In snystrom/memes: motif matching, comparison, and de novo discovery using the MEME Suite

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>"
)

memes uses the universalmotif package to simplify working with motif metadata. universalmotif objects can be represented in an alternative form, the unviversalmotif_df which allows users to manipulate motif metadata just as they would a normal R data.frame (in fact, they are just R data.frames).

These objects are useful for tidying motif metadata to prepare a motif database for use with memes, or performing any other data-driven tasks involving motifs. Here I describe one way these data structures can be used to construct a motif database for use with memes.

library(memes)
library(magrittr)
library(universalmotif)

The MotifDb package makes it easy to query thousands of motifs from public databases. Here, I will describe how to use one of these queries as input to memes functions, and how to manipulate the resulting motifs to prepare them for MEME suite tools.

I will use the motifs from the FlyFactorSurvey as an example. They can be accessed from MotifDb using the following query.

flyFactorDb <- MotifDb::MotifDb %>% 
  MotifDb::query("FlyFactorSurvey")

Use universalmotif::convert_motifs() to convert a MotifDb query into motif objects. In many cases, the resulting list can be used directly as input to memes functions, like runTomTom, or runAme.

flyFactorMotifs <- flyFactorDb %>% 
  convert_motifs()

But there are some issues with this database. For example, the following motif name is the FlyBase gene number, and the alternate name is the actual informative name of the PWM. The MEME Suite relies more heavily on the primary name, so it would be nice if the database used interpretable names.

flyFactorMotifs %>% 
  head(1)

The universalmotif function to_df() converts universalmotif lists into universalmotif_df format which can be used to update motif entries. This is particularly useful when dealing with several motifs at once.

flyFactor_data <- flyFactorMotifs %>% 
  to_df()

The columns of the universalmotif_df can be freely changed to edit the properties of the motifs stored in the motif column. Just like standard data.frames, additional columns can be added to store additional metadata. For more details on these objects see the help page: ?universalmotif::to_df.

# The following columns can be changed to update motif metadata
flyFactor_data %>% 
  names

using the universalmotif_df, we can quickly see that the issue with FBgn numbers only applies to certain entries. And TFs which are represented by multiple motifs in the database are assigned the same name. The MEME Suite tools which use a motif database (like TomTom and AME) require that the entries have unique primary identifiers, therefore the default names will not be appropriate.

flyFactor_data %>% 
  head(5)

However, the altname slots from the motifDb query are already unique, so we can make them the primary name.

length(flyFactor_data$altname) == length(unique(flyFactor_data$altname))

An easy way is to use dplyr::rename to swap the columns.

flyFactor_data %<>% 
  dplyr::rename("altname" = "name", 
                "name" = "altname")

The name column now contains the full motif name.

flyFactor_data %>% 
  head(3)

Next to solve the issue with the FBgn's. FBgn numbers are unique identifiers for a gene within a given FlyBase reference assembly. However, FBgn numbers are not stable over time (i.e. the same gene may have a different FBgn number between reference assemblies), therefore they are unreliable values to determine the correct gene symbol. FlyBase.org has a nice conversion tool which can be used to update FBgn numbers.

As of this writing in March 2021, the FBgn entries provided by the Fly Factor Survey database are out of date. In order to demonstrate an example of methods for tidying motif metadata, I won't use the FlyBase conversion tool, but will instead highlight some approaches which may be more generally useful when working with motif databases from disparate sources.

For this example, we will try to grab the correct gene name from the motif name, which is stored in the first field of the name, formatted as follows: "".

We use tidyr::separate to split out the first entry to the tifd column, then only use this value if the altname contains an FBgn.

flyFactor_data %<>% 
  # Critical to set remove = FALSE to keep the `name` column
  tidyr::separate(name, c("tfid"), remove = FALSE, extra = "drop") %>% 
  # Only use the tfid if the altname contains an FBgn
  dplyr::mutate(altname = ifelse(grepl("^FBgn", altname), tfid, altname))

Now, the first two entries are listed as "ab" instead of "FBgn0259750".

flyFactor_data %>% 
  head(3)

Next, because the FBgn's are out of date, we will remove them from the "names" to shorten up the motif names. This also makes the motif name more comparable to the original motif names from the FlyFactor Survey.

flyFactor_data %<>% 
  dplyr::mutate(name = gsub("_FBgn\\d+", "", name))

A few reality checks

It's worth taking a look at the instances where the altname and our parsed tfid do not match. This is a good way to ensure we haven't missed any important edge cases in the data. As new edge cases are encountered, we can develop new rules for tidying the data to ensure a high quality set of motifs.

Start by simply filtering for all instance where there is a mismatch between altname and tfid.

Carefully compare the altname, name, and tfid columns. Why might the values differ? Are there instances that make you question the data?

flyFactor_data %>% 
  dplyr::filter(altname != tfid) %>% 
  # I'm only showing the first 5 rows for brevity, but take a look at the full
  # data and see what patterns you notice
  head(5)

One thing that becomes obvious is that many motifs have mismatched altname/tfid values because of capitalization or hyphenation differences. You can use domain-specific knowledge to assess which one is correct. For Drosophila, "abd-A" is correct over "AbdA", for example.

After manually inspecting these rows, I determined that instances of different capitalization, hyphenation, or names that contain "." or "()" can be ignored. To further investigate the data, I will ignore capitalization and special character differences as follows:

flyFactor_data %>% 
  # calling tolower() on both columns removes capitalization as a difference
  dplyr::filter(tolower(altname) != tolower(tfid),
                # Select all altnames that do not contain "-", "." or "("
                !grepl("-|\\.|\\(", altname),
                ) %>% 
  # I'll visalize only these columns for brevity
  dplyr::select(altname, tfid, name, consensus) %>% 
  head(10)

Next, what is obvious is that several altnames set to "da" have a high number of mismatched tfids. For instance, amos_da_SANGER_10. When checking the FlyFactorSurvey page for da, it reveals only 1 motif corresponds to this factor. Checking the page for amos shows a match to amos_da_SANGER_10. Therefore, we can conclude that factors assigned the name of da are incorrectly assigned, and we should prefer our parsed tfid.

flyFactor_data %<>% 
  # rename all "da" instances using their tfid value instead
  dplyr::mutate(altname = ifelse(altname == "da", tfid, altname))

Now we've handled the "da" mismatches, we filter them out to identify new special cases.

flyFactor_data %>% 
  dplyr::filter(tolower(altname) != tolower(tfid),
                !grepl("-|\\.|\\(", altname)) %>% 
  dplyr::select(altname, tfid, name, consensus) %>% 
  head(10)

The next thing to notice about these data is that entries with "CG" prefixed tfids are often mismatched. This is because when the FlyFactor survey was conducted, many genes were unnamed, and thus assigned a CG from FlyBase. As time has gone on, some CG's have been named. Checking the FlyBase page for CG10267 reveals that it has been renamed "Zif". This matches with the altname, so we conclude that rows with a "CG" tfid can be safely skipped as their altname contains the new gene symbol.

flyFactor_data %>% 
  dplyr::filter(tolower(altname) != tolower(tfid),
                !grepl("-|\\.|\\(", altname),
                # Remove CG genes from consideration
                !grepl("CG\\d+", tfid)
                ) %>% 
  dplyr::select(altname, tfid, name, consensus)

The remaining rows (only 20 values) can be manually inspected for any discrepancies. I went through each entry by hand, looking up their motifs on FlyFactor, and their gene names on FlyBase to determine the best way to handle these motifs. Sometimes the best way to be sure your data are high quality is to carefully inspect it!

I determined from this that a few altnames need swapping, and one motif I will remove because it is unusual (Bgb has an identical motif to run, but the motif is marked "run" on the FlyFactor website).

I'll make those changes to the data:

swap_alt_id <- c("CG6272", "Clk", "Max", "Mnt", "Jra")
remove <- "Bgb"

flyFactor_data %<>% 
  dplyr::mutate(altname = ifelse(altname %in% swap_alt_id, tfid, altname)) %>% 
  dplyr::filter(!(altname %in% remove))

Finally, the remaining motif metadata is also OK based on my manual inspection.

flyFactor_data %>% 
  dplyr::filter(tolower(altname) != tolower(tfid),
                !grepl("-|\\.|\\(", altname),
                # Remove CG genes from consideration
                !grepl("CG\\d+", tfid)
                ) %>% 
  dplyr::select(altname, tfid, name, consensus)

Removing duplicate motif matrices

Just because the metadata for each entry is unique, this does not mean that the motif matrix for each entry is unique. There are many reasons why two different factors could have identical motifs: some biological, others technical. In the case of the FlyFactorSurvey, some entries are duplicated in MotifDb which should not be.

For instance, the following motif is a duplicate where the tidied metadata matches:

flyFactor_data %>% 
  dplyr::filter(consensus == "MMCACCTGYYV")

It is difficult to determine in a high-throughput way whether any matrix entries are identical in a large database, and it is not always possible to rely on metadata to determine matrix duplication.

In order to identify and remove duplicate motif matrices, memes provides remove_duplicate_motifs(), which can be used to deduplicate a list of motifs based solely on their motif matrices (i.e. it ignores motif name & other metadata). We will use this strategy to deduplicate the flyFactor data.

(NOTE: When working with other motif databases, it is critical to understand the data source to determine appropriate measures for handling duplicated entries.)

# This operation takes a while to run on large motif lists
flyFactor_dedup <- remove_duplicate_motifs(flyFactor_data)

Duplicate removal identifies and removes r nrow(flyFactor_data) - nrow(flyFactor_dedup) identical matrices.

# Rows before cleanup
nrow(flyFactor_data)
# Rows after cleanup
nrow(flyFactor_dedup)

Using the example from before now shows only 1 motif corresponding to this sequence.

flyFactor_dedup %>% 
  dplyr::filter(consensus == "MMCACCTGYYV")

Finally, now that the database has been tidied and deduplicated, the resulting data.frame can be converted back into a universalmotif list using to_list(). To discard the additional columns we created so they are not passed on to the universalmotif, set extrainfo = FALSE.

# extrainfo = FALSE drops the extra columns we added during data cleaning which are now unneeded
flyFactorMotifs_final <- to_list(flyFactor_dedup, extrainfo = FALSE)

The resulting universalmotif list object now reflects the changes we made to the data.frame and can now be exported as a .meme format file using universalmotif::write_meme or can be used directly as input to tools like runTomTom or runAme.

flyFactorMotifs_final %>% 
  head(1)

This cleaned-up version of the FlyFactorSurvey data is packaged with memes in system.file("extdata/flyFactorSurvey_cleaned.meme", package = "memes").

write_meme(flyFactorMotifs_final, "flyFactorSurvey_cleaned.meme")

Session Info

sessionInfo()

snystrom/memes documentation built on Oct. 12, 2024, 2:42 a.m.