Early-stage microbial discovery has historically been rate-limited by phenotypic screening: you grow an isolate on selective media, observe whether it survives a stress condition, then go back and ask why. That workflow produces confirmed candidates but at throughput that made large diversity surveys impractical for a small research group. Working with hundreds of isolates from a single field expedition was achievable; working with thousands was not, not if each one required manual assay time before you knew whether its genome was worth sequencing.
What changed is not simply computational speed, though that matters. What changed is that gene family databases for plant growth promotion and stress tolerance pathways have become specific enough that gene-based screening ahead of phenotype is now a legitimate first-pass filter rather than a speculative shortcut. We use this approach throughout our pipeline, and I want to explain what it actually involves and where its limits are.
From Isolate to Genome: the Input Side
The process starts with field isolates, not cultures downloaded from a public repository. After collection and plating on appropriate selective media, isolates that show initial growth at our target salinity or cold thresholds go through DNA extraction and whole-genome sequencing. We use short-read sequencing for initial screening because the turnaround time is shorter and the per-isolate cost is low enough to process batches of 50 to 100 genomes at a time. Assembly quality matters less at this stage than gene-calling quality: we care about the gene content, not the chromosome-level architecture.
Assembled genomes are annotated using Prokka or a comparable automated pipeline, which assigns functional annotations based on sequence similarity to curated protein databases. The annotation output is what the downstream screening step works from.
The Screening Query Layer
The core of our AI-assisted screening is a set of queries against annotated gene content. We maintain a curated panel of gene families that correspond to mechanisms we care about: halotolerance pathways (nhaA, ectA/B/C, betA, compatible solute transporter families), cold-shock response elements (cspA-family RNA chaperones, DEAD-box helicases, fatty acid desaturases linked to membrane fluidity), plant growth promotion functions (IAA biosynthesis genes, acdS for ACC deaminase, phosphate solubilization via pqq gene cluster, nitrogen fixation markers), and biocontrol-relevant pathways (antimicrobial peptide gene clusters, siderophore biosynthesis).
A simple Boolean query (strain X contains gene Y: yes/no) produces a first-tier ranking. But we push this further by asking about gene copy number, homolog diversity within the cluster, and whether flanking regulatory elements are recognizable. A strain with three independent ect homologs in distinct genomic contexts is almost certainly a more committed osmotic adjustment specialist than one with a single low-identity homolog. This kind of quantitative gene family scoring is where the machine-learning layer adds the most value: rather than manually interpreting a matrix of gene presence/absence scores for 200 isolates, a trained classifier can weight these features against each other and output a ranked prioritization that captures multi-gene patterns a simple threshold would miss.
What the Classifier Was Trained On
I want to be transparent about this because it matters for interpreting our results. The classifier we use internally was trained on a combination of publicly available genomes with known plant growth promotion phenotypes and a smaller set of our own validated isolates from earlier screening rounds where we had phenotypic ground truth. The training set is not large by deep learning standards: a few hundred labeled examples. We use a gradient-boosted tree approach rather than a neural network, partly because the feature space is interpretable (each input feature corresponds to a specific gene family score) and partly because the smaller training set is less likely to overfit to a model architecture with tens of millions of parameters.
The output is a probability score that we treat as a prioritization tool, not a binary classification. An isolate scoring above our current working threshold goes to the phenotypic screening queue. An isolate below threshold is not discarded from the archive, but it does not get immediate lab attention. We recalibrate the threshold every few months as our ground-truth phenotype dataset grows.
Where Genomic Screening Misses
Genomic inference has real gaps that we do not paper over. The clearest one is regulatory context: a strain may carry all the biosynthetic genes for ectoine production but have transcriptional repressors or sigma factor mismatches that mean those genes are never expressed under the conditions we care about. Gene presence is a necessary but not sufficient condition for functional activity. We catch some of this by looking at regulatory gene neighborhoods, but the prediction is inherently less reliable than a biochemical assay.
The second gap is horizontal gene transfer creating genomic islands that look functional but are not integrated into the organism's physiology in a useful way. A halotolerance gene cluster acquired through lateral transfer may be expressed but may not be the dominant stress response pathway in an organism whose core physiology is tuned for a different environment. These cases are harder to detect computationally and are part of why phenotypic validation is not optional, regardless of how high a strain scores genomically.
The third gap is novelty. Our classifier was trained on known gene families. Genuinely novel mechanisms of stress tolerance in understudied extremophile lineages from Patagonian soils may not have recognizable homologs in our panel. We may be systematically underprioritizing organisms with interesting biochemistry that happens to be divergent from what is already characterized. This is a real and uncomfortable limitation, and it is why we maintain a random sampling protocol alongside the priority-ranked queue: a fraction of every batch goes to phenotypic assay regardless of genomic score, specifically to catch what the classifier would miss.
Speed and Scale in Practice
The practical throughput gain from running genomic screening before phenotypic assays is substantial. With our current setup, we can process the gene family scoring for a batch of 100 assembled genomes in a few hours of compute time, compared to several weeks of bench assays to run the equivalent panel of phenotypic tests on the same number of isolates. This does not eliminate bench work: it concentrates bench work on the isolates that genomic analysis suggests are worth the time. In practice, genomic pre-screening reduces the number of isolates entering phenotypic queues by roughly two-thirds in a typical batch, which is the difference between our current team being able to run the pipeline and not.
It also surfaces diversity patterns that would be invisible in a phenotype-first workflow. When you score 200 isolates for 40 gene family features simultaneously, you can see which functional combinations appear together, which organism lineages tend to carry overlapping trait profiles, and which field collection sites show the most promising gene content. This shapes where we direct future sampling effort, which closes the loop between computational analysis and field strategy in a way that phenotypic screening alone cannot do at our current scale.
Where the Work Remains
Genomic screening is a fast and informative filter. It does not make the downstream work easier; it makes the downstream work more precisely targeted. A strain that scores well across our gene panel still needs to demonstrate actual plant growth promotion in root colonization assays, stress tolerance in direct germination trials, and stability over multiple generations before it is worth the investment of moving to greenhouse scale. We are not claiming that our AI pipeline produces candidates; it produces a ranked shortlist. The candidates are produced by everything that comes after.