Artificial intelligence promises to dramatically expand the number of potential biological crop-protection products researchers can discover. But for Soilytix co-founder and chief scientific officer Tim Rajakumar, that presents the industry with a problem as well as an opportunity.
“One of the biggest problems we hear is the validation bottleneck,” he told AgNavigator.
“AI is producing very large numbers of potentially testable candidates. But if they all still have to pass through expensive laboratory, greenhouse and field experiments, simply generating more candidates can make the bottleneck worse.”
The Hamburg-based biological intelligence and discovery company believes its newly launched LOAM genomic AI models can help tackle that problem.
LOAM was trained on 15,640 microbial genomes reconstructed from long-read sequencing of soil, sediment and water, representing approximately 67.5 billion DNA bases, to help scientists identify which genes and proteins found in soil are the most promising candidates for development into new biological crop-protection products.
The aim is not to replace experimental validation. Instead, Soilytix wants to progressively narrow the candidate pool before companies commit resources to expensive, lower-throughput tests.
“The need to enrich the pipeline before those expensive experiments begin is an idea that is resonating strongly with the companies we speak to,” said Rajakumar.
Hunting for natural disease suppression
Rajakumar gives the example of a crop-protection company looking for new antifungal peptides targeting a particular Fusarium species.
Soilytix could begin by searching its field database for soils in which the pathogen is present. Researchers would compare fields where disease occurred with those where Fusarium was detected but disease did not develop.
That difference provides what Rajakumar calls a “disease-suppression signal”, allowing researchers to concentrate their search on microbial groups potentially responsible for suppressing the pathogen.
Soilytix can then interrogate genomic information from those organisms and, where necessary, generate additional data from disease-suppressive soils using long-read metagenomic sequencing.
A version of LOAM fine-tuned for antifungal discovery can subsequently search those genomes for promising peptides.
Candidates are filtered computationally before a much smaller selection progresses into initial experimental screening using cell-free protein synthesis. The strongest candidates could then be handed to crop-protection partners for more extensive in-vitro, greenhouse and ultimately field testing.
“The Soilytix platform aims to combine agronomic field evidence, microbiology, long-read genomics and AI to progressively enrich the candidate pool before expensive, lower-throughput experimental testing,” Rajakumar explained.
‘Which product works in which soil?’
Candidate discovery is only one part of the challenge Soilytix wants to address.
Rajakumar said biologicals companies also regularly encounter inconsistent field performance, with products such as microbial inoculants, seed coatings and biostimulants working well in one location but delivering much weaker results elsewhere.
Soilytix believes some answers could come from an unlikely source: personalised medicine.
Rajakumar and Soilytix CEO Bruno Steinkraus previously worked with molecular biomarkers and patient stratification in oncology.
“The question then was: which patient is most likely to benefit from which treatment?” said Rajakumar. “We see a strong parallel in agriculture, where the question becomes: which biological product is most likely to work in which soil, crop and environmental context?”
The company already works with businesses conducting R&D trials on new products, analysing biological markers in soils to investigate why performance varies between trial locations.
The longer-term opportunity is to use this information not only during product development but to help companies determine where their products are most likely to work, potentially improving product positioning and sales.
Wet-lab validation is the next test
There is an important caveat to Soilytix’s proposition: its crop-protection discovery approach has not yet produced wet-lab validated candidates.
Rajakumar said its computational pipeline from field data through to in-silico validated antifungal peptide candidates is operational and LOAM has performed strongly in biological benchmarks.
The company is now collaborating with the University of Hamburg on its first in-vitro validation work.
That validation will be important to establishing whether Soilytix can translate computationally promising candidates into biological activity and, longer term, whether its approach genuinely reduces the number of candidates companies need to test.
2027 pilots to put the economics to the test
Soilytix plans to take the model into discovery pilots in 2027, targeting crop-protection companies with a defined biological discovery problem and existing capabilities to take promising leads through downstream development.
“We see ourselves as an upstream discovery platform that complements the downstream development infrastructure that crop-protection companies already have,” Rajakumar said.
Crucially, the company is not yet putting a number on how much money or time this enrichment process could save.
Rajakumar said Soilytix does not yet have sufficient prospective data to quantify how much its approach can improve the hit rate. Measuring that enrichment will be one of the key objectives of its first discovery pilots.
There could also be another benefit. When a promising biological candidate fails because of characteristics such as poor stability, solubility, expression or environmental persistence, experimental results could be fed back into Soilytix’s computational models to identify variants intended to retain the desired biological activity while improving their suitability as commercial products.
“The value will come from both sending fewer, more highly enriched candidates into expensive downstream testing and extracting more value from promising biological starting points once they have been identified,” Rajakumar said.
