance improvements were observed when more partitions were utilized (data not shown). Moreover, in each partition we synthetically oversampled the minority positive class, using the synthetic minority over-sampling technique51 (SMOTE) with a number of nearest neighbors k = 5. With the SMOTE approach we generated synthetic instances two times the cardinality of the positive class. We then randomly undersampled the majority negative class to obtain a three times larger set of negative examples. The resulting dataset was used to train a random forest (RF) classifier52 (forest size 10; larger forests do not significantly improve the performances; data not shown) that outputs a probability to estimate whether a given position in non-coding genome can cause a Mendelian disease if mutated. The overall process of over- and undersampling and the training of the RF was repeated for all the n partitions. Finally, the probabilities estimated by each RF were averaged and the resulting “consensus” probability of the hyperensemble represents the final ReMM score. Our method was implemented in Java using Weka.53
• #20: Benchmarking experiments for Genomiser were performed using 10,419 simulated rare disease genomes based on the 453 regulatory Mendelian mutations and 1,092 whole-genomes VCF files from the 1000 Genomes Project14 (05/02/2013 release). For autosomal-dominant diseases, one heterozygous mutation was added, and for autosomal-recessive diseases, either one homozygous mutation or two heterozygous mutations were added to the 1000 Genomes VCF file. For these experiments, the phenotypic (HPO) annotations for the corresponding disease in OMIM were taken on 8/7/2015 from the annotation files of the HPO team. To measure the ability of Genomiser to detect known disease-gene associations, we repeated the analysis with incomplete (maximum of three HPO annotations), noisy (two random HPO terms added), and imprecise (two of the original HPO annotations replaced by the more general parent terms in the ontology) annotations. These simulated genomes were run through the default settings of Genomiser. In the first step, genes and associated variants are removed where there is little similarity between observed phenotypes and direct or inferred knowledge from disease and model organism databases. Note in this step, distal (>20 kb from a gene) variants that reside in predicted enhancers from FANTOM5 and Ensembl are associated with the most phenotypically similar gene in the topological domain containing the enhancer rather than simply taking the closest gene. Distal variants that do not reside in a predicted enhancer are removed, followed by the exclusion of any that are common (>1% minor allele frequency [MAF]) in the 1000 Genomes Project, NHLBI Exome Sequencing Project (ESP), and Exome Aggregation Consortium (ExAC) datasets. Finally, the remaining variants are prioritized by a composite score of the minor allele frequency, phenotypic similarity, and pathogenicity (using the ReMM score for non-coding and the existing hiPHIVE method for coding and splice sequences). To assess our performance, we measured how often the seeded regulatory Mendelian variant was ranked first among the full set of the variants of the simulated Mendelian disease genomes. We note that the ReMM scores used in the Genomiser experiment were computed by ten-fold cross validation: the scores for the mutations included in each fold were obtained through a model trained on mutations not included in that fold, but only on those of the remaining nine. In other words, the ReMM score used in Genomiser for a specific mutation was obtained by a model not trained on this variant
• #21: Fully translational – from bench to bedside – group of stakeholders, contributors, and partners