PhenoSS: Phenotype semantic similarity-based approach for rare disease prediction and patient clustering
PhenoSS is an effective algorithm that makes disease prediction and performs patient clustering based on HPO concepts. PhenoSS uses the Gaussian copula technique by modeling the marginal prevalence of each HPO term for each disease and utilizes a multivariate normal distribution to link them together to account for term correlations. We utilized the OARD (open annotations for rare diseases) API for inferring the frequency of HPO terms in a diverse range of rare diseases. PhenoSS can calculate the phenotype similarity between any two patients for finding similar patients or clustering purposes, or between one patient and any candidate diseases for diagnosis support.
The toolkit is implemented in Python.
The file hpo_list contains the synthetic data for three randomly generated patients labeled 0_10, 1_10, 2_10.
0_10 HP_0004370;HP_0000280;HP_0002835;HP_0005274;HP_0000158;HP_0011470;HP_0001417;HP_0001270;HP_0008872;HP_0002015;HP_0000750;HP_0000157;
1_10 HP_0000483;HP_0002307;HP_0001090;HP_0001572;HP_0002342;HP_0011343;HP_0008760;HP_0001061;HP_0001249;HP_0000574;HP_0001417;HP_0002020;HP_0012810;HP_000
0540;HP_0001350;HP_0001270;HP_0002574;HP_0011231;HP_0000750;HP_0002155;HP_0000431;HP_0000718;
2_10 HP_0002194;HP_0001263;HP_0001684;HP_0001417;HP_0001270;HP_0001249;HP_0001670;HP_0001667;HP_0001629;HP_0001639;HP_0002474;HP_0010863;HP_0000750;
Using the following argument, we can calculate the similarity scores between patient 1_10 and each of the patients in the hpo_list. The first input argument is the file that contains the patient IDs and the HPO terms. The second input argument is the patient ID we are interested in.
python getdiff_one_patient.py hpo_list 1_10
You are recommended to submit one job for each patient to perform computing simultaneously.
The outputs of the argument can be found in the file 1_10_sim.
1_10 0_10 3.741497748992082
1_10 1_10 8.089540938721928
1_10 2_10 2.451702686307226
The following code will summarize the scores for all the patients and form the similarity matrix.
perl get_sim_mat.pl hpo_list
We can then perform the quality check. filter.pl automatically detects missing values in the similarity matrix and selects te largest subset of the patients such that there is no missing similarity scores for these patients and forms the similarity score matrix.
perl filter.pl
The similarity score matrix will then be saved into the file sim_mat_filter.
To better understand the results, we can perform the hierarchical clustering and plot the dendrograms in R.
library(dplyr)
x <- read.table("sim_mat_filter")
x$V3 <- 1/x$V3
numpat <- sqrt(dim(x)[1])
pat_mat <- matrix(x$V3, nrow = numpat)
patid <- x$V2[1:numpat]
colnames(pat_mat) <- patid
rownames(pat_mat) <- patid
hc <- hclust(dist(pat_mat), method = "ward.D2")
dend <- as.dendrogram(hc)
#groupCodes <- c(rep("NA10", 63), rep("NA15", 64))
#colorCodes <- c(NA10="blue", NA15="green")
#labels_colors(dend) <- colorCodes[groupCodes][order.dendrogram(dend)]
plot(dend)
PhenoSS extracts the diseases/phenotype frequencies from the Open Annotations for Rare Diseases (OARD) Database. It takes in HPO terms of a list of patients and outputs the ranks of possible underlying diseases.
Below is a sample input file:
P1 HP_0012759;HP_0000750;HP_0100022;HP_0000707;
P2 HP_0001270;HP_0012758;HP_0002066;HP_0011443;
P3 HP_0012758;HP_0002167;HP_0012638;HP_0000707;
To run PhenoSS, use the following command:
python phenoSS.py inputFile outputFile
The ranks will be stored in the output file.