Repo for Google Summer of Code Open Data Project 2019.

For a google docs version of this ReadMe see here

OpenWorm is an attempt to build the first biophysical simulation of the model organism C. elegans. C. elegans are an important model organism in biological research and a large amount of data has been generated surrounding them at a behavioural, physiological, neuronal and molecular level. As the OpenWorm project reaches critical milestones, it is increasingly important to be able to validate simulations by incorporating all available data. However, the publication of biological data is often weakly structured and has no standardised format making assimilation challenging. To validate existing simulations and enter the next phase of development will require a significant increase and improvement of the amount of data available.

A key improvement required will be to store data in a way that is most useful for future use. Developed in 2016, the FAIR principles set out a framework for addressing this goal and are already being followed by other major data collators. The FAIR principles emphasise the importance of metadata; information that provides context to the main research data being published. Metadata can range from the protocol and conditions used to generate research data to the version of the data being published and how this fits in with complementary publications. The availability of good metadata increases the reusability of research generated data and can give insight into differences between experimental results. The framework emphasises: Findability, especially with an emphasis on machines being able to find and use the data; Accessibility, involving licence checks and levels of authorisation; Interoperability, ensuring that data fits in with other data sources by using the same format, vocabulary and linking to similar work; and Reusability, ensuring provenance, usage rights and protocols are clearly described to be most useful for future work.

This Google Summer of Code project thus has two main aims. One: reformat existing data in line with the FAIR principles to make it more accessible for the OpenWorm community and for biological researchers as a whole. Two: increase the amount of data available by extracting features from already existing literature and data sources for easier access. To accomplish this I have focussed on 3 main sources of data: Worm movement videos, Calcium Imaging and Neuronal Coordinates.

The integration of movement data is important for both building and validating the OpenWorm models of C. elegans. Neuromechanical models have so far given insight into how movement can be shaped by neuronal activity, helped classify mutants and have led to the identification of a minimum subset of features useful for quantifying behaviour. Classification and quantification allows for easier comparison of neural simulations to empirical movement data. A large dataset of worm movement videos has already been collated and stored alongside metadata in Zenodo, a data storage site which emphasises open science and metadata accessibility. An interactive website, Open Worm Movement Database, exists to easily filter some metadata in the worm movement videos as required. Given these movement records are well maintained and appear to already contain easily accessible metadata it was decided that they would act as a good proof of principle for the two aims of the project.

The first task was to establish how well the data currently conforms to the FAIR principles. In addition to the 2016 publication, additional guidance on conforming to the FAIR framework exists here. Using this, existing metadata was grouped into categories and any missing metadata categories were flagged to be added at a later date.

Included already within OpenWorm is the codebase, PyOpenWorm. PyOpenWorm was designed specifically for the purpose of increasing searchability and accessibility of C. elegans research data. Currently incorporated data is taken from WormBase, WormAtlas and personal communications with researchers in the C. elegans community. It can be divided into three categories: Neuronal Information, Muscle Cells and the Connectome.

Once the movement data had been processed to ensure conformability with the FAIR framework the second task was to upload all data into PyOpenWorm in a format facilitating queriability by C. elegans researchers and model builders.

The integration of movement metadata into PyOpenWorm had six main phases as seen below.

First, following the FAIR framework, a set of categories were determined. These categories were decided to ensure the recommendations of the FAIR principles were followed whilst keeping the data as easily accessible as possible. These were greatly informed by the guidance already set out in WCON. Second, the metadata surrounding each movement record in Zenodo was harvested whilst still in its current format. Third, the data harvested was changed to a standardised format, with missing categories added and set to “None” to ensure uniformity across records. Fourth, the standardised metadata could now be integrated into the PyOpenWorm codebase following the instructions in the documentation. Fifth, after integration unit tests are performed to check that data has been added completely and correctly. Sixth, any FAIR categories not filled by the existing metadata are filled through additional searches.

Movement data and metadata is stored in Zenodo within the Community “Open Worm Movement Database”. As of August 2019, within the community are 15000 records. Clicking on one of these records gives a link to both the data itself and the metadata associated.

Document 1 shows the categories decided to implement the FAIR principles for each record.

Zenodo already provides interoperability for all its records with the Open Archives Initiatives, a protocol developed for metadata harvesting. Going back to the main Open Worm Movement Database Community page and clicking the link “OAI-PMH Interface” under “Harvesting API” on the right hand side allows us to access the metadata in the OAI format.

In this format we can use a premade OAI-PMH Metadata harvester, to cycle through each of the links using the token at the end of each page and harvest the metadata surrounding each record. I chose to use the prebuilt python package, Sickle.

Using Sickle we are able to access and extract data as needed from the “Open Worm Movement Database” Community.

import sickle
from sickle import Sickle
from sickle.iterator import OAIResponseIterator

#access Zenodo
sickle = Sickle("https://zenodo.org/oai2d")

#create object which contains all Zenodo records within the community "Open Worm Movement Database"
records = sickle.ListRecords(set='user-open-worm-movement-database', metadataPrefix='oai_dc')

Each record in records is in .xml format with the headers: OAI Identifier, Datestamp, setSpec, Author or Creator, Date, Description, Resource Identifier, Relation, Rights Management, Title, Resource Type.

The information stored within description is in yaml format. We are able to save the entire xml file using this script. This zipped folder of xml files can be accessed by running the script. However, as we are most interested in the information stored under Description, I have also created a zipped folder of yaml files using this script and accessible here.

As not every record contained the same data and headings, it was necessary to modify each yaml record into a standard format. This script finds the categories which are not common to all records and the script linked above adds them in the format, “category: None” as required to ensure all yaml records contain the same set of categories. Document 1 specifies the categories which were incorporated. Likewise, any headers which contain missing information designated by -N/A- are modified to store “None” as their value instead. Characters which cause difficulty in reading yaml files were also replaced for ease of processing. In addition, an identifier in the form of a doi was added to each yaml file and then inserted into the filename.

for r in recs:
   a = r.metadata["description"]
   d = r.metadata["identifier"][1]
   y = d.replace("/", "_")
   y = y.replace(".", "_")
   e = "\nidentifier: "
   var3 = e + d
   a = [s[1:] for s in a]
   a = [w.replace('@', '') for w in a]
   a = [w.replace('DA609: npr-1(ad609)X.', 'DA609 npr1(ad609)X.') for w in a]
   a = [w.replace('AX994: daf-22(m130)II; npr-1(ad609)X.', 'AX994 daf22(m130)II npr1(ad609)X.') for w in a]
   a = [w.replace('N2: lab reference strain.', 'N2 lab reference strain.') for w in a]
   a = [w.replace('AX994: daf-22(m130)II.', 'AX994 daf22(m130)II.') for w in a]
   a = [w.replace('OMG2: mIs12[myo-2p::GFP]II; npr-1(ad609)X. OMG19: rmIs349[myo3p::RFP]; npr-1(ad609)X.', 'OMG2 mIs12[myo2pGFP]II npr1(ad609)X. OMG19 rmIs349[myo3pRFP] npr1(ad609)X.') for w in a]
   a = [w.replace('OMG10: mIs12[myo-2p::GFP]II. OMG24: rmIs349[myo3p::RFP].', 'OMG10 mIs12[myo2pGFP]II. OMG24 rmIs349[myo3pRFP].') for w in a]
   a = [w.replace('-N/A-','None') for w in a]
   a = [w.replace('\n\t', '\n') for w in a]
   a = [w.replace('\n\n', '\n') for w in a]
   a = [w.replace('\t', '  ') for w in a]
   b = a[0]
   included_standard_elements = sorted([element for element in standardised_list if (element in b)])
   unincluded_standard_elements = list(set(standardised_list).difference(included_standard_elements))
   for i in unincluded_standard_elements:
       addel = "\n{}".format(i) + " : None"
       b = b + addel
   g = b.split('\n', 1)[1]
   c = var3 + g
   img_name = "yamlfile_{}.yaml".format(y)
   print("  Writing image {:s} in the archive".format(img_name))
   zf.writestr(img_name, c)

A full walk through of how PyOpenWorm can be installed and movement metadata incorporated is provided here in the google colab virtual environment.

Data is uploaded into PyOpenWorm in the form of classes. These classes must first be created and saved to the PyOpenWorm code base before any data can be uploaded in the required format. Further documentation regarding adding data to PyOpenWorm can be found here.

This script was used to create classes. Information was uploaded under the main class MovementMetadata which was in turn split into subclasses: Usage, Provenance, Collection, BioDetails and Software. In a separate script, I could then upload the information in the form of these classes creating a new context for each record, referenced by the doi each time.

These were then saved, committed and pushed to PyOpenWorm using the pow command. The pow command is specific to PyOpenWorm and provides a high level interface for working with PyOpenWorm managed data.

This ensures future users of PyOpenWorm are now able to access the movement metadata.

To check that all metadata was uploaded completely and correctly, unit tests were written. These can be run each time movement metadata information is accessed and ensure that the metadata being accessed is still as expected.

This script (still in progress) was written to check the data. Key checks include questioning if numerically all records have been uploaded, checking responses are returned in a reasonable amount of time and incorporating spot checks of 3 random records to ensure the information returned is correct.

The script will follow the format of previously incorporated tests: DataIntegrityTest and EvidenceQualityTest.

As can be seen in Document 1 to conform fully to the FAIR principles additional metadata needs to be harvested. The key categories missing include:

• Full description of the protocol
• Full description of the scope of the project
• Full description of the limitations of the project
• The addition of GO terms and taxon ID
• Ensuring a version number is linked to each data record
• Ensuring there is a reference to a RDF/DublinCite format of the metadata

Whilst a dictionary can be created for GO terms and Taxon ID (either by searching and replacing from website automatically or by changing manually), to include the full protocol, scope, description and limitations of record it will likely be important to scrape information from the abstract and the materials and methods section for the publication associated with each record which can be accessed via its associated doi.

1. Finish uploading metadata
2. Understand how to ensure that the contents that were uploaded with pow save can be reincorporated back into the main distribution of PyOpenWorm
3. Commit the necessary changes to personal github fork
4. Write tests 5 Upload missing metadata possibly by scraping abstracts and protocols of associated publications
5. Upload actual data in a format which is useful to Data Science competitions
6. Pull request the changes back to the master PyOpenWorm repo
7. Work with maintainers to get the PR approved

Having uploaded movement metadata into PyOpenWorm and developed a proof of principle for conforming to the FAIR principles, the next step was to apply this to other data sources. Calcium imaging of neurons is a key tool for understanding neuronal activity. On electrical activation of a neuron, calcium levels can rise up to 10-100x their 50-100nM baseline levels. By linking calcium to chemical or genetically encoded fluorescent indicators and imaging the change in fluorescence we are able to monitor neuronal activation. Recently imaging advances have meant it is now possible to record transient intracellular calcium levels at a single cell resolution in the head of a freely moving C. elegans. This will make important advances in linking neuronal activity to movement data. The primary aim of OpenWorm is to create a model of the C. elegans nervous system from which we are able to predict behaviour. Having access to data which empirically links neuronal activation and movement will provide invaluable insight for both model creation and validation.

Having created categories for movement metadata which conform to the FAIR principles, extending these categories makes them suitable for increasing accessibility of metadata concerning simultaneous neuronal and movement experiments. Unlike movement data sources, accessible in Zenodo, this information is mostly collected from individual research groups and is not yet collated into one source. In addition, as simultaneous tracking of movement and neuronal activity is a relatively new technique, it is necessary to supplement neuronal activity collected via calcium imaging with data from sedated worms. Once the data sources had been collated, I followed the same steps as before for the movement metadata. One, create categories to conform to the FAIR principles; Two, harvest metadata; Three, standardise the metadata as required; Four, upload to OpenWorm; Five, create unit tests to check incorporation; Six, extend metadata as required.

OSF, like Zenodo, provides a platform for data sharing and includes key datasets concerning neuronal activity in C. elegans. I decided to split datasets into two divisions: Dynamical - calcium imaging with the head free to move and Fixed - where the worms are sedated.

Key non-dynamical/fixed data sources include:

• Data collated by Manuel Zimmer

Key dynamical data sources include:

• Data collated by Andrew Leifer

Whilst these sources are currently sufficient to start planning integration categories which conform to the FAIR principles, ultimately more sources of data, especially dynamical, will be required. Using Octoparse, web scraping of publications citing the Leifer dataset was carried out but was largely unsuccessful at returning new datasets of interest.

To integrate dynamical neuronal activity, an extension to the FAIR categories for movement metadata was decided as shown in Document 2. These categories will need to be made more robust by evaluating more publications.

1. Gather more data sources and collate in a central location
2. Increase the robustness of the FAIR categories for integration into PyOpenWorm (possibly extend WCON documentation too)
3. Harvest metadata (possibly via paper scraping from key words)
4. Write classes and integrate
5. Write tests for integration

This year, 2019, researchers finally completed the whole animal connectome of both C. elegans sexes. This will improve modelling of the nervous system whilst more detailed neuronal architecture may increase our understanding of the function of neurons within circuits. However, whilst the connectivity was mapped, no atlas of neuronal 3D coordinates currently exists. Previous attempts to build a 3D atlas have successfully segmented, identified and mapped 357/558 nuclei in the L1 stage. Using DAPI and RFP stained confocal Z stacks, this project attempted to apply the principles set out in Long et al. to produce a reference atlas of the 3D neuronal coordinates to which C. elegans phenotypes could be quickly and easily compared. Key steps include: straightening worms, marking nuclei centroids, aligning the z stacks and transforming the coordinate space. With advent of NeuroPal, a multicolour transgene able to resolve unique neuron identities, the creation of a more robust atlas should become easier.

Specimens were one day old C. elegans. Using a DAPI stain, which attaches to AT rich regions of DNA, nuclei were visualised and the worm captured as a confocal stack.

Individual stacks were visualised in Fiji (Fiji Is Just ImageJ). Stacks were imported through File -> Import -> Image Sequence and then the first image of the stack was selected. Settings were left on default.

3. Straightening

Although each stack contained only a single worm these worms were often in curved or random orientations making comparison between worms harder. To make identification of neurons and transformation easier the worms were first straightened using Fiji’s Straighten option.

Once the stack was loaded, we scrolled through the stack until a central slice was chosen which showed the width and length of the worm’s body clearly. If no such slice exists the contrast can be adjusted using Image -> Adjust -> Brightness/Contrast. To reset the Brightness/Contrast settings to the original press Auto.

Once the worm body is visible select the polygon symbol on the Fiji tool panel and draw along the length of the worm following its curves by creating sections of straight lines. Connect the final point to the first.

To create a line rather than polygon: Edit -> Selection -> Area to Line. To create a smoother centre line: Edit -> Selection -> Fit Spline.

You now have a spline mimicking the worms centre. Fiji’s Straightening tool uses this to perform straightening via spline interpolation. The Straighten tool creates another finely spaced spline (1 pixel width) from your spline (ROI). You are asked to assign a width to the spline (ROI) you drew, generally enough to cover the width of the C. elegans. Then for each small spline segment it interpolates, using a line perpendicular to that segment which is the width of the original ROI and is centred on the spline, from the original image to the final result. This creates a straightened worm.

Nuclei were marked manually as the centroid of a spot of DAPI fluorescence. To do this each section of the worm was considered in turn and spots which were the brightest throughout the stack for the same xy coordinates were considered the centroid. Only somatic/neuronal nuclei were marked. Nuclei within the gonad and embryonic region were left unmarked as these would not remain stable between worms. For example, everything in blue would be left unmarked in the figure below.

To accomplish the marking of neurons, we used the multipoint selection and zoomed in on each section. If neurons were difficult to visualise the contrast of the image could again be adjusted and reset as before: Image -> Adjust -> Brightness/Contrast

This eventually led to a stack of points (ROIs) representing the centre of each somatic/neuronal nuclei.

The next step is to save these points for further use. Pressing Ctrl + M (or alternatively Analyse -> Measure) allows you to save each yellow point’s x, y, z coordinates to a csv, alongside other information such as Area, Min and Max etc. To load the ROIs at a later date onto the image stack, they must also be saved as a .roi file through File -> Save As -> Selection. To load the .roi file at any time: first import the image stack (File -> Import -> Image Sequence, select the first image in the sequence), then load the .roi file through the ROI manager. Analyse -> Tools -> ROI manager. Once the box appears select More, then Open and then select the .roi file which you want to load.

However some worms may be bigger than others or in a different coordinate space. To solve this problem to subsequently compare the ROIs selected in each worm to decide which are stable and which may be unstable. To do this we used a plugin called BigWarp which uses an affine transform to transform one worm into the space of another. An affine transform preserves points, straight lines and planes and keeps parallel lines parallel.

BigWarp cannot deal with stacks so we define the coordinate space using two corresponding slices. Currently, we do not know which slice corresponds to another slice in a corresponding stack. For the sake of testing the script, I compared slice 13 of one RFP stack which I designated the reference to slice 13 of all other DAPI stacks.

Open the corresponding slices from each image through File -> Open and then open them both in BigWarp: Plugins -> BigDataViewer -> BigWarp. Specify which image will be your target/reference image and which will be your moving image (i.e. the image you transform). With both images on the screen, click on the moving image to make it active, press F8 and select Affine in the pop up window then close the pop up image. Press the space bar to enter landmark mode.

Select a point on the moving image and then select the corresponding point on the target image. Ensure that you are considering dorso-ventral and anterior-posterior orientation. Worms are not necessarily imaged in the same orientation. The excretory pore is a good marker for identifying dorso-ventral orientation. These are your landmark points. Continue with this until all your landmark points are selected.

Save the landmark file as a .csv file: File -> Export landmarks

Close all windows except for the orignal Fiji Toolbar. Press the [ key on your keyboard to open up a script interpreter

Open excel and save a blank worksheet as a .csv file.

Within the space under the New_ tab copy and paste this script. Set Language->groovy. Press Run and fill in the boxes in the pop up window.

Under landmark file, select the .csv file which you used to store the landmark coordinates from BigWarp. Create a copy of the .csv file you earlier created from the ROI (yellow dots) xyz locations for the moving image. Delete all columns other than X, Y and Slice and delete the header row. Use this newly formatted file for input points. For output points select the blank .csv file that you just created. Press OK.

Open the previously blank .csv file. This should now have 3 columns. The first corresponding to the transformed X coordinates of your ROIs, the second corresponding to the transformed Y coordinates of your ROI and the third a column of zeros. Replace the column of zeros with the Slice column from your original ROI file.

These are your transformed worm coordinates which can then be plotted using this script.

However, there are major flaws with this approach. Worms do not lie flat along the Z plane and so will need to be straightened along the Z axis. Current suggestions include identifying the four landmark points as the first visualisation of the head/tail and last visualisation of the head/tail through the image stack and applying a transform. Another issue is the reliability of centroid marking. This was done manually and may not lead to the correct centroid identification. The key issue, however, is that there no easy/obvious fiducial marker was identified to anchor worm stacks to each other. Currently, alignment is heavily reliant on being able to identify patterns of nuclei shared between worms.

Possible Fiducial Markers discussed for finding corresponding slices between stacks were:

1. Pharynx/other internal structures: unlikely to be useful in this as not fixed.
2. Vulva/Excretory Pore: may be more useful but not visible right now in current images (excretory pore is visible).
3. Discussed the use of myo-3:GFP as in Myers paper. Peter Sun, in Scott Emmons Lab, has created images with different fluorescence identifying different types of cells. Scott raised the point that the inside of the worm will be more squishy than outside.
4. Right now looking along the lines of being able to identify similar patterns in nuclei to align stacks: to aid in this future images will reduce the number of nuclei by switching from DAPI to less sensitive RFP.
5. No start or end point of worm stack/slices per stack seem to be more different than would be expected. This is an issue for identifying an accurate cuticle and an accurate cuticle might be a nice fiducial marker.

All 15 000 records are in the correct format to be uploaded to PyOpenWorm. The metadata still currently needs to be integrated fully and tests still need to be written. Missing metadata will need to be harvested and added to each record. In addition, a link to the movement data itself will be added.

More data sources need to be collated and added to a centralised storage site, following the example set for movement metadata. It may be worth creating a metadata record in WCON for each dataset? The FAIR categories currently decided will need to be made robust to the addition of more datasets. Once completed however, the integration with PyOpenWorm should follow the same principle as the integration of Movement Metadata.

Multiple confocal Z stacks have now been collected of 1 day old C. elegans. Currently, I am manually marking the xyz position of the centroid of each individual fluorescence which is believed to correspond to a somatic or neuronal nuclei. To increase standardisation and reduce the reliance on the human annotator, it would be good to find the centroid of each fluorescence automatically. Whilst worms have been straightened along the xy coordinate space, currently worms do not lie flat along the z plane. Straightening of the worm along the z dimension will need to be performed. Currently, there is no way to align confocal Z stacks slices to slices in other Z stacks. This makes it difficult to identify corresponding nuclei between stacks. Either a fiducial marker will need to be incorporated to act as an anchor between stacks or patterns of nuclei will need to be identified manually or via machine learning in 3D plots. The variability in annotation of the centroid increases the difficulty of identifying patterns of nuclei and so far manual identification has been unsuccessful.