Contents

Note: the most recent version of this tutorial can be found here and a short overview slide show here.

1 Introduction

The R package eiR provides an index for chemical compound databases allowing one to quickly find similar compounds in a very large database. To create this index, r reference compounds are selected to represent the database. Then each compound in the database is embedded into d-dimensional space based on their distance to each reference compound. This requires time linear in the size of the database, but only needs to be done once for a database. Within this space, Locality Sensitive Hashing (LSH) is employed to allow sub-linear time nearest neighbor lookups (Dong, Wang, Charikar, et al. 2008; Dong, Wang, Josephson, et al. 2008) that nearest neighbors can be found without doing a linear scan through the entire compound database. Additional compounds can be added to the database in time linear to the number of new compounds. No time is spend processing existing compounds, as long as the set of reference compounds remains the same. Given the ability to quickly find nearest neighbors, this method enables fast clustering with the Jarvis-Pattrick algorithm as well (Jarvis and Patrick 1973). For details on the whole process see (Cao, Jiang, and Girke 2010).

This library uses an SQL back-end (SQLite by default) to store chemical compound definitions, either in SDF or SMILE format, as well as descriptors. Several different kinds of descriptors can be stored for each compound, for example, one could compute and store atom-pair and fingerprint descriptors for each compound. The SQLite database, if used, is stored in a directory called data. The eiInit function is used to create a new database, it can import data from SDF or SMILE formated files, or an SDFset object.

Once a database has been created, an embedding must also be created (Dimitris K. Agrafiotis, Dmitrii N. Rassokhin, Victor S. Lobanov 2001). In this step the reference compounds are chosen and each compound is embedded into the new space. This step creates a new directory called run-r-d, where r and d are the corresponding values. This is the most costly step of the process and is handled by the eiMakeDb function. This step can be parallelized by providing a SNOW cluster to the eiMakeDb function.

Given an embedded database, queries can be run against it with the eiQuery function. Additional compounds can also be added to an existing database and embedding using eiAdd. Performance tests can be run using the eiPerformanceTest function, and Jarvis-Patrick clustering can be done with the eiCluster function.

eiR also provides some mechanisms to allow the user to extend the set of descriptor formats used and to define new distance functions. See Section Customization for details.

2 Initialization

An initial compound database must be created with the following command:

library(eiR) 
data(sdfsample) 
eiInit(sdfsample[1:99],priorityFn=randomPriorities) 
## Loading required package: RSQLite
## [1] "createing db"
## loaded 99 compounds
## [1] "99  loaded by eiInit"
## setting priorities for 4 groups.
##  [1] 295 211 264 272 258 273 228 261 206 299 288 203 220 278 286 244 257 263 270 274 249 239 214 226
## [25] 202 298 296 242 262 235 291 256 207 225 230 280 283 210 289 248 259 290 246 293 265 271 234 236
## [49] 241 245 253 205 240 217 294 223 255 221 281 251 238 268 252 212 247 277 266 250 269 282 254 204
## [73] 260 285 216 227 215 229 279 267 276 243 237 275 218 224 233 284 297 292 287 208 232 231 213 201
## [97] 219 209 222

EiInit can take either an SDFset, or a filename. If a filename is given it must be in either SDF or SMILE format and the format must be specified with the format paramter. It might complain if your SDF file does not follow the SDF specification. If this happens, you can create an SDFset with the read.SDFset command and then use that instead of the filename.

Descriptors will also be computed at this time. The default descriptor type is atompair. Other types can be used by setting the descriptorType parameter. Currently available types are “ap” for atompair, and “fp” for fingerprint. The set of available descriptors can be extended, see Section Customization. EiInit will create a folder called ‘data’. Commands should always be executed in the folder containing this directory (ie, the parent directory of “data”), or else specify the location of that directory with the dir option.

EiInit also has some options to support loading data in parallel. This is only possible if the SQL database being used supports parallel writes. For now, this means only PostgreSQL. To use this feature, the inputs parameter must be set to an array of filenames, cl set to a SNOW cluster, and connSource set to a function which will return a new database connection when called with no parameters. Then each file will be loaded on a different node of the given cluster, in parallel. The connSource function must actually create a new connection on each call, simply returning the same reference to an existing connection will not work. This is because that function will be called on potentially different machines which will have to establish their own connection to the database.

Compounds which already exist in the database will be skipped over, so it is safe to re-run eiInit on an input file which has already been partially loaded. By default it discovers duplicates by comparing the entire compound definition. However, if you want two compounds with the same name to be considered equal even if the definition is different, you can set updateByName to be true. In this mode, if a compound being loaded is found to exist in the database already, by name, but has a different definition, the compound will be updated with the new definition and any associated descriptors and/or features will be re-computed.

eiInit will return a list of compound id numbers which represent the compounds just inserted. These numbers can be used to issue queries later.

3 Creating a Searchable Database

In this step the compounds in the data directory will be embedded in another space which allows for more efficient searching. The main two parameters are refs and d. refs can be either a list of compound ids to use as references, or else an integer value indicating the number of references to use, which will then be randomly chosen. We will use r to represent the number of reference compounds. d is the dimension of the embedding space. We have found in practice that setting d to around 100 works well. r should be large enough to “represent” the full compound database. Since creating a database is the longest running step, a SNOW cluster can be provided to parallelize the task. A conSource function is also required in this case, as described in Section Initialization.

To help tune these values, eiMakeDb will pick numSamples non-reference samples which can later be used by the eiPerformanceTest function.

eiMakdDb does its job in a job folder, named after the number of reference compounds and the number of embedding dimensions. For example, using 300 reference compounds to generate a 100-dimensional embedding (r=300,d=100) will result in a job folder called run-300-100. The embedding result is the file matrix.r.d. In the above example, the output would be run-300-100/matrix.300.100.

Since more than one type of descriptor can be stored for each compound, the desired descriptor type must be given to this function with the descriptorType parameter. The default value is “ap”, for atompair. You can also specify a custom distance function that must be able to take two descriptors in the format specified and return a distance value. The default distance method used is 1-Tanimoto(descriptor1,descriptor2).

The return value is called the Run Id. This value is needed by other functions to identify the data set and embedding parameters to use.

r<- 60 
d<- 40 
runId <- eiMakeDb(r,d) 
## found 95 un-embedded descriptors for runId 1
## Regenerating matrix file...
## Regenerating matrix file...

4 Queries

Queries can be given in several formats, defined by the format parameter. The default format is “sdf”. The queries parameter can be either an SDF file or an SDFset under this format. Other valid values for format are “name” and “compound_id”. Under these two formats the queries parameter is expected to be a list of compound names (as returned by sdfid on an SDFset), or a list of compound id numbers from the database, such as what is returned by the eiInit function.

The runId parameter is required to determine which embedded database to use. As with eiMakeDb, the distance parameter may be given if desired, it will default to the Tanimoto Coefficient otherwise. Finally, the parameter K is the number of results that will be returned. In some cases, particularly if K is small, you may need to set it to a larger value and then trim down the result set yourself. This is because LSH is not an exact algorithm. Internally, it actually searches for xK neighbors, where x is referred to as the expansion ratio, generally set to 2. This allows it to pick the best K matches, according to the true distance function, out of a larger set of candidates. When K is small though, sometimes that expansion ratio is not quite enough.

Note also that this function returns distance values and not similarities. Similarities can be computed by setting asSimilarity to TRUE. This assumes that whatever distance function is currently in use returns values between 0 and 1. This is true for the default distance functions for “ap” and “fp” descriptors.

Then you can perform a query as follows:

#find compounds similar to each query
result=eiQuery(runId,sdfsample[45],K=10,asSimilarity=TRUE)
## creating annoy index
## loading matrix file
## done. num items in index: 95
## found 1 queries
print(result)
##     query target similarity target_ids
## 1  650046 650046  1.0000000        245
## 2  650046 650054  0.5133418        251
## 3  650046 650011  0.4651163        211
## 4  650046 650092  0.3922652        286
## 5  650046 650069  0.3897316        264
## 6  650046 650008  0.3256659        208
## 7  650046 650100  0.3027888        294
## 8  650046 650080  0.2791878        275
## 9  650046 650002  0.2322581        202
## 10 650046 650067  0.2299197        262
#Compare to traditional similarity search: 
data(apset)
print(cmp.search(apset,apset[45],type=3,cutoff=4,quiet=TRUE))
##   index    cid    scores
## 1    45 650046 1.0000000
## 2    51 650054 0.5133418
## 3    11 650011 0.4651163
## 4    86 650092 0.3922652
cid(sdfsample)=sdfid(sdfsample)
plot(sdfsample[result$target[1:4]],regenCoords=TRUE,print=FALSE)

The result will be a data frame with four columns. The first is a query id, the second is a target, or hit, id, the third is the distance (or similarity if asSimilary was true) between the query and the target, and the fourth column is the compound id number of the target. Lsh parameters can be passed in as well, see Section Performance Tests for more details.

5 Adding New Compounds

New Compounds can be added to an existing database, however, the reference compounds cannot be changed. To add new compounds, use the eiAdd function. This function is very similar to the eiQuery function, except instead of a queries parameter, there is an additions parameter, defining the compounds to be added. The format of the value of this parameter works the same as in the eiQuery function. For example, to add one compound from an SDFset you would do:

 eiAdd(runId,sdfsample[100]) 
## loaded 1 compounds
## [1] "1  loaded by eiInit"
## setting priorities for 0 groups.
## found 1 un-embedded descriptors for runId 1
## Regenerating matrix file...
## [1] 300

The returned value is a list of the compounds ids that were just added. This function will also write out a new matrix file in the run directory. For large databases, this can take a significant amount of time.

6 Performance Tests

The eiPerforamceTest function will run several tests using some sample data to evaluate the performance of the current embedding. It takes the usual runId parameter, as well as on optional distance function. It also takes several LSH parameters, though the defaults are usually fine. To evaluate the performance you can run:

 eiPerformanceTest(runId,K=22) 
## computing embedded distances on test quries
## writing top 50000 to file ./run-60-40/eucsearch.60-40
## A connection with                                         
## description "./run-60-40/eucsearch.60-40"
## class       "file"                       
## mode        "w"                          
## text        "text"                       
## opened      "opened"                     
## can read    "no"                         
## can write   "yes"                        
## [1] "getting test query ids"
## writing top 50000 to file ./data/chemical-search.results
## A connection with                                            
## description "./data/chemical-search.results"
## class       "file"                          
## mode        "w"                             
## text        "text"                          
## opened      "opened"                        
## can read    "no"                            
## can write   "yes"
## average embedding performance: 0.488017941949345
## creating annoy index
## loading matrix file
## done. num items in index: 96
## found 9 queries
## average indexing performance: 0.865712299059789
## [1] 0.9735127 0.9120858 0.8423713 0.7517938 0.8968346 0.9418114 0.8600895 0.7584060 0.8545056

This will perform two different tests. The first tests the embedding results in similarity search. The way this works is by approximating 1,000 random similarity searches (determined by data/test_queries.iddb) by nearest neighbor search using the coordinates from the embedding results. The search results are then compared to the reference search results (chemical-search.results.gz).

The comparison results are summarized in two types of files. The first type lists the recall for different k values, k being the number of numbers to retrieve. These files are named as “recall-ratio-k”. For example, if the recall is 70% for top-100 compound search (70 of the 100 results are among the real top-100 compounds) then the value at line 100 is 0.7. Several relaxation ratios are used, each generating a file in this form. For instance, recall.ratio-10 is the file listing the recalls when relaxation ratio is 10. The other file, recall.csv, lists recalls of different relaxation ratios in one file by limiting to selected k value. In this CSV file, the rows correspond to different relaxation ratios, and the columns are different k values. You will be able to pick an appropriate relaxation ratio for the k values you are interested in.

The second test measures the performance of the Locality Sensitive Hash (LSH). The results for lsh-assisted search will be in run-r-d/indexed.performance. It’s a 1,000-line file of recall values. Each line corresponds to one test query. LSH search performance is highly sensitive to your LSH parameters (K, W, M, L, T). The default parameters are listed in the man page for eiPerformanceTest. When you have your embedding result in a matrix file, you should follow instruction on http://lshkit.sourceforge.net/dd/d2a/mplsh-tune_8cpp.html to find the best values for these parameters.

7 Clustering

Compounds can be clustered in near linear time using the Jarvis-Patrick clustering algorithm by taking advantage of the near constant time nearest neighbor lookup provided by the LSH index. Clustering is done with the eiCluster function. It takes a runId to identify the data set and embedding, and two parameters for the Jarvis-Patrick algorithm: K is the number of neighbors to fetch for each compound, and minNbrs is the minimum number of neighbors two compounds must have in common in order to be joined into the same cluster. A cutoff value can also be given to set a maximum distance between neighbors. Any two compounds farther apart than this cutoff will never be considered neighbors. This parameter is helpful in preventing compounds which are very different from almost every other compound from being considered similar to other distant compounds simply because they happened to be closest.

By default eiCluster will cluster the entire dataset specified by runId. If you want to only cluster a subset of those compounds, you can provide their compound id values to the compoundIds parameter.

    clustering <- eiCluster(runId,K=5,minNbrs=2,cutoff=0.5)
    byCluster(clustering)

8 Customization

eiR can be extended to understand new descriptor types and new distance functions. New distance functions can be set in two different ways. Any function that takes a distance parameter can be given a new distance function that will be used for just that call. If no distance function is given, it will fetch a default distance function that has been defined for the given descriptor type. This default value can be changed using the setDefaultDistance function, which takes the descriptor type and a distance function. Once this function has been called, the new distance function will be used for that descriptor type by all functions using a distance function. The built-in defaults are defined as follows:

setDefaultDistance("ap", function(d1,d2) 1-cmp.similarity(d1,d2) ) 
setDefaultDistance("fp", function(d1,d2) 1-fpSim(d1,d2) ) 

New descriptor types can also be added using the addTransform function. These transforms are basically just ways to read descriptors from compound definitions, and to convert descriptors between string and object form. This conversion is required because descriptors are stored as strings in the SQL database, but are used by the rest of the program as objects.

There are two main components that need to be added. The addTransform function takes the name of the transform and two functions, toString, and toObject. These have slightly different meanings depending on the component you are adding. The first component to add is a transform from a chemical compound format, such as SDF, to a descriptor format, such as atom pair (AP), in either string or object form. The toString function should take any kind of chemical compound source, such an SDF file, an SDF object or an SDFset, and output a string representation of the descriptors. Since this function can be written in terms of other functions that will be defined, you can usually accept the default value of this function. The toObject function should take the same kind of input, but output the descriptors as an object. The actual return value is a list containing the names of the compounds (in the names field), and the actual descriptor objects ( in the descriptors field).

The second component to add is a transform that converts between string and object representations of descriptors. In this case the toString function takes descriptors in object form and returns a string representation for each. The toObject function performs the inverse operation. It takes descriptors in string form and returns them as objects. The objects returned by this function will be exactly what is handed to the distance function, so you need to make sure that the two match each other.

For example, to allow atom pair descriptors to be extracted from and SDF source we would make the following call:

addTransform("ap","sdf",
    toObject = function(input,conn=NULL,dir="."){
        sdfset=if(is.character(input) && file.exists(input)){
            read.SDFset(input)
        }else if(inherits(input,"SDFset")){
            input
        }else{
            stop(paste("unknown type for 'input', 
                or filename does not exist. type found:",class(input)))
        }
        list(names=sdfid(sdfset),descriptors=sdf2ap(sdfset))
    }
)

addTransform("ap",  
    toString = function(apset,conn=NULL,dir="."){
        unlist(lapply(ap(apset), function(x) paste(x,collapse=", ")))
    },
    toObject= function(v,conn=NULL,dir="."){ 
        if(inherits(v,"list") || length(v)==0)
            return(v)

        as( if(!inherits(v,"APset")){
                names(v)=as.character(1:length(v));  
                read.AP(v,type="ap",isFile=FALSE)
            } else v,
            "list")  
    }
)

9 Version Information

sessionInfo()

R version 4.0.0 (2020-04-24) Platform: x86_64-pc-linux-gnu (64-bit) Running under: Ubuntu 18.04.4 LTS

Matrix products: default BLAS: /home/biocbuild/bbs-3.11-bioc/R/lib/libRblas.so LAPACK: /home/biocbuild/bbs-3.11-bioc/R/lib/libRlapack.so

locale: [1] LC_CTYPE=en_US.UTF-8 LC_NUMERIC=C LC_TIME=en_US.UTF-8
[4] LC_COLLATE=C LC_MONETARY=en_US.UTF-8 LC_MESSAGES=en_US.UTF-8
[7] LC_PAPER=en_US.UTF-8 LC_NAME=C LC_ADDRESS=C
[10] LC_TELEPHONE=C LC_MEASUREMENT=en_US.UTF-8 LC_IDENTIFICATION=C

attached base packages: [1] stats graphics grDevices utils datasets methods base

other attached packages: [1] ChemmineOB_1.26.0 RSQLite_2.2.0 eiR_1.28.0 DBI_1.1.0 ChemmineR_3.40.0 [6] BiocStyle_2.16.0

loaded via a namespace (and not attached): [1] Category_2.54.0 bitops_1.0-6 bit64_0.9-7 RcppAnnoy_0.0.16
[5] doParallel_1.0.15 RColorBrewer_1.1-2 progress_1.2.2 httr_1.4.1
[9] gespeR_1.20.0 tools_4.0.0 R6_2.4.1 DT_0.13
[13] affyio_1.58.0 cellHTS2_2.52.0 BiocGenerics_0.34.0 colorspace_1.4-1
[17] tidyselect_1.0.0 gridExtra_2.3 prettyunits_1.1.1 bit_1.1-15.2
[21] curl_4.3 compiler_4.0.0 preprocessCore_1.50.0 graph_1.66.0
[25] glmnet_3.0-2 Biobase_2.48.0 bookdown_0.18 scales_1.1.0
[29] DEoptimR_1.0-8 mvtnorm_1.1-0 robustbase_0.93-6 genefilter_1.70.0
[33] affy_1.66.0 RBGL_1.64.0 askpass_1.1 rappdirs_0.3.1
[37] snowfall_1.84-6.1 stringr_1.4.0 digest_0.6.25 splots_1.54.0
[41] rmarkdown_2.1 rrcov_1.5-2 base64enc_0.1-3 pkgconfig_2.0.3
[45] htmltools_0.4.0 dbplyr_1.4.3 limma_3.44.0 htmlwidgets_1.5.1
[49] rlang_0.4.5 shape_1.4.4 hwriter_1.3.2 dplyr_0.8.5
[53] RCurl_1.98-1.2 magrittr_1.5 Matrix_1.2-18 Rcpp_1.0.4.6
[57] munsell_0.5.0 S4Vectors_0.26.0 lifecycle_0.2.0 vsn_3.56.0
[61] stringi_1.4.6 yaml_2.2.1 MASS_7.3-51.6 zlibbioc_1.34.0
[65] plyr_1.8.6 BiocFileCache_1.12.0 grid_4.0.0 blob_1.2.1
[69] parallel_4.0.0 crayon_1.3.4 lattice_0.20-41 splines_4.0.0
[73] annotate_1.66.0 hms_0.5.3 magick_2.3 locfit_1.5-9.4
[77] knitr_1.28 pillar_1.4.3 rjson_0.2.20 reshape2_1.4.4
[81] codetools_0.2-16 biomaRt_2.44.0 stats4_4.0.0 XML_3.99-0.3
[85] glue_1.4.0 evaluate_0.14 BiocManager_1.30.10 prada_1.64.0
[89] png_0.1-7 vctrs_0.2.4 foreach_1.5.0 gtable_0.3.0
[93] openssl_1.4.1 purrr_0.3.4 assertthat_0.2.1 ggplot2_3.3.0
[97] xfun_0.13 xtable_1.8-4 rsvg_1.3 survival_3.1-12
[101] pcaPP_1.9-73 snow_0.4-3 tibble_3.0.1 iterators_1.0.12
[105] AnnotationDbi_1.50.0 memoise_1.1.0 IRanges_2.22.0 ellipsis_0.3.0
[109] GSEABase_1.50.0

10 Funding

This software was developed with funding from the National Science Foundation: ABI-0957099, 2010-0520325 and IGERT-0504249.

References

Cao, Y, T Jiang, and T Girke. 2010. “Accelerated Similarity Searching and Clustering of Large Compound Sets by Geometric Embedding and Locality Sensitive Hashing.” Bioinformatics 26 (7):953–59. https://doi.org/10.1093/bioinformatics/btq067.

Dimitris K. Agrafiotis, Dmitrii N. Rassokhin, Victor S. Lobanov. 2001. “Multidimensional Scaling and Visualization of Large Molecular Similarity Tables.” In Journal of Computational Chemistry. http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.84.2995.

Dong, Wei, Zhe Wang, Moses Charikar, and Kai Li. 2008. “Efficiently Matching Sets of Features with Random Histograms.” In Proceedings of the 16th ACM International Conference on Multimedia, 179–88. MM ’08. New York, NY, USA: ACM. https://doi.org/10.1145/1459359.1459384.

Dong, Wei, Zhe Wang, William Josephson, Moses Charikar, and Kai Li. 2008. “Modeling LSH for Performance Tuning.” In Proceedings of the 17th ACM Conference on Information and Knowledge Management, 669–78. ACM. https://doi.org/10.1145/1458082.1458172.

Jarvis, RA, and EA Patrick. 1973. “Clustering Using a Similarity Measure Based on Shared Near Neighbors.” IEEE Trans. Comput. 22 (11). IEEE Computer Society Washington, DC, USA:1025–34.