Marcus Povey

As part of the development of ARIA, we’ve spent a lot of time integrating data from a wide range of scientific resources. Whether we’re harvesting experimental metadata from a visit, tracking the proposal as it passes through our access platform, or linking to outputs and depositions in repositories, it quickly became clear that each technology or technique had their own view of the world.

Think about a protein for a moment.

You might be interested in its structure, which might have one set of identifiers, models and resources. You might be interested in the way it moves or the way it binds, which has another set. You may be interested in the chemistry. You may be used to a specific imagine technology (e.g. x-ray), and so might be unfamiliar with what is available from NMR. You may be interested in AI models. Again, all separate identifiers, keys and resources.

It might have a UniProt accession, several experimentally determined structures in the Protein Data Bank, predicted models in AlphaFold, NMR data in the BMRB, cryo-EM maps in EMDB, domain annotations from InterPro, publications referenced by DOI and PubMed, and authors identified by ORCID.

To put it simply, the data is there, but the connections are not. In order to get a complete picture of the science and to be able to reuse and build on what has been done before, it is necessary to be able to find these data. Valuable research sitting somewhere on a hard drive is useless unless we can know what it is, where it is and how it connects to the underlaying fundamental research.

Now consider a fragment screening campaign.

A screening project might begin with a protein target identified by UniProt. Experimental structures deposited in the PDB reveal fragment binding sites. Ligands are represented using SMILES, InChI and ChEBI identifiers. Publications describe the work. Funding is linked to Horizon Europe projects. The researchers involved have ORCID identifiers and belong to organisations identified by ROR.

Each of these repositories are authoritative within its own domain, but none of them are (or should be) responsible for mapping the entire landscape.

Thinking beyond identifier lookup

At Instruct, myself and my team have been building out a data management and visualisation platform for proposals in our platform, linking them to experimental metadata from the facilities, and research outputs. The end goal is that this will allow the researcher to better understand the science, more easily expand on what has been done before, and produce useful training data for AI enhanced drug discovery.

This platform requires good search, but it became clear to me that what people would be looking for may well be different depending on the background of the researcher. Initially I started thinking about building another identifier resolution service. The idea was straightforward, given a UniProt accession, discover the associated PDB entries, AlphaFold models and other related identifiers. However, the more I explored the problem, the more I came to realise that the individual identifiers, while important, were not the interesting part.

The interesting part was the network the relationships between the identifiers represented, and instead of asking “What is the PDB identifier for this protein”, we can begin to ask “What do we know about this biological entity, and how can Instruct enhance that?”

This changes the framing from resolver to knowledge graph. The knowledge graph does not replace these resources, instead, it records the relationships between them and makes those relationships easier to discover and reuse. This aligns closely with the FAIR principles we’re all meant to be following as good scientific citizens. Existing repositories continue to curate their specialist domains, while the discovery layer improves interoperability and reuse across the wider ecosystem.

Introducing Constellation

So, as an experiment, I built a thingy.

Constellation (working title) is a simple tool – given an identifier (PDB, UniProt, etc), it will go away and query the respective endpoints and pull out information about the key (e.g. details on the protein). It then looks at the return data, and if it finds keys in there, it adds them to the search queue.

Over time, this builds up a graph of related information, with the type of relationship between nodes preserved to provide valuable context. A single UniProt accession might discover several PDB structures. Those structures might reveal ligands. Those ligands might resolve to ChEBI and ChEMBL records. Publications associated with the structures can be linked through DOI and PubMed. Authors can be connected through ORCID, institutions through ROR, and funding through CORDIS. Not only will this allow you to identify differing keys that refer to the same biological entity which can be used as additional search terms, but it allows you to expand what you know and provide valuable context to data sets as well as aid reproducibility.

Long term readers of this blog may remember I actually started kicking this idea around some time ago.

Supporting Instruct

Take for example a fragment screening campaign. Such a campaign naturally brings together structural biology, chemistry and metadata from multiple independent resources.

A fragment itself may be represented by SMILES, InChI, InChIKey, PubChem and ChEBI identifiers. The target protein has a UniProt accession. Crystal structures showing fragment binding are deposited in the PDB. NMR data may exist in the BMRB. Experimental metadata, publications and project information all live elsewhere.

The landscape is very fragmented, if you excuse the pun, and so connecting these data together can be tricky. With a graph, at least, we only need to “know” about a single entry node. For some of this data, we may be directly informed by our facilities as part of Instruct’s normal data management activities, but with a graph we can expand on what we know, providing a more complete picture.

After research has been published, we may be able to “backfill” records to measure impact and reach of the research conducted by our facilities, since again, all we need is to know about one “node” on the graph.

Looking ahead

It’s still very much an experiment, but it’s already earning its keep, enriching our project data pages with cross-domain search and resource discovery that didn’t exist before. The next steps are to push further into chemistry and molecular dynamics identifiers, and to build out better ways of measuring impact and capturing citations.

Go back to that protein we started with, the identifiers were always there, scattered across a dozen authoritative sources that each did their own job well. What was missing was the map between them. That’s what Constellation is trying to be: not another repository, but the connective tissue that lets researchers ask “what do we know about this” instead of “what’s the PDB ID for this,” and actually get an answer.