Predicting effectors is the easy bit. Knowing which one to chase is not.

Fusarium oxysporum is a species complex spanning lifestyles from devastating crop pathogen to harmless soil saprotroph. Within that complex, isolates that look all but identical will take out banana, tomato, onion and brassica, each strain locked to its own host. Sitting alongside those pathogens are endophytic Fusarium strains that live in the same plants and cause no disease at all. Some are close enough to be near-indistinguishable by sequence. So “this gene is in the pathogen” is not a useful filter. Almost everything is in the pathogen. Most of it is in the harmless relative growing next to it.

Over the past year, a contract research collaboration between the Client Company, Rothamsted Research and KnetMiner Ltd has worked that problem at scale. The Client funded the work and set the scientific questions, and the Rothamsted Bioinformatics team built a 71-genome comparative pangenome, a full secreted-protein and effector prediction pipeline, and RNA-seq expression profiling across three species, in and out of their hosts, totalling 22 comparisons.

The pain of having all the data with no answer in sight

Despite the months of comprehensive bioinformatics analyses, there was still no single answer to the question that mattered: of the hundreds of predicted effectors, which handful should go into the lab? Instead, the researchers were left with a directory with seventy-one folders of EffectorP outputs, a 40 MB orthogroup table, twenty-two DESeq2 spreadsheets, BLAST outputs, and a naming convention that changes depending on which tool wrote the file. The transcriptomics uses gene IDs, the effector prediction uses protein IDs, and the orthology pipeline prefixes both with its own genome codes.

Although the data can answer a thousand different questions a biologist would pose, finding the answer is not a query anybody can run. Every filter that would narrow the list (conserved in the pathogens, absent from the endophytes, induced during infection, related to a known virulence gene) lives in a different file. It takes days of deciphering different formats and outputs, and of joining tables into large, clunky spreadsheets of genes of interest, then repeating the analysis from scratch the next time someone changes the question slightly. So the shortlist ends up chosen on the evidence that was easiest to join, not the evidence that mattered.

One gene, every layer of evidence attached

This is what a knowledge graph is for. Each analysis stops being a file and instead becomes edges on a gene.

The orthology work lands especially cleanly, because KnetMiner 1.5 rebuilt its own cross-species links on hierarchical orthologous groups.

Then you can just ask it

This is where the KnetMiner team came in: integrating the new data, generating and deploying the knowledge graph, and onboarding and training the researchers to use it. Researchers can now use manual and agentic search to query the data in structured, logical ways. They type a question and the agent plans the search, picks its own tools, queries the graph and the APIs, and answers from what is actually in the data rather than what a language model happens to remember.

On this dataset, that means a data query that once took days becomes a single sentence:

Then come the follow-ups, which is where the real value lies:

None of these requires knowledge of the file layout, the ID conventions or the join logic.

Watch: KnetMiner Agentic in under two minutes. A demo on KnetMiner’s fungi knowledge graph, from a plain-language question to ranked genes, evidence and the network.

The genes are also connected to the full KnetMiner fungi knowledge graph, which connects to model organisms, PubMed, UniProt and around twenty other sources, all within the same conversation.

Maintaining usable data

Many research groups have a version of this problem. The pangenome was run by one person, the RNA-seq by another. Each analysis is sound, but none of it can be queried together, and the institutional memory of how to join it up walks out of the door with whoever did the analysis.

Integration is the step that turns that into an asset with a shelf life. It is also the step most easily left undone, because it is nobody’s paper.

This collaborative project is a great demonstration of the impact when integration does happen. From the connected data and the powerful search, we were able to generate a prioritised candidate gene list that is now moving forward for further testing.

About the collaboration

The project was contract research funded by the Client Company, who defined the scientific questions. The bioinformatics, covering the pan-genome analysis, effector prediction and gene expression analysis, was carried out by Dr Erika Kroll and Dr Ange Zoclanclounon of the Rothamsted Research Bioinformatics team. The KnetMiner team integrated the results, generated and deployed the knowledge graph, and onboarded and trained the researchers. The Client had early access to the integrated graph, which is now also available to other organisations on a subscription basis.

If you have multi-omics output sitting in a directory that only one person knows how to navigate, and you want to see what it looks like as a knowledge graph you can interrogate in plain language, we would be happy to talk. Contact us.

‍