search

biocore / redbiom

Sample search by metadata and features
Other
45 stars 20 forks source link
bioinformatics biom microbiome qiime

readme

redbiom

What is this?

Redbiom is a cache service for sample metadata and sample data. It allows for rapidly:

redbiom is designed to handle biological and technical replicates. Specifically, it allows for a one to many relationship between a sample's metadata and its data, both within and between preparation types. Additional information about redbiom can be found in our mSystems article.

This repository defines the de facto redbiom data representation, and one possible interface into the resource. Other interfaces (e.g., Javascript) are possible to define. Please see the Design section below for details about how other interfaces can be written.

By default, redbiom will search against qiita.ucsd.edu:7329. This can be changed at runtime by setting the REDBIOM_HOST environmental variable, e.g., export REDBIOM_HOST=http://qiita.ucsd.edu:7329. The default host is read-only and administrative functions like loading data will not work against it.

If you intend to load your own data, you must setup a local instance (please see the server installation instructions below). In addition, you must explicitly set the REDBIOM_HOST environment variable.

Citation

To cite redbiom, please refer to:

redbiom: a Rapid Sample Discovery and Feature Characterization System. Daniel McDonald, Benjamin Kaehler, Antonio Gonzalez, Jeff DeReus, Gail Ackermann, Clarisse Marotz, Gavin Huttley, Rob Knight. mSystems Jun 2019, 4 (4) e00215-19; DOI: 10.1128/mSystems.00215-19

Very brief examples

A few quick examples of what can be done. More complex and detailed examples can be found later in the document.

Get all the samples in which the word "beer" is found:

$ redbiom search metadata beer | head
10105.Ingredient.18
1976.Beer.286
1689.261
10105.Ingredient.19
1976.Beer.262
10105.Ingredient.23
1976.Beer.228
10105.Ingredient.17
10395.000041066
10105.Ingredient.24

Get the closed reference OTU picking 16S V4 data for those samples (more on what ctx and context is in the longer examples below):

$ export ctx=Pick_closed-reference_OTUs-Greengenes-Illumina-16S-V4-5c6506
$ redbiom search metadata beer | head | redbiom fetch samples --context $ctx --output beer_example.biom
$ redbiom search metadata beer | head | redbiom fetch sample-metadata --context $ctx --output beer_example.txt

Find the feature IDs (Greengenes OTU IDs in this case) associated with S. aureus (and for example purposes, an arbitrary 10):

$ redbiom search taxon --context $ctx s__aureus | head
943389
1023241
862312
1102743
870118
969777
1086805
976010
951488
951136

...and then find samples which contain those 10 S. aureus features:

$ redbiom search taxon --context $ctx s__aureus | head | redbiom search features --context $ctx | wc -l
   21577

Installation

General requirements

Redbiom depends on Python (tested on 3.5 and 3.6), BIOM (tested on >= 2.1.5), Pandas (tested on 0.19.0), Click (required >= 6.7), nltk (tested on 3.2.2), joblib (tested on 0.9.3), and scipy (whatever BIOM is happy with).

Client

If you would like to use redbiom as only a client (which is the general case), then the following instructions apply. Note that we need to install numpy separately as one of the dependencies, BIOM-Format, imports numpy within its installation process.

$ pip install numpy
$ pip install redbiom

Alternatively, you can install redbiom through conda:

$ conda install -c conda-forge redbiom

Server

If you would like to run your own resource, and load data locally or private data, then the following instructions apply.

In addition to the general requirements, redbiom server needs Redis (tested with 2.8.17 and 3.2.6) and Webdis (just clone the repo). It is not necessary to have super user access to establish a redbiom server.

For Redis, the following has worked on OSX and multiple flavors of Linux without issue.

$ wget http://download.redis.io/releases/redis-3.2.6.tar.gz
$ tar xzf redis-3.2.6.tar.gz
$ pushd redis-3.2.6
$ make
$ ./src/redis-server --daemonize
$ popd

Webdis packages its dependencies with the exception of libevent. It is entirely likely that libevent is already available on your system. If so, the following should work. If libevent is not available, compilation will die quickly. However, libevent is in all the common repositories (e.g., yum, apt, brew, etc), and compiling by source is straight forward. 

$ git clone https://github.com/nicolasff/webdis.git
$ pushd webdis
$ make
$ ./webdis &
$ popd

Last, redbiom itself can be installed as a normal Python package.

$ pip install numpy
$ pip install redbiom

Testing

The test framework is setup to by default only execute tests against localhost, specifically, 127.0.0.1:7379. However, the repository, by default, is setup to communicate with a remote Webdis server. If you wish to execute the test suite, please export REDBIOM_HOST=http://127.0.0.1:7379.

Terminology and notation

In redbiom, the word "context" refers to a way in which the sample data were processed. Data are loaded into contexts and searches for samples by feature happen within contexts.

To support the one to many relationship between a sample's metadata and its data, within a context, a sample's IDs are prefixed by a "tag" which can be specified at load. Internally, within a context, these IDs are of the form <tag>_<sample-id>. The use of the _ character ensures that they are not valid QIIME sample IDs, and is necessary so we can appropriately differentiate these IDs. Methods which write data will coerce these invalid QIIME IDs into valid IDs of the form <sample-id>.<tag>. IMPORTANT: if you run your own resource, it is important to specify --tag on load of sample data to differentiate BIOM tables in which the sample IDs between the tables may not be mutually exclusive.

Commands which write data will notify the user if there were ambiguities. An ambiguitiy means that there was a sample ID which mapped to multiple redbiom IDs within the output. The IDs written are unique because of the reasons noted above.

Command structure

Redbiom relies on Click to provide a tiered command line interface. An example of the first tier is below, and with the exception of admin, are composed of verbs:

$ redbiom --help
Usage: redbiom [OPTIONS] COMMAND [ARGS]...

Options:
  --version  Show the version and exit.
  --help     Show this message and exit.

Commands:
  admin      Update database, etc.
  fetch      Sample data and metadata retrieval.
  search     Feature and sample search support.
  select     Select items based on metadata
  summarize  Summarize things.

The actual commands to execute are contained within a submodule. For instance, below are the commands associated with "search":

$ redbiom search --help
Usage: redbiom search [OPTIONS] COMMAND [ARGS]...

  Feature and sample search support.

Options:
  --help  Show this message and exit.

Commands:
  metadata      Find samples or categories.
  features      Find samples containing features.
  taxon         Find features associated with a taxon

The intention is for commands to make sense in English. The general command form is "redbiom ", however this form is not strictly enforced.

In general, these commands are intended to be composable via Unix pipes. For example:

redbiom search metadata antibiotics | redbiom fetch samples --context <foo> --output my_table.biom

Examples

Search for samples by metadata

By default, redbiom is setup to query against Qiita. First, let's search for some samples by metadata. Specifically, what we're going to do is identify what samples exist in Qiita in which any of their sample metadata contains the stem of the word beer. This returns quite a few samples, so for the sake of the example, we're only going to show the first 10 using head: 

$ redbiom search metadata beer | head
10105.Ingredient.18
1976.Beer.286
1689.261
10105.Ingredient.19
1976.Beer.262
10105.Ingredient.23
1976.Beer.228
10105.Ingredient.17
10395.000041066
10105.Ingredient.24

$ redbiom search metadata beer | wc -l
  416

Now that we have some samples, let's pull out their sample data. Qiita contains a huge amount of data, which are logically partitioned by the sample preparations and processing parameters -- these partitions are denoted as contexts in redbiom. In order to pull out the data, we need to specify the context to operate in. There are a lot of contexts, so let's filter to only those which are 16S and V4 using grep. We're also going to cut the first three columns of data as the fourth one is a voluminous description of the processing parameters. And last, let's sort the results by the number of samples represented in the context. Unfortunately, the grep removes the column headers, so we'll run a second summarize command and just grab the header:

$ redbiom summarize contexts | cut -f 1,2,3 | grep 16S-V4 | grep Greengenes-Illumina |  sort -k 2 -n
Pick_closed-reference_OTUs-Greengenes-Illumina-16S-V4-41ebc6    100 14434
Pick_closed-reference_OTUs-Greengenes-Illumina-16S-V4-250nt-66a626  174 2686
Pick_closed-reference_OTUs-Greengenes-Illumina-16S-V4-200nt-a5e305  7009    16070
Pick_closed-reference_OTUs-Greengenes-Illumina-16S-V4-58196d    8468    34789
Pick_closed-reference_OTUs-Greengenes-Illumina-16S-V4-125nt-65468f  27100   43261
Pick_closed-reference_OTUs-Greengenes-Illumina-16S-V4-150nt-bd7d4d  145308  73089
Pick_closed-reference_OTUs-Greengenes-Illumina-16S-V4-90nt-44feac   173749  74298
Pick_closed-reference_OTUs-Greengenes-Illumina-16S-V4-100nt-a243a1  173809  75990
Pick_closed-reference_OTUs-Greengenes-Illumina-16S-V4-5c6506    200552  84164

$ redbiom summarize contexts | head -n 1
ContextName SamplesWithData FeaturesWithData    Description

To reduce typing later, let's just pick a context and store it as an environment variable:

$ export ctx=Pick_closed-reference_OTUs-Greengenes-Illumina-16S-V4-5c6506

...and now we can grab some data:

$ redbiom search metadata beer | redbiom fetch samples --context $ctx --output example.biom
$ biom summarize-table -i example.biom | head
Num samples: 203
Num observations: 5,265
Total count: 5,187,346
Table density (fraction of non-zero values): 0.026

Counts/sample summary:
Min: 1.000
Max: 208,223.000
Median: 11,172.000
Mean: 25,553.429

We probably also want to get the sample metadata:

$ redbiom search metadata beer | redbiom fetch sample-metadata --output example.txt --context $ctx

You might note that the total number of samples found by the metadata search is not the same as the number of samples found by the sample data fetch. The sample information is distinct from the sample preparation, and data processing: just because there is sample information does not mean a given sample has (for instance) 16S V4 sequence data associated with it.

The query structures for metadata are fairly permissive, and there are actually two types of queries that can be performed. The structure is as follows: <set operations> where <value restrictions>. The <set operations> work by finding all samples with that contain a given word, which can be combined together. For the set queries, & performs an intersection of the sample IDs, | a union, and - a difference:

$ redbiom search metadata "soil & europe where ph < 7" | wc -l
5824

IMPORTANT: just because a sample may have a word associated with it, does not mean that word is used as you may expect. In the example below, we're counting the number of samples by their described sample_type value. We are working to improve the search functionality, and it is important for users to scrutinize their results:

$ redbiom search metadata "soil & europe where ph < 7" | redbiom summarize samples --category sample_type  | head
XXQIITAXX   1489
soil    1186
fresh water 724
water   722
Soil    595
cheese  435
peat    192
biofilm 139
wetland soil    78
belly   41

Search by feature

We can also use redbiom to search for samples containing features of interest. Let's operate off our example table from the metadata search above. What we're going to do is find all samples in Qiita that contain any of the a handful of the feature IDs. In this particular example, let's just grab 10 arbitrary IDs:

$ biom table-ids -i example.biom --observations | head
4449525
4420570
471180
815819
4235445
1108951
519367
12364
4454153
4227110

...and then let's pipe them back into redbiom to search for other samples in our context which contain those same features:

$ biom table-ids -i example.biom --observations | head | redbiom search features --context $ctx | wc -l
   43133

$ biom table-ids -i example.biom --observations | head | redbiom search features --context $ctx | head
3759_10172.3338
2923_10317.000017653
2096_1716.McG.PAPrS17
2015_1034.CHB1
2150_755.SSFA.L1.D30.07.06.11.lane1.NoIndex.L001
2150_755.LSSF.ALPHA.D20.14.07.11.lane1.NoIndex.L001
26483_10317.000007237
3788_10119.MT.741
2112_1774.527.Skin.Puer
2102_1734.BD.ERD510

Search by taxon

One thing you might want to do is find features based on taxonomy. We can do this by searching for a taxon:

$ redbiom search taxon g__Roseburia --context $ctx | wc -l
     121

What we get back are the feature IDs that are of that taxon. We can then take those feature IDs and feed them back into redbiom. So for instance, let's say we wanted to find all samples which contain a Roseburia feature:

$ redbiom search taxon g__Roseburia --context $ctx | redbiom search features --context $ctx | wc -l
   84884

IMPORTANT not all contexts necessarily have taxonomy, and taxonomy may not make sense for a context (e.g., if it contains KEGG Orthologous group features).

Retrieving pre-selected samples

In additional to allowing you to search based on specific metadata or features, you can also retrieve a list of samples based on the sample ID. For instance, we might want to get a list of all the samples with cider associated with them, and then potentially access only these samples later, after a database update.

To do this, we can pulldown a list of the first five samples with cider associated.

$ redbiom search metadata cider | head -5 > cider.txt
$ head cider.txt
    11261.CW91.R1.T7
    11261.CW130.S.F2.T7
    11261.CW120.F1.T4
    11261.CW75.R3.T1
    11261.CW125.F3.T5

Then, we can use this list of samples to retrieve the biom table. The text file simply needs to be a list of sample IDs in the databse, one per line.

$ redbiom fetch samples --from cider.txt --context $ctx --output cider.biom
$ biom summarize-table -i cider.biom | head
Num samples: 5
Num observations: 281
Total count: 173,579
Table density (fraction of non-zero values): 0.396

Counts/sample summary:
 Min: 25,936.000
 Max: 38,900.000
 Median: 36,602.000
 Mean: 34,715.800

Summarizations

We found a lot of samples that contain Roseburia. That isn't too surprising since Qiita contains a lot of fecal samples. How many? In this next example, we're taking all of the feature IDs associated with Roseburia, then finding all of the samples which contain that taxon, followed by binning each sample by their sample_type category value, and finally we're taking just the top 10 entries. You can see that the metadata are a bit noisy.

$ redbiom search taxon g__Roseburia --context $ctx | redbiom search features --context $ctx | redbiom summarize samples --category sample_type | head
Stool   21029
stool   18879
feces   16333
skin    4242
XXQIITAXX   3080
control blank   1253
saliva  1012
surface 995
tanker milk 984
biopsy  930

We can still work through the noise though. Let's take our samples we found that contain Roseburia, and only select the ones that appear to obviously be fecal. Instead of summarizing as we did in our last example, we're going to "select" the samples in which sample_type is either "Stool" or "stool". (as this command is getting long, we'll break it up with \):

$ redbiom search taxon g__Roseburia --context $ctx | \
    redbiom search features --context $ctx | \
    redbiom select samples-from-metadata --context $ctx "where sample_type in ('Stool', 'stool')" | \
    wc -l
   39908

And last, we can grab the data for those samples. Fetching data for 24,667 samples can take a few minutes, so for the purpose of the example, let's just grab the ones associated with skin. Please note the "ambiguity" on the output, more in a second on that:

$ redbiom search taxon g__Roseburia --context $ctx | \
    redbiom search features --context $ctx | \
    redbiom select samples-from-metadata --context $ctx "where sample_type=='skin'" | \
    redbiom fetch samples --context $ctx --output roseburia_example.biom
16 sample ambiguities observed. Writing ambiguity mappings to: roseburia_example.biom.ambiguities

Ambiguities can arise if the same sample was processed multiple times as might happen with a technical replicate. It is the same physical sample, but it may have been processed multiple times. The .ambiguities file is in JSON and contains a mapping of what IDs map to the same sample.

Load some data (i.e., if you are running your own server)

To make use of this cache, we need to load things. Loading can be done in parallel. First, we need to set the server to be writable.

$ redbiom admin scripts-writable

Next, we'll load up metadata. This will create keys in Redis which describe all of the columns associated with a sample (e.g., metadata:categories:<sample_id>, hash buckets for each category and sample combination (e.g., metadata:category:<category_name> as the hash and <sample_id> as the field), a set of all known categories (e.g., metadata:categories-represented), and a set of all known sample IDs (e.g., metadata:samples-represented):

$ redbiom admin load-sample-metadata --metadata path/to/qiime/compat/mapping.txt

redbiom supports one to many mappings between sample metadata and actual sample data. This is done as there may be multiple types of processing performed on the same data (e.g., different nucleotide trims). Or, a physical sample may have been run through multiple protocols (e.g., 16S, WGS, etc). So before we load any data, we need to create a context for the data to be placed. The following action will add an entry into the state:contexts hash bucket keyed by name and valued by description:

$ redbiom admin create-context --name deblur-100nt --description "16S V4 Caporaso et al data deblurred at 100nt"

Last, let's load up all of the BIOM table data. We'll only store the non-zero values, and we'll encode the sample data into something simple so that it goes in as just a string to Redis. Important: we only support storing count data right now, not floating point. The keys created are of the form <context_name>:sample:<redbiom_id>. To reduce space, we reindex the feature IDs as things like sOTUs tend to be very long in name. The mapping is stable over all tables loaded (ie the same feature has the same index), and is stored under <context_name>:feature-index. Because we need to update the index, this operation cannot be done in parallel however the code is setup with a redis-based mutex so it's okay to queue up multiple loads.

$ redbiom load-sample-data --context deblur-100nt --table /path/to/biom/table.biom

Caveats

Redbiom is still in heavy active development. At this time, there are still some important caveats.

Design

Python and testing

There are a few design decisions in place which deviate from some other typical Python projects. First off, the majority of imports are deferred. The motivating force here is to minimize overhead on load as to provide a responsive user interface -- deferred imports are the most straight forward way to accomplish that goal.

The test harness is broken into multiple components, and are driven by a Makefile. This was done initially to be pragmatic as it was easier to write integration tests than unit tests for the click commands. These tests can be found in test.sh which is composed of "positive" tests and test_failures.sh which is composed of "negative" tests. The difference being that the positive tests will fail if any command results in a nonzero exit status, whereas the negative tests expect a nonzero exit status (and really, the decision was to avoid unsetting "-e"). Additional tests which validate some of the Redis contents can be found in redbiom/tests/test_rest.py. These are neither unit tests nor integration tests, but simply exercise the behind-the-scenes REST interface. Last, there are a suite of unit tests placed under redbiom/tests/.

Redis data organization

Because redbiom is currently in alpha, and its data model is subject to change, we are holding off an indepth description of it. That being said, the API methods in general outline the Redis commands issued within their docstrings and can be used to guide interaction.

The key structures used are in the following forms: