PaperlayerView exact revision on arXiv

Source-linked AI summary

Phen-Gen: combining phenotype and genotype to analyze rare disorders

Asif Javed, Saloni Agrawal, Pauline C. Ng

arXiv:1502.07829v1q-bio.GN

TL;DR

Rare-disorder gene prioritization remains challenging, including because many cases lack a known molecular cause and symptom-driven methods do not span the genome. Phen-Gen combines sequencing, symptoms, and prior disease and gene knowledge in a Bayesian framework; it ranked causal genes first in 88% of simulations and outperformed previously published methods by 19–58%.

  • Problem

    Rare-disorder diagnostic yield remains moderate, many cases lack a known cause, and existing symptom-driven methods do not extend across the entire genome.

    Source: Introduction · p. 2

  • Method

    Phen-Gen combines sequencing data and patient symptoms with prior disease knowledge and gene interactions in a Bayesian framework.

    Source: Introduction · p. 2

  • Results

    Phen-Gen ranked causal genes first in 88% of simulations and improved by 19–58% over eXtasy, VAAST, and PHEVOR; it ranked an implicated gene first in 8/11 patients.

    Source: Introduction · p. 2

  • Takeaways & Limitations

    Phen-Gen assigns damaging probabilities to regulatory variants and is presented as applicable to both coding and noncoding variation.

    Source: Introduction · p. 2, Comparison · p. 5

    The real-patient evaluation was constrained by noisy variant calls and exclusion of seven families whose reported variants did not match the provided inheritance pattern.

    Source: Performance using real dataset · p. 25

Abstract

from arXiv · show

We introduce Phen-Gen, a method which combines patient disease symptoms and sequencing data with prior domain knowledge to identify the causative gene(s) for rare disorders.

Introduction

Phen-Gen combines sequencing, symptoms, disease knowledge, and gene interactions in a Bayesian framework, with exome-centric and genome-wide analysis paths. Simulations and patient data show strong causal-gene ranking, including for novel genes and noncoding variants.

Introduction

  • Phen-Gen combines patient sequencing and symptoms with disease knowledge and gene interactions in a Bayesian framework, using exome-centric and genome-wide paths.The introduction identifies limited integration of prior knowledge and the lack of symptom-driven methods extending across the genome as issues motivating the method.Source: Introduction · p. 2
  • Phen-Gen ranked causal genes first in 88% of simulations and improved on eXtasy, VAAST, and PHEVOR by 19–58%.It reported a 13–16% improvement over PHIVE and ranked the causal gene first in 8/11 patients.Source: Introduction · p. 2, Simulated datasets · p. 3

Prediction framework

  • In variant-level evaluation, at least 84% of mutations in all three coding categories received damaging probabilities of 0.9 or higher.The genomic predictor scored 80% of HGMD regulatory variants and assigned 38% a damaging probability of at least 0.5.Source: Prediction framework · p. 3

Simulated datasets

  • For unknown disease-gene links, Phen-Gen ranked the causal gene first in 71% of combined unknown and masked simulations.For unknown disorders, it identified the true gene in 43% of dominant and 92% of recessive cases.Source: Simulated datasets · pp. 3–4
  • Phen-Gen correctly identified the genetic cause in 49% of noncoding-variant cases, with phenotypic information contributing substantially to prediction.The authors note that predictions may improve as regulatory roles of genomic sites are further characterized.Source: Simulated datasets · p. 4

Real datasets

In a study of trio families with children experiencing intellectual disability, Phen-Gen ranked reported genes in the top 10 for 13/16 genes. Applying the original study’s variant filters improved rankings.

Real datasets

  • With the original study’s criteria, Phen-Gen ranked an implicated gene first for 8/11 patients and placed all reported genes in the top five.Without those criteria, 13/16 reported genes ranked in the top 10.Source: Real datasets · p. 4
  • The evaluation used 11 families with variants implicated in recessive or X-mode inheritance from a study of 100 father-mother-child trios.Each child in the study had intellectual disability symptoms.Source: Real datasets · p. 4
  • Unlike the original study, Phen-Gen allowed rare dbSNP variants with minor allele frequency below 1%.The original study removed mutations reported in dbSNP or observed in its in-house dataset from further evaluation.Source: Real datasets · p. 4

Comparison

Phen-Gen outperformed VAAST, eXtasy, VAAST+PHEVOR, and PHIVE in reported comparisons, and it assigns damaging probabilities to many regulatory variants. The authors emphasize its application to both coding and noncoding variation.

Comparison

  • For recessive disorders, Phen-Gen ranked the causal gene first in 97% of cases, versus 78% for VAAST and 30% for eXtasy.The comparison also found first-rank results of 82% for Phen-Gen, 62% for VAAST, and 24% for eXtasy in dominant disorders.Source: Comparison · p. 5
  • Phen-Gen outperformed VAAST, eXtasy, and VAAST+PHEVOR by 19–58% in simulations.It identified causal genes in 90% of dominant and 88% of recessive simulations, compared with 66% and 68% for VAAST+PHEVOR.Source: Comparison · p. 5
  • Compared with PHIVE, Phen-Gen showed 16 and 13 percentage-point advantages for dominant and recessive causal-gene identification, respectively.PHIVE’s reported first-rank rates were 66% for dominant and 83% for recessive cases.Source: Comparison · p. 5
  • Phen-Gen assigned damaging probabilities to 80% of HGMD regulatory variants, compared with 18% of damaging regulatory variants falling within FunSeq’s first filtration criteria.The passage describes FunSeq’s result as an enrichment of damaging mutations.Source: Comparison · p. 5
  • Phen-Gen integrates patient symptoms and is presented as applicable to both coding and noncoding variation.The authors describe symptom information as advantageous for understanding disease mechanisms.Source: Comparison · p. 5

Figure legend

The figure and table legends describe comparison setups and evaluated categories, while the methods passages detail gene and variant processing, phenotype scoring, and gene-interaction evidence. The reported network validation tests whether disease-associated gene pairs are represented in the network.

Figure legend

  • Figure 1 compares Phen-Gen with VAAST, eXtasy, and VAAST+PHEVOR by how often each narrows the true gene to 1, 5, or 10 candidates.Table 1 reports simulated coding-predictor performance by causal-variant class and known versus unknown disease-gene links; its top three rows compare complete, genotype-only, and phenotype-only approaches. The gene set comprised 26,803 protein-coding genes drawn from four popular gene lists.Source: Figure legend · p. 8, Table 1 · p. 10, Gene list · p. 11

Nonsynonymous variants

  • Coding variants are classified by consequence and assigned deleteriousness probabilities; nonsynonymous estimates combine SIFT and PolyPhen-2 using empirical distributions and Bayes’ rule.The resulting probability continuum can be thresholded to 91% accuracy with less than 2.1% false positives, compared with MutationTaster’s 87% accuracy and 4% false positive rate on the same data.Source: Variant prediction · p. 11, Nonsynonymous variants · pp. 11–12

Splice site variants

  • Splice-disrupting mutations are defined within 8 bp of donor sites or 3 bp of acceptor sites; this definition misses 8% of HGMD mutations, with another 1% outside the gene list.The 2 bp at each intron end are treated separately.Source: Splice site variants · p. 12

Small insertions and deletions

Genomic variants

  • The genomic predictor combines conservation, ENCODE regulatory annotations, and proximity to genes; two positive and two neutral datasets address sparse regulatory-mutation data.The estimated prior probability of a damaging mutation in annotated regions is 0.0688. Additional annotations were limited by small positive sets and the resulting risk of sparse, noisy bins.Source: Genomic variants · pp. 13–14

Pooling

  • Gene-level loss of function pools damaging variants by parental chromosome, excludes common variants, and compares inheritance patterns with disease inheritance.When phase is unavailable, the two highest variants are assumed to lie on opposite chromosomes; noncoding mutations are assigned to the nearest gene within 50 kb.Source: Pooling · p. 14

Phenotypic prediction

  • Phenotype scoring maps symptoms to known disorders, assigns disease probabilities to implicated genes, and combines probabilities across disorders for pleiotropic genes.The disease probabilities derive from multiple-testing-corrected matches, while random walk with restart spreads part of each gene’s phenotypic probability across its interaction partners.Source: Phenotypic prediction · pp. 15–16

Gene-gene interaction network construction

  • The weighted gene-interaction network integrates pathway, data-mining, ontology, and co-expression evidence; 31% of HGMD and 39% of OMIM same-disorder gene pairs had a direct edge.Disease-implicated genes also had more edges than expected, with reported P ≤10^-166 for HGMD and P ≤1.8×10^-192 for OMIM genes.Source: Gene-gene interaction network… · pp. 16–17

Combined

Phen-Gen combines phenotype and genotype predictions for each gene under an independence assumption. The evaluation uses simulated patients, including single-variant, compound-heterozygous, and heterogeneous-symptom scenarios.

Combined

  • The Combined model treats phenotypic and genotypic probability estimates as independent observations of a gene’s underlying probability of involvement.It combines the two predictions for each gene within a Bayesian framework.Source: Combined · p. 17

Computational efficiency

  • For computational efficiency, Phen-Gen scans variant and annotation files together, filters through a 19% genome sieve, and stores some preprocessing data in binary form.Interval forests define regions of interest, and noncoding database overlap information uses one byte per interval.Source: Computational efficiency · p. 17

Simulating in silico patients

  • Single-variant simulations inserted HGMD causal mutations into 1000 Genomes individuals and paired them with symptoms drawn from OMIM disease descriptions.The study used 765 diseases and evaluated each causal variant in turn across 1,092 individuals.Source: Simulating in silico patients · p. 18
  • Compound-heterozygous simulations inserted two reported variants in trans and included recessive disorders with at least two variants in or near the same gene.The genomic predictor was used when either spiked variant was noncoding.Source: Simulating in silico patients · p. 18
  • To test robustness to phenotypic heterogeneity, the study generated five symptom histories per disorder by sampling symptoms according to their reported frequencies.Each individual genome was evaluated with each of the five histories.Source: Simulating in silico patients · pp. 18–19

VAAST, PHEVOR and eXtasy simulations

Across simulations, Phen-Gen outperformed eXtasy, VAAST, and VAAST+PHEVOR in the reported comparisons. Its advantage was also reported for heterogeneous symptom profiles.

VAAST, PHEVOR and eXtasy simulations

  • 82% for dominant and 97% for recessive disorders: Phen-Gen identified the etiological gene, compared with VAAST’s 62% and 78% narrowing the gene to within 14 genes.VAAST’s rates are averages across dominant and recessive cases.Source: VAAST, PHEVOR and eXtasy simulations · pp. 23–24
  • 90% for dominant and 88% for recessive simulations: Phen-Gen identified the causal gene, compared with 66% and 68% for VAAST+PHEVOR on the same cases.The PHEVOR comparison used 100 simulations; its server accepts at most five symptoms.Source: VAAST, PHEVOR and eXtasy simulations · p. 24
  • 91% for dominant and 98% for recessive disorders: Phen-Gen narrowed the causal variant to within 10 variants, compared with eXtasy’s 69% and 76%.eXtasy was assessed both with all variants and with rare variants only.Source: VAAST, PHEVOR and eXtasy simulations · p. 24
  • 19–58%: Phen-Gen outperformed eXtasy, VAAST, and VAAST+PHEVOR in the direct comparison.Restricting Phen-Gen to symptoms accepted by eXtasy reduced true-gene identification by 1%.Source: VAAST, PHEVOR and eXtasy simulations · p. 25
  • Phen-Gen retained a reported relative performance advantage over VAAST on the 44-disorder phenotypically heterogeneous dataset.The evaluation tested cases where disease symptoms were not completely specific.Source: VAAST, PHEVOR and eXtasy simulations · p. 25

Performance using real dataset

The real-data evaluation was constrained by noisy variant calls and the inability to reproduce bench validation, leaving eleven families for analysis. Phenotypic ranking results were significant, while five variants remained unresolved by bioinformatics alone.

Performance using real dataset

  • No patient-sample access prevented reproducing bench validation, and seven of eighteen initially considered families were excluded, leaving eleven families.The exclusions reflected variants not observed in the provided data in the reported inheritance pattern.Source: Performance using real dataset · pp. 25–26
  • 16/21 genes agreed with the original study after applying its filtering criteria, while five variants remained possible false positives requiring further validation or functional analysis.Two of the five were homozygous in the patient and heterozygous in two other families; the other three appeared only in their respective family, and two of those were novel indels.Source: Performance using real dataset · p. 26
  • Phenotypic-only rank sums for the 58 patients’ reported genes were lower than expected by chance (P ≤ 0.0036).The analysis ranked genes using patient symptoms alone and summed ranks across patients.Source: Performance using real dataset · p. 26
  • Rank sums were lower than expected by chance for known genes (P ≤ 2.44×10^-8), candidate genes (P ≤ 0.0051), and unknown genes (P ≤ 0.0192).The authors suggest the unknown-gene set may contain true positives and that corroboration could improve diagnostic yield.Source: Performance using real dataset · pp. 26–27

Comparison with PHIVE

Phen-Gen’s reported efficacy was 13–16% higher than PHIVE’s reported results, but a comprehensive same-dataset comparison was not possible. The discussion also notes limitations in PHIVE’s phenotype and genotype predictions.

Comparison with PHIVE

  • 13–16% higher efficacy: Phen-Gen’s reported results exceeded PHIVE’s reported 66% dominant and 83% recessive causal-gene identification rates.The comparison is not based on a comprehensive shared-dataset evaluation; PHIVE used 869 diseases versus Phen-Gen’s 765.Source: Comparison with PHIVE · p. 27
  • PHIVE’s mouse knockout phenotype links cover about one-third of human genes, and mouse models may not represent human-specific disease morphology.The authors note that integrating animal-model information with human disease matches could still improve Phen-Gen predictions.Source: Comparison with PHIVE · p. 27
  • PHIVE’s performance declined by almost half on in-house exome simulations, which the authors attributed to reliance on allele-frequency information unavailable for novel variants.They also raise possible overfitting to curated 1000 Genomes data and note that practical indel calls may be more error-prone.Source: Comparison with PHIVE · p. 28

Comparison with FunSeq

Phen-Gen assigns probabilities to many regulatory variants that FunSeq’s initial filter excludes, and phenotype information improves regulatory-variant prioritization in simulations.

Comparison with FunSeq

  • FunSeq’s initial filter captures 18% of damaging regulatory variants, while Phen-Gen assigns probabilities to 80% of them.The remaining variants would be annotated as benign by that FunSeq filter; Phen-Gen assigns them lower probabilities based on genotypic information.Source: Comparison with FunSeq · p. 28
  • 49% of simulated regulatory cases with phenotype information identify the causal variant, with a 30 percentage points advantage over a genotype-only approach for predicting the causal gene.The passage attributes Phen-Gen’s predictive power to phenotypic corroboration.Source: Comparison with FunSeq · p. 28
  • 69% of cases place the disease gene in Phen-Gen’s top 10 candidate list.Source: Comparison with FunSeq · p. 28

Prior probability for genomic predictor

The genomic predictor uses an estimated probability that an annotated variant is damaging as its prior, derived from estimates of functional genome coverage and variant conservation.

Prior probability for genomic predictor

Incorporating pedigree information

Phen-Gen filters candidate genes and variants according to inheritance pattern and can restrict analysis to variants consistent with the family pedigree.

Incorporating pedigree information

  • For dominant inheritance, genes with one or more damaging mutations are evaluated, retaining variants present in cases and absent from controls and genes whose predicted probability is at least as high in cases.Source: Incorporating pedigree information · p. 29
  • For recessive or compound heterozygous inheritance, genes need at least two damaging mutations, with a further requirement for one variant from each parent when both parents are analyzed.The criteria also require variants to be present in cases and permit one or zero alternate-allele copies in controls; the filtering is described as reducing false positives while allowing compound heterozygosity.Source: Incorporating pedigree information · p. 29
  • Users can restrict analysis to pedigree-consistent variants; inconsistent variants may be allowed when de novo mutations are likely, though a pedigree-based caller is recommended to prune false de novo calls in real datasets.Source: Incorporating pedigree information · p. 29

Evaluating null distribution of genes

Phen-Gen’s predicted damaging-variant genes were compared with published loss-of-function and intolerance datasets. Genes above its one-percent null-distribution cutoff were enriched in both datasets despite methodological and dataset differences.

Evaluating null distribution of genes

  • Genes above Phen-Gen’s one-percent null-distribution cutoff were enriched in both comparison datasets (P ≤ 8×10-3 and 10-2).The comparisons were with loss-of-function genes and Residual Variation Intolerance Score datasets, despite methodological and dataset differences.Source: Evaluating null distribution of genes · pp. 29–30
  • Across masked and unknown simulations, Phen-Gen ranked causal variants first in 71% of cases.Supplementary Table 2 covers performance in novel disease-gene discovery.Source: Evaluating null distribution of genes · p. 34, Table 2 · p. 35
  • Using ESP, dbSNP, and 1000 Genomes to define common variants conferred a 13–21% advantage over using only ESP.The comparison used simulated patients with OMIM-listed symptoms and HGMD-reported variants.Source: Table 3 · p. 36
  • Compound-heterozygous mutation-pair performance was similar to the 96% observed for recessive coding-variant simulations.The table covers mutation pairs in coding and genomic regions.Source: Table 4 · p. 37
  • With variable symptoms, performance dropped 5–7% versus known-disease results; indel simulations were absent and only one genomic variant was available.The known-disease comparison was 92% for dominant and 96% for recessive cases.Source: Table 5 · p. 38, Evaluating null distribution of genes · p. 38
  • Regulatory-variant simulations evaluated Phen-Gen’s genomic predictor, with table entries reporting the percentage of simulations in each category.The simulations used OMIM-listed disease symptoms and HGMD regulatory disease-causal variants.Source: Table 6 · p. 39
  • The real-patient table compares Phen-Gen’s 1% MAF cutoff with the original study’s removal of dbSNP, in-house, or family-common variants.It covers recessive and X-linked implicated variants in familial data.Source: Table 7 · p. 40
Loading 1502.07829v1…