Contents

1 Installation

Install the AnVIL package with

if (!requireNamespace("BiocManager", quietly = TRUE))
    install.packages("BiocManager", repos = "https://cran.r-project.org")
BiocManager::install("AnVIL")

Once installed, load the package with

library(AnVILGCP)
library(AnVIL)

2 Quick start

2.1 Up to speed with AnVIL

The AnVIL project is an analysis, visualization, and informatics cloud-based space for data access, sharing and computing across large genomic-related data sets.

The AnVIL project supports use of R through Jupyter notebooks and RStudio. Support for RStudio is preliminary as of April 2020.

This package provides access to AnVIL resources from within the AnVIL cloud, and also from stand-alone computing resources such as a user’s laptop.

Use of this package requires AnVIL and Google cloud computing billing accounts. Consult AnVIL training guides for details on establishing these accounts.

The remainder of this vignette assumes that an AnVIL account has been established and successfully linked to a Google cloud computing billing account.

2.2 Use in the AnVIL cloud

In the AnVIL cloud environment, clone or create a new workspace. Click on the Cloud Environment button at the top right of the screen. Choose the R / Bioconductor runtime to use in a Jupyter notebook, or RStudio to use in RStudio. When creating a Jupyter notebook, choose R as the engine.

A new layout is being introduced in Fall of 2022. If the workspace has an ‘Analyses’ tab, navigate to it and look for the ‘Environment Configuration’ button to the right of the screen. For a Jupyter notebook-based environment, select jupyter ‘Environment Settings’ followed by Customize and the R / Bioconductor application configuration. RStudio is available by clicking on the RStudio / Bioconductor ‘Environment Settings’ button.

For tasks more complicated than manipulation and visualization of tabular data (e.g., performing steps of a single-cell work flow) the default Jupyter notebook configuration of 1 CPU and 3.75 GB of memory will be insufficient; the RStudio image defaults to 4 CPU and 15 GB of memory

2.3 Local use

Local use requires that the gcloud SDK is installed, and that the billing account used by AnVIL can be authenticated with the user. These requirements are satisfied when using the AnVIL compute cloud. For local use, one must

  • Install the gcloud sdk (for Linux and Windows, cloudml::gcloud_install() provides an alternative way to install gcloud).

  • Define an environment variable or option() named GCLOUD_SDK_PATH pointing to the root of the SDK installation, e.g,

    dir(file.path(Sys.getenv("GCLOUD_SDK_PATH"), "bin"), "^(gcloud|gsutil)$")
## [1] "gcloud" "gsutil"

    Test the installation with gcloud_exists()

    ## the code chunks in this vignette are fully evaluated when
## gcloud_exists() returns TRUE
gcloud_exists()
## Warning in lifeCycle(newpackage = "AnVILGCP", package = "AnVIL", title = "gcloud"): 'gcloud_exists' is deprecated.
##   Use 'AnVILGCP::gcloud_exists' instead.
##   See help('gcloud-deprecated').
## [1] FALSE

2.4 Graphical interfaces

Several commonly used functions have an additional ‘gadget’ interface, allowing selection of workspaces (avworkspace_gadget(), DATA tables (avtable_gadget()) and workflows avworkflow_gadget() using a simple tabular graphical user interface. The browse_workspace() function allows selection of a workspace to be opened as a browser tab.

3 For end users

3.1 Fast binary package installation

The AnVIL cloud compute environment makes use of Docker containers with defined installations of binary system software. Bioconductor has arranged to build ‘binary’ R packages that work out of the box with the BiocManager::install() function. Binary packages (when available and current) install without requiring compilation, and are faster to install than packages built from source.

BiocManager::install("GenomicFeatures")

Thus BiocManager::install() can be used as an improved method for installing CRAN and Bioconductor binary and source packages.

Because package installation is fast, it can be convenient to install packages into libraries on a project-specific basis, e.g., to create a ‘snapshot’ of packages for reproducible analysis. Use

add_libpaths("~/my/project")

as a convenient way to prepend a project-specific library path to .libPaths(). New packages will be installed into this library.

3.2 Working with Google cloud-based resources

The AnVIL package implements functions to facilitate access to Google cloud resources.

Using gcloud_*() for account management

The gcloud_*() family of functions provide access to Google cloud functions implemented by the gcloud binary. gcloud_project() returns the current billing account.

gcloud_account() # authentication account
gcloud_project() # billing project information

A convenient way to access any gcloud SDK command is to use gcloud_cmd(), e.g.,

gcloud_cmd("projects", "list") %>%
    readr::read_table() %>%
    filter(startsWith(PROJECT_ID, "anvil"))

This translates into the command line gcloud projects list. Help is also available within R, e.g.,

gcloud_help("projects")

Use gcloud_help() (with no arguments) for an overview of available commands.

Using gsutil_*() for file and bucket management

The gsutil_*() family of functions provides an interface to google bucket manipulation. The following refers to publicly available 1000 genomes data available in Google Cloud Storage.

src <- "gs://genomics-public-data/1000-genomes/"

gsutil_ls() lists bucket content; gsutil_stat() additional detail about fully-specified buckets.

gsutil_ls(src)

other <- paste0(src, "other")
gsutil_ls(other, recursive = TRUE)

sample_info <- paste0(src, "other/sample_info/sample_info.csv")
gsutil_stat(sample_info)

gsutil_cp() copies buckets from or to Google cloud storage; copying to cloud storage requires write permission, of course. One or both of the arguments can be cloud endpoints.

fl <- tempfile()
gsutil_cp(sample_info, fl)

csv <- readr::read_csv(fl, guess_max = 5000L, col_types = readr::cols())
csv

gsutil_pipe() provides a streaming interface that does not require intermediate disk storage.

pipe <- gsutil_pipe(fl, "rb")
readr::read_csv(pipe, guess_max = 5000L, col_types = readr::cols()) %>%
    dplyr::select("Sample", "Family_ID", "Population", "Gender")

gsutil_rsync() synchronizes a local file hierarchy with a remote bucket. This can be a powerful operation when delete = TRUE (removing local or remote files), and has default option dry = TRUE to indicate the consequences of the sync.

destination <- tempfile()
stopifnot(dir.create(destination))
source <- paste0(src, "other/sample_info")

## dry run
gsutil_rsync(source, destination)

gsutil_rsync(source, destination, dry = FALSE)
dir(destination, recursive = TRUE)

## nothing to synchronize
gsutil_rsync(source, destination, dry = FALSE)

## one file requires synchronization
unlink(file.path(destination, "README"))
gsutil_rsync(source, destination, dry = FALSE)

localize() and delocalize() provide ‘one-way’ synchronization. localize() moves the content of the gs:// source to the local file system. localize() could be used at the start of an analysis to retrieve data stored in the google cloud to the local compute instance. delocalize() performs the complementary operation, copying local files to a gs:// destination. The unlink = TRUE option to delocalize() unlinks local source files recursively. It could be used at the end of an analysis to move results to the cloud for long-term persistent storage.

3.3 Using av*() to work with AnVIL tables and data

Tables, reference data, and persistent files

AnVIL organizes data and analysis environments into ‘workspaces’. AnVIL-provided data resources in a workspace are managed under the ‘DATA’ tab as ‘TABLES’, ‘REFERENCE DATA’, and ‘OTHER DATA’; the latter includes ‘’Workspace Data’ and ‘Files’, with ‘Files’ corresponding to a google cloud bucket associated with the workspace. These components of the graphical user interface are illustrated in the figure below.

The AnVIL package provides programmatic tools to access different components of the data workspace, as summarized in the following table.

Workspace AnVIL function
TABLES avtables()
REFERENCE DATA None
OTHER DATA avbucket()
Workspace Data avdata()
Files avfiles_ls(), avfiles_backup(), avfiles_restore()

Data tables in a workspace are available by specifying the namespace (billing account) and name (workspace name) of the workspace. When on the AnVIL in a Jupyter notebook or RStudio, this information can be discovered with

avworkspace_namespace()
avworkspace_name()

It is also possible to specify, when not in the AnVIL compute environment, the data resource to work with.

## N.B.: IT MAY NOT BE NECESSARY TO SET THESE WHEN ON ANVIL
avworkspace_namespace("pathogen-genomic-surveillance")
avworkspace_name("COVID-19")

Using avtable*() for accessing tables

Accessing data tables use the av*() functions. Use avtables() to discover available tables, and avtable() to retrieve a particular table

avtables()
sample <- avtable("sample")
sample

The data in the table can then be manipulated using standard R commands, e.g., to identify SRA samples for which a final assembly fasta file is available.

sample %>%
    select("sample_id", contains("fasta")) %>%
    filter(!is.na(final_assembly_fasta))

Users can easily add tables to their own workspace using avtable_import(), perhaps as the final stage of a pipe

my_cars <-
    mtcars |>
    as_tibble(rownames = "model") |>
    mutate(model = gsub(" ", "_", model))
job_status <- avtable_import(my_cars)

Tables are imported ‘asynchronously’, and large tables (more than 1.5 million elements; see the pageSize argument) are uploaded in pages. The job status is a tibble summarizing each page; the status of the upload can be checked with

avtable_import_status(job_status)

The transcript of a session where page size is set intentionally small for illustration is

(job_status <- avtable_import(my_cars, pageSize = 10))
## pageSize = 10 rows (4 pages)
##   |======================================================================| 100%
## # A tibble: 4 × 5
##    page from_row to_row job_id                               status
##   <int>    <int>  <int> <chr>                                <chr>
## 1     1        1     10 a32e9706-f63c-49ed-9620-b214746b9392 Uploaded
## 2     2       11     20 f2910ac2-0954-4fb9-b36c-970845a266b7 Uploaded
## 3     3       21     30 e18adc5b-d26f-4a8a-a0d7-a232e17ac8d2 Uploaded
## 4     4       31     32 d14efb89-e2dd-4937-b80a-169520b5f563 Uploaded
(job_status <- avtable_import_status(job_status))
## checking status of 4 avtable import jobs
##   |======================================================================| 100%
## # A tibble: 4 × 5
##    page from_row to_row job_id                               status
##   <int>    <int>  <int> <chr>                                <chr>
## 1     1        1     10 a32e9706-f63c-49ed-9620-b214746b9392 Done
## 2     2       11     20 f2910ac2-0954-4fb9-b36c-970845a266b7 Done
## 3     3       21     30 e18adc5b-d26f-4a8a-a0d7-a232e17ac8d2 ReadyForUpsert
## 4     4       31     32 d14efb89-e2dd-4937-b80a-169520b5f563 ReadyForUpsert
(job_status <- avtable_import_status(job_status))
## checking status of 4 avtable import jobs
##   |======================================================================| 100%
## # A tibble: 4 × 5
##    page from_row to_row job_id                               status
##   <int>    <int>  <int> <chr>                                <chr>
## 1     1        1     10 a32e9706-f63c-49ed-9620-b214746b9392 Done
## 2     2       11     20 f2910ac2-0954-4fb9-b36c-970845a266b7 Done
## 3     3       21     30 e18adc5b-d26f-4a8a-a0d7-a232e17ac8d2 Done
## 4     4       31     32 d14efb89-e2dd-4937-b80a-169520b5f563 Done

The Terra data model allows for tables that represent samples of other tables. The following create or add rows to participant_set and sample_set tables. Each row represents a sample from the corresponding ‘origin’ table.

## editable copy of '1000G-high-coverage-2019' workspace
avworkspace("anvil-datastorage/1000G-high-coverage-2019")
sample <-
    avtable("sample") %>%                               # existing table
    mutate(set = sample(head(LETTERS), nrow(.), TRUE))  # arbitrary groups
sample %>%                                   # new 'participant_set' table
    avtable_import_set("participant", "set", "participant")
sample %>%                                   # new 'sample_set' table
    avtable_import_set("sample", "set", "name")

The TABLES data in a workspace are usually provided as curated results from AnVIL. Nonetheless, it can sometimes be useful to delete individual rows from a table. Use avtable_delete_values().

Using avdata() for accessing Workspace Data

The ‘Workspace Data’ is accessible through avdata() (the example below shows that some additional parsing may be necessary).

avdata()

Using avbucket() and workspace files

Each workspace is associated with a google bucket, with the content summarized in the ‘Files’ portion of the workspace. The location of the files is

bucket <- avbucket()
bucket

The content of the bucket can be viewed with

avfiles_ls()

If the workspace is owned by the user, then persistent data can be written to the bucket.

## requires workspace ownership
uri <- avbucket()                             # discover bucket
bucket <- file.path(uri, "mtcars.tab")
write.table(mtcars, gsutil_pipe(bucket, "w")) # write to bucket

A particularly convenient operation is to back up files or directories from the compute node to the bucket

## backup all files and folders in the current working directory
avfiles_backup(getwd(), recursive = TRUE)

## backup all files in the current directory
avfiles_backup(dir())

## backup all files to gs://<avbucket()>/scratch/
avfiles_backup(dir, paste0(avbucket(), "/scratch"))

Note that the backup operations have file naming behavior like the Linux cp command; details are described in the help page gsutil_help("cp").

Use avfiles_restore() to restore files or directories from the workspace bucket to the compute node.

3.4 Using avnotebooks*() for notebook management

Python (.ipynb) or R (.Rmd) notebooks are associated with individual workspaces under the DATA tab, Files/notebooks location.

Jupyter notebooks are exposed through the Terra interface under the NOTEBOOKS tab, and are automatically synchronized between the workspace and the current runtime.

R markdown documents may also be associated with the workspace (under DATA Files/notebooks) but are not automatically synchronized with the current runtime. The functions in this section help manage R markdown documents.

Available notebooks in the workspace are listed with avnotebooks(). Copies of the notebooks on the current runtime are listed with avnotebooks(local = TRUE). The default location of the notebooks is ~/<avworkspace_name()>/notebooks/.

Use avnotebooks_localize() to synchronize the version of the notebooks in the workspace to the current runtime. This operation might be used when a new runtime is created, and one wishes to start with the notebooks found in the workspace. If a newer version of the notebook exists in the workspace, this will overwrite the older version on the runtime, potentially causing data loss. For this reason, avnotebooks_localize() by default reports the actions that will be performed, without actually performing them. Use avnotebooks_localize(dry = FALSE) to perform the localization.

Use avnotebooks_delocalize() to synchronize local versions of the notebooks on the current runtime to the workspace. This operation might be used when developing a workspace, and wishing to update the definitive notebook in the workspace. When dry = FALSE, this operation also overwrites older workspace notebook files with their runtime version.

3.5 Using avworkflows_*() for workflows

See the vignette “Running an AnVIL workflow within R”, in this package, for details on running workflows and managing output.

3.6 Using avworkspace_*() for workspaces

avworkspace() is used to define or return the ‘namespace’ (billing project) and ‘name’ of the workspace on which operations are to act. avworkspace_namespace() and avworkspace_name() can be used to set individual elements of the workspace.

avworkspace_clone() clones a workspace to a new location. The clone includes the ‘DATA’, ‘NOTEBOOK’, and ‘WORKFLOWS’ elements of the workspace.

3.7 Using drs_*() for resolving DRS (Data Repository Service) URIs

The Data Repository Service (DRS) is a GA4GH standard that separates a resource location (e.g., google bucket of a VCF file) from the URI that identifies the resource. A URI with the form drs://... is submitted to the Terra / AnVIL DRS, and translated to bucket (e.g., gs://...) or https://... URIs. One use case for DRS is when the location (e.g., google bucket) of the resouce moves. In this case the DRS identifier does not change, so no changes are needed to code or data resources that referenced the object. A second use case is when access to a resource is restricted. The DRS URI in conjunction with appropriate credentials can then be translated to a ‘signed’ https URL that encodes authentication information, allowing standard software like a web browser, or R commands like download.file() or VariantAnnotation::readVcf() to access the resource. A Terra support article provides more information, though not about DRS in R!

The following DRS URIs identify a 1000 Genomes VCF file and it’s index

uri <- c(
    vcf = "drs://dg.ANV0/6f633518-f2de-4460-aaa4-a27ee6138ab5",
    tbi = "drs://dg.ANV0/4fb9e77f-c92a-4deb-ac90-db007dc633aa"
)

Information about the URIs can be discovered with drs_stat()

tbl <- drs_stat(uri)
## # A tibble: 2 × 9
##   drs      fileName   size gsUri accessUrl timeUpdated hashes       bucket name
##   <chr>    <chr>     <dbl> <chr> <chr>     <chr>       <list>       <chr>  <chr>
## 1 drs://d… NA21144… 7.06e9 gs:/… NA        2020-07-08… <named list> fc-56… CCDG…
## 2 drs://d… NA21144… 4.08e6 gs:/… NA        2020-07-08… <named list> fc-56… CCDG…

Column names indicate the information that is avaialable, e.g., the google object (gsUri) and size (size) of the object, and the object’s file name (fileName)

drs_cp() provides a convient way to translate DRS URIs to gs:// URIs, and to copy files from their cloud location to the local disk or another bucket, e.g.,

drs_cp(uri, "/tmp")     # local temporary directory
drs_cp(uri, avbucket()) # workspace bucket

drs_access_url() translates the DRS URI to a standard HTTPS URI, but with additional authentication information embedded. These HTTPS URIs are usually time-limited. They can be used like regular HTTPS URIs, e.g,

suppressPackageStartupMessages({
    library(VariantAnnotation)
})
https <- drs_access_url(uri)
vcffile <- VcfFile(https[["vcf"]], https[["tbi"]])
scanVcfHeader(vcffile)
## class: VCFHeader
## samples(1): NA21144
## meta(3): fileformat reference contig
## fixed(2): FILTER ALT
## info(16): BaseQRankSum ClippingRankSum ... ReadPosRankSum VariantType
## geno(11): GT AB ... PL SB

variants <- readVcf(vcffile, param = GRanges("chr1:1-1000000"))
nrow(variants)
## [1] 123077

The buckets are both ‘requester pays’ (see gsutil_requesterpays(uri)), so these queries are billed to the current project.

4 For developers

4.1 Set-up

4.2 Service APIs

AnVIL applications are exposed to the developer through RESTful API services. Each service is represented in R as an object. The object is created by invoking a constructor, sometimes with arguments. We illustrate basic functionality with the Terra() service.

Currently, APIs using the OpenAPI Specification (OAS) Version 2 (formerly known as Swagger) are supported. AnVIL makes use of the rapiclient codebase to provide a unified representation of the API protocol.

Construction

Create an instance of the service. This consults a Swagger / OpenAPI schema corresponding to the service to create an object that knows about available endpoints. Terra / AnVIL project services usually have Swagger / OpenApi-generated documentation, e.g., for the Terra service.

terra <- Terra()

Printing the return object displays a brief summary of endpoints

terra

The schema for the service groups endpoints based on tag values, providing some level of organization when exploring the service. Tags display consists of endpoints (available as a tibble with tags(terra)).

terra %>% tags("Status")

Invoke endpoints

Access an endpoint with $; without parentheses () this generates a brief documentation string (derived from the schema specification. Including parentheses (and necessary arguments) invokes the endpoint.

terra$status
terra$status()

Some arguments appear in the ‘body’ of a REST request. Provide these as a list specified with .__body__ = list(...); use args() to discover whether arguments should be present in the body of the request. For instance,

args(terra$createBillingProjectFull)

shows that all arguments should be included in the .__body__= argument. A more complicated example is

args(terra$overwriteWorkspaceMethodConfig)

where the same argument name appears in both the URL and the body. Again, the specification of the body arguments should be in .__body__ = list(). As a convenience, arguments appearing only in the body can also be specified in the ... argument of the reqeust.

operations() and schemas() return a named list of endpoints, and of argument and return value schemas. operations(terra)$XXX() can be used an alternative to direct invocation terra$XXX(). schemas() can be used to construct function arguments with complex structure.

empty_object() is a convenience function to construct an ‘empty’ object (named list without content) required by some endpoints.

Process responses

Endpoints return objects of class response, defined in the httr package

status <- terra$status()
class(status)

Several convenience functions are available to help developers transform return values into representations that are more directly useful.

str() is invoked for the side-effect of displaying the list-like structure of the response. Note that this is not the literal structure of the response object (use utils::str(status) for that), but rather the structure of the JSON response received from the service.

str(status)

as.list() returns the JSON response as a list, and flatten() attempts to transform the list into a tibble. flatten() is effective when the response is in fact a JSON row-wise representation of tibble-like data.

lst <- status %>% as.list()
lengths(lst)
lengths(lst$systems)
str(lst$systems)

Test endpoints

Testing endpoints is challenging. Endpoints cannot be evaluated directly because they required credentialed access, and because remote calls involve considerable latency and sometimes bandwidth. Traditional ‘mocks’ are difficult to implement because of the auto-generated nature of endpoints from APIs. Simply checking for identical API YAML files (e.g., using md5sums) only indicates a change in the file without assessing whether the R code invoking the endpoint is the same (e.g., because arguments were added, removed, or renamed).

The approach adopted here is to take a ‘snapshot’ of the current API. This is then compared to the updated API. Endpoints that are used in the code but that have been removed or have updated arguments are then manually checked for conformance to the updated API. Once endpoints are brought into line with the new API, the snapshot is updated to reflect the new API.

Non-exported functions in the AnVIL package facilitate these steps. For instance, AnVIL:::.api_test_write(Terra(), "Terra") creates a snapshot of the current API. This is saved as tests/testthat/api-Terra.rds. The service is then updated (following the README of inst/services/terra) and the updated API compared to the original with AnVIL::.api_test_check(Terra(), "Terra"). The result is a list of functions that are common to both APIs, or added, removed, or updated (different arguments) in the new API. A static example is

> .api_test_check(Terra(), "Terra") |> lengths()
        common          added        removed        updated  common_in_use
           135             24              3             11              9
removed_in_use updated_in_use
             0              3

with the removed_in_use and updated_in_use endpoints

> .api_test_check(Terra(), "Terra")[c("removed_in_use", "updated_in_use")]
$removed_in_use
character(0)

$updated_in_use
[1] "cloneWorkspace"         "entityQuery"            "flexibleImportEntities"

requiring manual inspection. Manual inspection means that each use in the AnVIL R package code is examined and updated to match the new API. Once the R code is aligned with the new API, .api_test_write() is re-run. The commit consists of the updated API files in inst/services, updated R code, and the updated snapshot.

Unit tests (in test_api.R) are implemented to fail when the removed_in_use or updated_in_use fields are not zero-length.

4.3 Service implementations

The AnVIL package implements and has made extensive use of the following services:

  • Terra (https://api.firecloud.org/; Terra()) provides access to terra account and workspace management, and is meant as the primary user-facing ‘orchestration’ API.

  • Leonardo (https://leonardo.dev.anvilproject.org/; Leonardo()) implements an interface to the AnVIL container deployment service, useful for management Jupyter notebook and RStudio sessions running in the AnVIL compute cloud.

  • Rawls (https://rawls.dsde-prod.broadinstitute.org; Rawls()) implements functionality that often overlaps with (and is delegated to) the Terra interface; the Rawls interface implements lower-level functionality, and some operations (e.g., populating a DATA TABLE) are more difficult to accomplish with Rawls.

The Dockstore service (https://dockstore.org/swagger.json, Dockstore()) is available but has received limited testing. Dockstore is used to run CWL- or WDL-based work flows, including workflows using R / Bioconductor. See the separate vignette ‘Dockstore and Bioconductor for AnVIL’ for initial documentation.

4.4 Extending the Service class to implement your own RESTful interface

The AnVIL package provides useful functionality for exposing other RESTful services represented in Swagger. To use this in other packages,

  • Add to the package DESCRIPTION file

    Imports: AnVIL

  • Arrange (e.g., via roxygen2 @importFrom, etc.) for the NAMESPACE file to contain

    importFrom AnVIL, Service
importMethodsFrom AnVIL, "$"   # pehaps also `tags()`, etc
importClassesFrom AnVIL, Service

  • Implement your own class definition and constructor. Use ?Service to provide guidance on argument specification. For instance, to re-implement the terra service.

    .MyService <- setClass("MyService", contains = "Service")

MyService <-
    function()
{
    .MyService(Service(
        "myservice",
        host = "api.firecloud.org",
        api_url = "https://api.firecloud.org/api-docs.yaml",
        authenticate = FALSE
    ))
}

Use api_reference_url and api_reference_md5sum of Service() as a mechanism to provide some confidence that the service created by the user at runtime is consistent with the service intended by the developer.

5 Support, bug reports, and source code availability

For user support, please ask for help on the Bioconductor support site. Remember to tag your question with ‘AnVIL’, so that the maintainer is notified. Ask for developer support on the bioc-devel mailing list.

Please report bugs as ‘issues’ on GitHub.

Retrieve the source code for this package from it’s canonical location.

git clone https://git.bioconductor.org/packages/AnVIL

The package source code is also available on GitHub

Appendix

Acknowledgments

Research reported in this software package was supported by the US National Human Genomics Research Institute of the National Institutes of Health under award number U24HG010263. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Institutes of Health.

Session info