Nitrogen-fixing symbiosis between legumes and rhizobia feeds a large share of the planet, yet the molecular interface of this partnership remains largely dark. Earlier, alfalfa mutants dnf1 through dnf7 were found with defects in nitrogen fixation, and for some genes the role is already known: DNF1 and DNF7 encode peptides that push bacteria to differentiate into bacteroids. The DNF2 protein looks like a phospholipase C, but its molecular function stayed a mystery. On the bacterial side, the nif and fix genes responsible for nitrogenase are known, along with only a few effectors, and a recent transposon screen pointed to hundreds of candidates. Secreted proteins are considered key mediators of the cross-kingdom dialogue, yet in rhizobia only a handful have been functionally characterized. The work sits at the edge of its field: the closest neighbors in the archive explore neural network predictions of protein interactions, food web resilience, magnetoreception in robins and energy transfer in plants, but no one had applied AlphaFold3 to the full rhizobial secretome against nodulation regulators.
First, a candidate secretome of S. meliloti 2011 was assembled: from the proteome, 640 proteins with a signal peptide and no transmembrane helices were selected, and 13 previously described effectors were added. Expression was checked against public transcriptomic data from nodules at 2-4 weeks after inoculation: 651 of 653 genes turned out to be active. On the plant side, 333 protein forms from 227 known nodulation regulators were taken. Then each bacterial protein x plant protein pair was run through AlphaFold3, over 217,000 complexes in total, on a Cambridge computing cluster; the screen consumed more than 10,872 GPU hours. For ranking, the metric 0.5 x pTM + 0.5 x ipTM was used, with a high-confidence threshold of 0.8. Selected interactions were tested in Nicotiana benthamiana leaves using split-luciferase and co-immunoprecipitation, and the role of bacterial genes was probed with deletion mutants and an acetylene test for nitrogenase activity.
At the strict threshold of 0.8, AlphaFold3 predicted 308 interactions between 189 secreted bacterial proteins and 68 alfalfa proteins (89 variants). The network turned out not to be uniform but hub-based: the DNF2 protein became the main node with 32 predicted partners, followed by NPD5 with 20. Among bacterial proteins, SRP485 (a putative substrate-binding protein of a phosphonate ABC transporter) and SRP86 (a protein with a domain of unknown function, DUF2155) stood out. Experiments confirmed the physical interaction of DNF2 with SRP86 and SRP485: the split-luciferase signal was strong, and co-immunoprecipitation worked. DNF2 localizes to the peribacteroid space, surrounding differentiating bacteria. Single deletions of srp86, srp256 and srp485 caused only a slight reduction in shoot length and a small rise in the share of white nodules, but the double mutant srp86/srp485 produced a mass of small white nodules, yellow leaves and a drop in nitrogenase activity. Transcriptomic analysis showed that the double mutant reproduces 52% of the genes suppressed in dnf2, and only 27% of the defense genes activated in dnf2. In other words, losing the bacterial proteins breaks the symbiotic program but does not trigger the acute immune response seen in the plant mutant.
The work turns DNF2 from a protein of unclear function into a cross-kingdom hub that binds several bacterial proteins at once and sustains the symbiotic accommodation program. Unlike neighboring studies, where AlphaFold was applied to individual effectors or pathogenic systems, here the full rhizobial secretome was run against the repertoire of plant regulators for the first time, yielding a map of hundreds of candidates. This is not just a list but a testable resource: two predicted contacts have been confirmed experimentally and functionally. The division of labor within the DNF2-SRP module is also new: bacterial proteins are needed to maintain leghemoglobins, NCR peptides and nitrogen fixation, whereas immune suppression remains DNF2's own job. Such a scenario was invisible in earlier data from the plant mutant alone.
The open database of predicted interactions lets any researcher check their own protein pair without costly screening. The next step is to test the remaining contacts of DNF2 and other hubs, especially given that AlphaFold3 can produce false positives due to mismatched localization or expression timing. The conservation of DNF2 and SRP485 in pea and soybean hints that the module can be tested in other host-symbiont pairs. A separate question is how exactly DNF2-SRP regulates lipid signatures on the symbiotic membrane, since phosphatidylinositol-4,5-bisphosphate accumulates during symbiosis but not during oomycete infection.
The resource will be useful to plant molecular biologists, microbiologists and agrobiotechnologists looking for candidate genes to improve nitrogen fixation. The method is transferable to other cross-kingdom interfaces, from pathogens to insect symbionts.
Test the remaining predicted partners of DNF2 and other hubs using co-immunoprecipitation and deletion analysis. Investigate exactly how DNF2-SRP affects the lipid composition of the peribacteroid membrane.
For symbiosis biology, the question remains unresolved: how does a plant distinguish a beneficial rhizobium from a pathogen at the molecular level, and why does immunity break down in dnf2 but not in srp86/srp485? The work shows that the defensive and symbiotic branches of DNF2 are separable, but the mechanism of that separation is unknown. Also open is the question of which bacterial signals trigger bacteroid differentiation and how the secretome of different strains affects compatibility with the host.
The screen of 217,000 protein pairs took more than 10,872 GPU hours, which is over a year of continuous work for a single NVIDIA A100-class graphics card. Meanwhile, the entire rhizobial secretome is active in the nodule: 651 of 653 genes are transcribed simultaneously, as if the bacterium keeps a whole arsenal of tools at the ready.
Not published on arXiv. Here are its page, the PDF and the formalization: you can check it yourself, not only read our account.