Mutational signature deciphering - Evaluation report for different NMF algorithms in mutational signature deciphering.

Mutational signature deciphering

Evaluation report for different NMF algorithms in mutational signature deciphering.

8 Mar 2016 | Version 1.0, contributed by Li Xiangchun

Overview

Introduction

Nonnegative matrix factorization (NMF) has been successfully applied to extract mutational signatures underlying tumor development and progression. Alexandrov L.B. and colleagures used brunet algorithm to develop a mutational signature deciphering framework. This framework has been successfully applied to 7042 cancers and extracted more than 20 signatures. However, the performance and accuracy of the brunet algorithm has not been compared with other NMF algorithms such as als, mult, sNMFR and gdclsNMF etc.

Summary

In this report, I extended this mutational signature analysis framework and conducted systematic comparison among different NMF algorithms. The result showed that brunet is worse among other algorithms in terms of reconstruction errors and stabilities. Moreover, brunet fluctuated more frequently as compared with the others. The other four algorithms except brunet have comparable performance with varying computational time spent. The sNMFR took the most time during the evaluation process. Herein, I would recommend users to the als algorithm with respect to its performance and time cost.

Mutational signature deciphering framework: https://github.com/lixiangchun/decipherMutationalSignatures.

Results

Figures & Tables

Table 1. Evalutating the stability of different NMF algorithms on exomic somatic mutations of 544 gastric cancers

Rank	als	brunet	gdclsNMF	mult	sNMFR
1	1	1	1	1	1
2	0.997	0.999	0.999	0.998	0.998
3	0.996	0.613	0.997	0.996	0.997
4	0.987	0.72	0.998	0.99	0.995
5	0.982	0.725	0.998	0.988	0.992
6	0.978	0.906	0.993	0.969	0.988
7	0.891	0.969	0.714	0.792	0.858
8	0.58	0.827	0.628	0.742	0.784
9	0.743	0.734	0.649	0.631	0.573
10	0.8	0.585	0.623	0.461	0.704
11	0.708	0.525	0.755	0.389	0.669
12	0.811	0.512	0.804	0.216	0.723
13	0.733	0.532	0.802	0.31	0.696
14	0.838	0.431	0.793	0.315	0.577
15	0.721	0.376	0.743	0.256	0.533
16	0.636	0.446	0.726	0.156	0.514

Figure 1. Stability comparison among different NMF algorithms.

Table 2. Evalutating the stability of different NMF algorithms on exomic somatic mutations of 544 gastric cancers

Rank	als	brunet	gdclsNMF	mult	sNMFR
1	2391.52	2725.58	2395.03	2391.52	2391.52
2	1665.58	2486.25	1669.83	1666.31	1665.53
3	1222.22	2293.07	1229.19	1221.02	1219.99
4	852.56	1519.6	858.2	852.87	851.89
5	609.42	1237.08	615.44	609.22	607.98
6	474.87	936.32	480.04	474.48	472.06
7	456.48	608.06	618.83	469.46	455.93
8	728.42	580.3	1007.36	449.5	457.56
9	444.92	623.06	874.52	454.74	912.4
10	406.86	565.14	975.22	502.89	485.24
11	449.86	934.81	498.28	595.65	514.47
12	350.26	665.88	460.96	819.98	428.16
13	405.41	748.19	527.29	708.96	434
14	319.53	926.25	454.87	663.93	733.23
15	352.77	1077.31	374.9	751.15	965.44
16	407.35	677.27	370.51	754.5	734.07

Figure 2. Reconstruction error comparison among different NMF algorithms.

Table 3. Evaluating time cost of different NMF algorithms on exomic somatic mutations of 544 gastric cancers. Time present in hour.

Time	als	brunet	gdclsNMF	mult	sNMFR
real	33.1	2.4	32.9	63.2	227.4
user	42.8	9.7	54	191.5	252.5
sys	14.5	1.4	1.5	65.9	4.98

Stability and reconstruction error

Stability comparison among 5 NMF algorithms

Reconstruction error comparison among 5 NMF algorithms

Mutational signature plot (brunet)

Mutational signature 1

Mutational signature 2

Mutational signature 3

Mutational signature 4

Mutational signature 5

Mutational signature 6

Mutational signature 7

Mutational signature plot (als)

Mutational signature 1

Mutational signature 2

Mutational signature 3

Mutational signature 4

Mutational signature 5

Mutational signature 6

Methods & Data

Input

All input files are listed and can be downloaded header: data.tar.gz.

Types file = types
Subtypes file = subtypes
Sample name file = sampleNames
Mutational context by sample file = originalGenomes
Shell script file = als.sh

Running example on BGI server


#!/bin/bash

# Mutation types
typesFile=../../types
            
# Mutational categories for mutation types in `typesFile`
subtypesFile=../../subtypes
            
# Sample names, i.e. column names of input matrix in `originalGenomesFile`
sampleNamesFile=../../sampleNames
            
# Rows represent mutation type and columns sample names. Each cell entry
#+is the number of corresponding mutation types in the sample. Note that
#+the row number of this file must be equal row numbers of both `typesFile`
#+and `subtypesFile`.
originalGenomesFile=../../originalGenomes
            
# removes weak mutation types, i.e. reduces the dimmensions, default value
#+in the original code is 0.01, users can set it to 0.0 if they want to use
#+this framework as consensus clustering for gene expression analysis.
removeWeakMutationTypes=0.01
            
# Algorithm to use, i.e. 1. brunet, 2. mult, 3. als, 4. gdclsNMF, 5. sNMFR
# The original framework uses `brunet`.
algorithm='als'
            
# The min number of signatures
minNumberOfSignature=1
            
# The max number of signatures
maxNumberOfSignature=16
            
# Number of iterations to run nmf for each rank
iterationsPerCore=100
            
# Number of processes to use, if >1 a matlabpool is open for parallel computing.
numberOfWorkers=1
            
## Set to larger number is recommended, e.g. 500000
maxIterPerNmfRun=5000
            
# The following 2 variables must be set appropriately.
DECIPHER_MUTATIONAL_SIGNATURE_PATH=/ifshk1/BC_CANCER/03user/lixiangchun/iCGA/v0.02/decipherMutationalSignatures
MCRroot=/ifshk1/BC_CANCER/03user/lixiangchun/Software/INSTALL/MCR_R2013a/INSTALL/v81
            
time -p bash $DECIPHER_MUTATIONAL_SIGNATURE_PATH/run_decipherMutationalSignatures.sh $MCRroot \
$typesFile $subtypesFile $sampleNamesFile $originalGenomesFile $removeWeakMutationTypes $algorithm \
$minNumberOfSignature $maxNumberOfSignature $iterationsPerCore $numberOfWorkers $maxIterPerNmfRun

References

[1] Li Xiangchun et al., Distinct subtypes of gastric cancer defined by molecular characterization include novel mutational signatures with prognostic capability, Cancer Research (2016)

Made with Nozzle

Total visited times