Package: MsBackendSql
Authors: Johannes Rainer [aut, cre] (https://orcid.org/0000-0002-6977-7147), Chong Tang [ctb], Laurent Gatto [ctb] (https://orcid.org/0000-0002-1520-2268)
Compiled: Fri Oct 27 03:27:25 2023

1 Introduction

The Spectra Bioconductor package provides a flexible and expandable infrastructure for Mass Spectrometry (MS) data. The package supports interchangeable use of different backends that provide additional file support or different ways to store and represent MS data. The MsBackendSql package provides backends to store data from whole MS experiments in SQL databases. The data in such databases can be easily (and efficiently) accessed using Spectra objects that use the MsBackendSql class as an interface to the data in the database. Such Spectra objects have a minimal memory footprint and hence allow analysis of very large data sets even on computers with limited hardware capabilities. For certain operations, the performance of this data representation is superior to that of other low-memory (on-disk) data representations such as Spectra’s MsBackendMzR backend. Finally, the MsBackendSql supports also remote data access to e.g. a central database server hosting several large MS data sets.

2 Installation

The package can be installed with the BiocManager package. To install BiocManager use install.packages("BiocManager") and, after that, BiocManager::install("MsBackendSql") to install this package.

3 Creating and using MsBackendSql SQL databases

MsBackendSql SQL databases can be created either by importing (raw) MS data from MS data files using the createMsBackendSqlDatabase or using the backendInitialize function by providing in addition to the database connection also the full MS data to import as a DataFrame. In the first example we use the createMsBackendSqlDatabase function which takes a connection to an (empty) database and the names of the files from which the data should be imported as input parameters creates all necessary database tables and stores the full data into the database. Below we create an empty SQLite database (in a temporary file) and fill that with MS data from two mzML files (from the r Biocpkg("msdata") package).

library(RSQLite)

dbfile <- tempfile()
con <- dbConnect(SQLite(), dbfile)

library(MsBackendSql)
fls <- dir(system.file("sciex", package = "msdata"), full.names = TRUE)
createMsBackendSqlDatabase(con, fls)

By default the m/z and intensity values are stored as BLOB data types in the database. This has advantages on the performance to extract peaks data from the database but would for example not allow to filter peaks by m/z values directly in the database. As an alternative it is also possible to the individual m/z and intensity values in separate rows of the database table. This long table format results however in considerably larger databases (with potentially poorer performance). Note also that the code and backend is optimized for MySQL/MariaDB databases by taking advantage of table partitioning and specialized table storage options. Any other SQL database server is however also supported (also portable, self-contained SQLite databases).

The MsBackendSql package provides two backends to interact with such databases: the (default) MsBackendSql class and the MsBackendOfflineSql, that inherits all properties and functions from the former, but which does not store the connection to the database within the object but connects (and disconnects) to (and from) the database in each function call. This allows to use the latter also for parallel processing setups or to save/load the object (e.g. using save and saveRDS). Thus, for most applications the MsBackendOfflineSql might be used as the preferred backend to SQL databases.

To access the data in the database we create below a Spectra object providing the connection to the database in the constructor call and specifying to use the MsBackendSql as backend using the source parameter.

sps <- Spectra(con, source = MsBackendSql())
sps
## MSn data (Spectra) with 1862 spectra in a MsBackendSql backend:
##        msLevel precursorMz  polarity
##      <integer>   <numeric> <integer>
## 1            1          NA         1
## 2            1          NA         1
## 3            1          NA         1
## 4            1          NA         1
## 5            1          NA         1
## ...        ...         ...       ...
## 1858         1          NA         1
## 1859         1          NA         1
## 1860         1          NA         1
## 1861         1          NA         1
## 1862         1          NA         1
##  ... 34 more variables/columns.
##  Use  'spectraVariables' to list all of them.
## Database: /tmp/RtmpcoutSc/file11b102d077dab

As an alternative the MsBackendOfflineSql backend could be used instead, which supports serializing the data to disk and allows, if supported by the SQL database, also parallel processing. Thus, for most use cases the MsBackendOfflineSql should be used instead. See further below for more information on that backend..

Spectra objects allow also to change the backend to any other backend (extending MsBackend) using the setBackend function. Below we use this function to first load all data into memory by changing from the MsBackendSql to a MsBackendMemory.

sps_mem <- setBackend(sps, MsBackendMemory())
sps_mem
## MSn data (Spectra) with 1862 spectra in a MsBackendMemory backend:
##        msLevel     rtime scanIndex
##      <integer> <numeric> <integer>
## 1            1     0.280         1
## 2            1     0.559         2
## 3            1     0.838         3
## 4            1     1.117         4
## 5            1     1.396         5
## ...        ...       ...       ...
## 1858         1   258.636       927
## 1859         1   258.915       928
## 1860         1   259.194       929
## 1861         1   259.473       930
## 1862         1   259.752       931
##  ... 34 more variables/columns.
## Processing:
##  Switch backend from MsBackendSql to MsBackendMemory [Fri Oct 27 03:27:36 2023]

With this function it is also possible to change from any backend to a MsBackendSql in which case a new database is created and all data from the originating backend is stored in this database. To change the backend to an MsBackendOfflineSql we need to provide the connection information to the SQL database as additional parameters. These are the same as we would need to connect to the database through a dbConnect call and includes the database driver to be used (parameter drv) as well as additional parameters such as the database name and eventually the user name, host etc (see ?dbConnect for more information). In the simple example below we store the data into a SQLite database and thus only need to provide the database name, which corresponds SQLite database file. In our example we store the data into a temporary file. Importantly, we also need to disable parallel processing by specifying BPPARAM = SerialParam() since (most) SQL databases don’t provide parallel data insertion.

sps2 <- setBackend(sps_mem, MsBackendOfflineSql(), drv = SQLite(),
                   dbname = tempfile())
## Warning in .create_from_spectra_data(dbcon, data = data, ...): Replacing
## original column "spectrum_id_"
sps2
## MSn data (Spectra) with 1862 spectra in a MsBackendOfflineSql backend:
##        msLevel precursorMz  polarity
##      <integer>   <numeric> <integer>
## 1            1          NA         1
## 2            1          NA         1
## 3            1          NA         1
## 4            1          NA         1
## 5            1          NA         1
## ...        ...         ...       ...
## 1858         1          NA         1
## 1859         1          NA         1
## 1860         1          NA         1
## 1861         1          NA         1
## 1862         1          NA         1
##  ... 34 more variables/columns.
##  Use  'spectraVariables' to list all of them.
## Database: /private/tmp/RtmpcoutSc/file11b1069acb7de
## Processing:
##  Switch backend from MsBackendSql to MsBackendMemory [Fri Oct 27 03:27:36 2023]
##  Switch backend from MsBackendMemory to MsBackendOfflineSql [Fri Oct 27 03:27:37 2023]

Similar to any other Spectra object we can retrieve the available spectra variables using the spectraVariables function.

spectraVariables(sps)
##  [1] "msLevel"                  "rtime"                   
##  [3] "acquisitionNum"           "scanIndex"               
##  [5] "dataStorage"              "dataOrigin"              
##  [7] "centroided"               "smoothed"                
##  [9] "polarity"                 "precScanNum"             
## [11] "precursorMz"              "precursorIntensity"      
## [13] "precursorCharge"          "collisionEnergy"         
## [15] "isolationWindowLowerMz"   "isolationWindowTargetMz" 
## [17] "isolationWindowUpperMz"   "peaksCount"              
## [19] "totIonCurrent"            "basePeakMZ"              
## [21] "basePeakIntensity"        "ionisationEnergy"        
## [23] "lowMZ"                    "highMZ"                  
## [25] "mergedScan"               "mergedResultScanNum"     
## [27] "mergedResultStartScanNum" "mergedResultEndScanNum"  
## [29] "injectionTime"            "filterString"            
## [31] "spectrumId"               "ionMobilityDriftTime"    
## [33] "scanWindowLowerLimit"     "scanWindowUpperLimit"    
## [35] "spectrum_id_"

The MS peak data can be accessed using either the mz, intensity or peaksData functions. Below we extract the peaks matrix of the 5th spectrum and display the first 6 rows.

peaksData(sps)[[5]] |>
head()
##            mz intensity
## [1,] 105.0347         0
## [2,] 105.0362       164
## [3,] 105.0376         0
## [4,] 105.0391         0
## [5,] 105.0405       328
## [6,] 105.0420         0

All data (peaks data or spectra variables) are always retrieved on the fly from the database resulting thus in a minimal memory footprint for the Spectra object.

print(object.size(sps), units = "KB")
## 91.4 Kb

The backend supports also adding additional spectra variables or changing their values. Below we add 10 seconds to the retention time of each spectrum.

sps$rtime <- sps$rtime + 10

Such operations do however not change the data in the database (which is always considered read-only) but are cached locally within the backend object (in memory). The size in memory of the object is thus higher after changing that spectra variable.

print(object.size(sps), units = "KB")
## 106 Kb

Such $<- operations can also be used to cache spectra variables (temporarily) in memory which can eventually improve performance. Below we test the time it takes to extract the MS level from each spectrum from the database, then cache the MS levels in memory using $msLevel <- and test the timing to extract these cached variable.

system.time(msLevel(sps))
##    user  system elapsed 
##   0.010   0.001   0.012
sps$msLevel <- msLevel(sps)
system.time(msLevel(sps))
##    user  system elapsed 
##   0.004   0.000   0.004

We can also use the reset function to reset the data to its original state (this will cause any local spectra variables to be deleted and the backend to be initialized with the original data in the database).

sps <- reset(sps)

To use the MsBackendOfflineSql backend we need to provide all information required to connect to the database along with the database driver to the Spectra function. Which parameters are required to connect to the database depends on the SQL database and the used driver. In our example the data is stored in a SQLite database, hence we use the SQLite() database driver and only need to provide the database name with the dbname parameter. For a MySQL/MariaDB database we would use the MariaDB() driver and would have to provide the database name, user name, password as well as the host name and port through which the database is accessible.

sps_off <- Spectra(dbfile, drv = SQLite(),
                   source = MsBackendOfflineSql())
sps_off
## MSn data (Spectra) with 1862 spectra in a MsBackendOfflineSql backend:
##        msLevel precursorMz  polarity
##      <integer>   <numeric> <integer>
## 1            1          NA         1
## 2            1          NA         1
## 3            1          NA         1
## 4            1          NA         1
## 5            1          NA         1
## ...        ...         ...       ...
## 1858         1          NA         1
## 1859         1          NA         1
## 1860         1          NA         1
## 1861         1          NA         1
## 1862         1          NA         1
##  ... 34 more variables/columns.
##  Use  'spectraVariables' to list all of them.
## Database: /private/tmp/RtmpcoutSc/file11b102d077dab

This backend provides the exact same functionality than MsBackendSql with the difference that the connection to the database is opened and closed for each function call. While this leads to a slightly lower performance, it allows to to serialize the object (i.e. save/load the object to/from disk) and to use it (and hence the Spectra object) also in a parallel processing setup. In contrast, for the MsBackendSql parallel processing is disabled since it is not possible to share the active backend connection within the object across different parallel processes.

Below we compare the performance of the two backends. The performance difference is the result from opening and closing the database connection for each call. Note that this will also depend on the SQL server that is being used. For SQLite databases there is almost no overhead.

library(microbenchmark)
microbenchmark(msLevel(sps), msLevel(sps_off))
## Warning in microbenchmark(msLevel(sps), msLevel(sps_off)): less accurate
## nanosecond times to avoid potential integer overflows
## Unit: milliseconds
##              expr       min       lq     mean   median       uq      max neval
##      msLevel(sps)  9.993463 11.34527 12.22585 11.98268 12.80209 21.97235   100
##  msLevel(sps_off) 13.026520 15.47904 16.67661 16.28370 17.60114 24.42415   100
##  cld
##   a 
##    b

4 Performance comparison with other backends

The need to retrieve any spectra data on-the-fly from the database will have an impact on the performance of data access function of Spectra objects using the MsBackendSql backends. To evaluate its impact we next compare the performance of the MsBackendSql to other Spectra backends, specifically, the MsBackendMzR which is the default backend to read and represent raw MS data, and the MsBackendMemory backend that keeps all MS data in memory (and is thus not suggested for larger MS experiments). Similar to the MsBackendMzR, also the MsBackendSql keeps only a limited amount of data in memory. These on-disk backends need thus to retrieve spectra and MS peaks data on-the-fly from either the original raw data files (in the case of the MsBackendMzR) or from the SQL database (in the case of the MsBackendSql). The in-memory backend MsBackendMemory is supposed to provide the fastest data access since all data is kept in memory.

Below we thus create Spectra objects from the same data but using the different backends.

sps <- Spectra(con, source = MsBackendSql())
sps_mzr <- Spectra(fls, source = MsBackendMzR())
sps_im <- setBackend(sps_mzr, backend = MsBackendMemory())

At first we compare the memory footprint of the 3 backends.

print(object.size(sps), units = "KB")
## 91.4 Kb
print(object.size(sps_mzr), units = "KB")
## 386.5 Kb
print(object.size(sps_im), units = "KB")
## 54494.3 Kb

The MsBackendSql has the lowest memory footprint of all 3 backends because it does not keep any data in memory. The MsBackendMzR keeps all spectra variables, except the MS peaks data, in memory and has thus a larger size. The MsBackendMemory keeps all data (including the MS peaks data) in memory and has thus the largest size in memory.

Next we compare the performance to extract the MS level for each spectrum from the 4 different Spectra objects.

library(microbenchmark)
microbenchmark(msLevel(sps),
               msLevel(sps_mzr),
               msLevel(sps_im))
## Unit: microseconds
##              expr       min         lq        mean     median         uq
##      msLevel(sps) 10644.666 11685.4715 12643.63535 12289.6680 12808.0105
##  msLevel(sps_mzr)   533.164   615.1640   691.07058   661.8835   717.0900
##   msLevel(sps_im)    16.441    23.9645    42.66706    42.1275    59.2245
##        max neval cld
##  34345.413   100 a  
##   1899.079   100  b 
##    139.441   100   c

Extracting MS levels is thus slowest for the MsBackendSql, which is not surprising because both other backends keep this data in memory while the MsBackendSql needs to retrieve it from the database.

We next compare the performance to access the full peaks data from each Spectra object.

microbenchmark(peaksData(sps, BPPARAM = SerialParam()),
               peaksData(sps_mzr, BPPARAM = SerialParam()),
               peaksData(sps_im, BPPARAM = SerialParam()), times = 10)
## Unit: milliseconds
##                                         expr         min          lq
##      peaksData(sps, BPPARAM = SerialParam())  248.091820  253.627763
##  peaksData(sps_mzr, BPPARAM = SerialParam()) 1267.011561 1289.551147
##   peaksData(sps_im, BPPARAM = SerialParam())    2.676726    2.953968
##         mean      median          uq        max neval cld
##   364.906453  261.955232  301.315314  774.58832    10 a  
##  1391.216051 1316.951509 1403.519585 1931.69745    10  b 
##     5.240899    3.813164    4.037475   20.08102    10   c

As expected, the MsBackendMemory has the fasted access to the full peaks data. The MsBackendSql outperforms however the MsBackendMzR providing faster access to the m/z and intensity values.

Performance can be improved for the MsBackendMzR using parallel processing. Note that the MsBackendSql does not support parallel processing and thus parallel processing is (silently) disabled in functions such as peaksData.

m2 <- MulticoreParam(2)
microbenchmark(peaksData(sps, BPPARAM = m2),
               peaksData(sps_mzr, BPPARAM = m2),
               peaksData(sps_im, BPPARAM = m2), times = 10)
## Unit: microseconds
##                              expr        min         lq        mean      median
##      peaksData(sps, BPPARAM = m2) 274568.267 279636.113  432217.174  299449.711
##  peaksData(sps_mzr, BPPARAM = m2) 899201.586 946905.086 1322345.501 1037981.748
##   peaksData(sps_im, BPPARAM = m2)    684.741    883.427    1658.999    1503.634
##           uq         max neval cld
##   380529.897  978873.893    10  a 
##  1159497.876 3085820.146    10   b
##     1908.878    3849.326    10  a

We next compare the performance of subsetting operations.

microbenchmark(filterRt(sps, rt = c(50, 100)),
               filterRt(sps_mzr, rt = c(50, 100)),
               filterRt(sps_im, rt = c(50, 100)))
## Unit: microseconds
##                                expr      min        lq      mean   median
##      filterRt(sps, rt = c(50, 100)) 4617.502 5402.4060 6234.8081 5867.654
##  filterRt(sps_mzr, rt = c(50, 100)) 3351.791 3772.8815 4302.7581 4115.416
##   filterRt(sps_im, rt = c(50, 100))  700.895  869.4255  993.3919  969.363
##        uq       max neval cld
##  6742.983 13255.464   100 a  
##  4455.962 16952.926   100  b 
##  1030.822  3021.536   100   c

The two on-disk backends MsBackendSql and MsBackendMzR show a comparable performance for this operation. This filtering does involves access to a spectra variables (the retention time in this case) which, for the MsBackendSql needs first to be retrieved from the backend. The MsBackendSql backend allows however also to cache spectra variables (i.e. they are stored within the MsBackendSql object). Any access to such cached spectra variables can eventually be faster because no dedicated SQL query is needed.

To evaluate the performance of a pure subsetting operation we first define the indices of 10 random spectra and subset the Spectra objects to these.

idx <- sample(seq_along(sps), 10)
microbenchmark(sps[idx],
               sps_mzr[idx],
               sps_im[idx])
## Unit: microseconds
##          expr     min       lq      mean    median        uq      max neval cld
##      sps[idx] 178.350 191.4700  256.6670  207.8495  237.4515 3928.333   100 a  
##  sps_mzr[idx] 955.710 999.2725 1061.0472 1035.6395 1093.4700 1567.799   100  b 
##   sps_im[idx] 307.828 321.3375  381.0417  337.0200  359.7955 3807.752   100   c

Here the MsBackendSql outperforms the other backends because it does not keep any data in memory and hence does not need to subset these. The two other backends need to subset the data they keep in memory which is in both cases a data frame with either a reduced set of spectra variables or the full MS data.

At last we compare also the extraction of the peaks data from the such subset Spectra objects.

sps_10 <- sps[idx]
sps_mzr_10 <- sps_mzr[idx]
sps_im_10 <- sps_im[idx]

microbenchmark(peaksData(sps_10),
               peaksData(sps_mzr_10),
               peaksData(sps_im_10),
               times = 10)
## Unit: microseconds
##                   expr       min         lq       mean     median         uq
##      peaksData(sps_10)  5545.988   5748.118   9700.723   9143.205  12887.571
##  peaksData(sps_mzr_10) 95309.953 100864.592 102340.129 102949.340 104517.159
##   peaksData(sps_im_10)   510.860    518.076   1445.172    714.466   2597.309
##         max neval cld
##   17383.303    10 a  
##  107143.291    10  b 
##    4520.496    10   c

The MsBackendSql outperforms the MsBackendMzR while, not unexpectedly, the MsBackendMemory provides fasted access.

5 Other properties of the MsBackendSql

The MsBackendSql backend does not support parallel processing since the database connection can not be shared across the different (parallel) processes. Thus, all methods on Spectra objects that use a MsBackendSql will automatically (and silently) disable parallel processing even if a dedicated parallel processing setup was passed along with the BPPARAM method.

Some functions on Spectra objects require to load the MS peak data (i.e., m/z and intensity values) into memory. For very large data sets (or computers with limited hardware resources) such function calls can cause out-of-memory errors. One example is the lengths function that determines the number of peaks per spectrum by loading the peak matrix first into memory. Such functions should ideally be called using the peaksapply function with parameter chunkSize (e.g., peaksapply(sps, lengths, chunkSize = 5000L)). Instead of processing the full data set, the data will be first split into chunks of size chunkSize that are stepwise processed. Hence, only data from chunkSize spectra is loaded into memory in one iteration.

6 Summary

The MsBackendSql provides an MS data representations and storage mode with a minimal memory footprint (in R) that is still comparably efficient for standard processing and subsetting operations. This backend is specifically useful for very large MS data sets, that could even be hosted on remote (MySQL/MariaDB) servers. A potential use case for this backend could thus be to set up a central storage place for MS experiments with data analysts connecting remotely to this server to perform initial data exploration and filtering. After subsetting to a smaller data set of interest, users could then retrieve/download this data by changing the backend to e.g. a MsBackendMemory, which would result in a download of the full data to the user computer’s memory.

7 Session information

sessionInfo()
## R version 4.3.1 (2023-06-16)
## Platform: aarch64-apple-darwin20 (64-bit)
## Running under: macOS Ventura 13.6.1
## 
## Matrix products: default
## BLAS:   /Library/Frameworks/R.framework/Versions/4.3-arm64/Resources/lib/libRblas.0.dylib 
## LAPACK: /Library/Frameworks/R.framework/Versions/4.3-arm64/Resources/lib/libRlapack.dylib;  LAPACK version 3.11.0
## 
## locale:
## [1] C/en_US.UTF-8/en_US.UTF-8/C/en_US.UTF-8/en_US.UTF-8
## 
## time zone: America/New_York
## tzcode source: internal
## 
## attached base packages:
## [1] stats4    stats     graphics  grDevices utils     datasets  methods  
## [8] base     
## 
## other attached packages:
## [1] microbenchmark_1.4.10 RSQLite_2.3.1         MsBackendSql_1.2.0   
## [4] Spectra_1.12.0        ProtGenerics_1.34.0   BiocParallel_1.36.0  
## [7] S4Vectors_0.40.1      BiocGenerics_0.48.0   BiocStyle_2.30.0     
## 
## loaded via a namespace (and not attached):
##  [1] sandwich_3.0-2         sass_0.4.6             MsCoreUtils_1.14.0    
##  [4] lattice_0.21-8         hms_1.1.3              digest_0.6.33         
##  [7] grid_4.3.1             evaluate_0.21          bookdown_0.34         
## [10] mvtnorm_1.2-2          fastmap_1.1.1          blob_1.2.4            
## [13] Matrix_1.6-0           jsonlite_1.8.7         progress_1.2.2        
## [16] mzR_2.36.0             DBI_1.1.3              survival_3.5-5        
## [19] multcomp_1.4-25        BiocManager_1.30.22    TH.data_1.1-2         
## [22] codetools_0.2-19       jquerylib_0.1.4        cli_3.6.1             
## [25] rlang_1.1.1            crayon_1.5.2           Biobase_2.62.0        
## [28] splines_4.3.1          bit64_4.0.5            cachem_1.0.8          
## [31] yaml_2.3.7             tools_4.3.1            parallel_4.3.1        
## [34] memoise_2.0.1          ncdf4_1.21             vctrs_0.6.3           
## [37] R6_2.5.1               zoo_1.8-12             lifecycle_1.0.3       
## [40] fs_1.6.2               IRanges_2.36.0         bit_4.0.5             
## [43] clue_0.3-64            MASS_7.3-60            cluster_2.1.4         
## [46] pkgconfig_2.0.3        bslib_0.5.0            data.table_1.14.8     
## [49] Rcpp_1.0.11            xfun_0.39              knitr_1.43            
## [52] htmltools_0.5.5        rmarkdown_2.23         compiler_4.3.1        
## [55] prettyunits_1.1.1      MetaboCoreUtils_1.10.0