AI Reads Microbial Genomes to Forecast Life in Earth’s Saltiest Places

Some of the most forbidding landscapes on the planet are, in fact, crowded with life. From the shores of the Dead Sea to the shimmering solar salt pans where seawater is evaporated to industrial salinity, hypersaline environments host thriving communities of bacteria and archaea that have evolved remarkable biochemical strategies for coping with salt concentrations that would rapidly kill most known organisms. Yet a fundamental obstacle has long stood in the way of understanding these extremophiles: the overwhelming majority of microorganisms on Earth resist every attempt at laboratory cultivation. Researchers estimate that more than 81 percent of the world’s microbes cannot be grown under standard laboratory conditions, which means that for most species, scientists have no direct way to measure how salty an environment they can tolerate, what substrates they prefer, or how they would behave in a given habitat. A new computational tool developed in China now promises to bypass that bottleneck entirely by reading the answer directly out of an organism’s genome.

The tool, called SuSha, was created by a team at Nanjing Tech University and is described in a peer-reviewed paper published in the KeAi journal Engineering Microbiology. At its core, SuSha is a machine learning framework designed to predict a microorganism’s salt tolerance from its genetic blueprint alone, without ever requiring the organism to be cultured in a dish. The work addresses a problem that has grown increasingly urgent as metagenomics has transformed microbiology. Modern sequencing technologies allow researchers to reconstruct partial or even near-complete genomes of uncultured microbes directly from environmental DNA samples, generating vast catalogs of genomic data from habitats ranging from ocean water to deep subsurface brines. But a genome sequence on its own says little about how the organism actually lives. Inferring ecological traits, such as salinity preference, from sequence data has remained a stubborn challenge.

Earlier attempts to solve this problem typically relied on the presence or absence of individual genes associated with salt tolerance. Microbes that thrive in salty conditions often carry genes for specialized ion pumps, compatible solute synthesis pathways, and other osmoadaptation mechanisms, and researchers reasoned that detecting such genes could serve as a proxy for salt tolerance. The difficulty, as the SuSha team recognized, is that this single-gene approach frequently fails to capture an organism’s true behavior in complex natural habitats. Salt adaptation is not the product of one gene acting in isolation; it is a systems-level property that emerges from the coordinated action of hundreds of proteins, each of which must remain stable and functional at particular salt concentrations. A microbe might carry a salt-tolerance gene yet be poorly adapted overall, or conversely, might tolerate salt through strategies that a narrow gene screen would miss entirely.

SuSha takes a fundamentally different approach by examining the whole genome through the lens of protein chemistry. The researchers started with a training set of 123 bacterial and archaeal species whose salinity preferences had already been well established through prior study. For each of these organisms, they analyzed the amino acid composition of the encoded proteome, the complete set of proteins the genome specifies. This choice is grounded in evolutionary biochemistry: proteins from salt-loving organisms tend to exhibit distinctive compositional biases, such as an enrichment of acidic amino acids and characteristic patterns in the overall balance of amino acid classes, because their proteins must fold and function properly in briny cytoplasm. From this analysis the team built a 24-dimensional fingerprint for each genome, capturing both the frequencies of individual amino acids and broader chemical properties of the protein complement. The result is a quantitative signature that reflects the evolutionary pressure a lineage has experienced, distilled into a form that machine learning algorithms can interpret.

On top of this fingerprint, the researchers deployed an ensemble learning strategy. Rather than betting on a single algorithm, SuSha combines several machine learning methods, including random forest, bagging, and extra trees, into a single integrated model whose predictions aggregate the strengths of its components. This ensemble approach proved effective. In benchmark testing, SuSha achieved a global accuracy of 0.765 and an area under the receiver operating characteristic curve, or AUC, of 0.941, outperforming any of the individual models on their own. An AUC approaching 0.94 indicates that the framework distinguishes salt-tolerant from salt-sensitive organisms with a high degree of reliability across the full spectrum of salinity preferences, from organisms that prefer near-freshwater conditions to those that flourish in near-saturated brines.

The real test, however, came outside the training data. The team applied SuSha to 2,678 genomes that had been reconstructed from metagenomic samples collected across six regions worldwide, spanning freshwater, marine, and hypersaline habitats. These metagenome-assembled genomes represent uncultured organisms whose lifestyles are largely unknown, making them an ideal proving ground for a tool whose entire purpose is to infer traits without cultivation. The results were striking. In crystallizer ponds, where salinity exceeds 35 percent and only the most extreme halophiles can survive, SuSha correctly identified more than 90 percent of the resident microbes as extreme halophiles. At the opposite end of the spectrum, in the freshwater sediments of Lake Baikal, the world’s deepest lake, the tool accurately flagged salt-sensitive organisms while avoiding false alarms, correctly refraining from labeling freshwater specialists as salt tolerant.

One of the more sophisticated aspects of the framework is that it does not apply a single one-size-fits-all algorithm across all environments. Instead, SuSha switches between strategies depending on the habitat being analyzed. For oceans and freshwater environments, it employs a robust tree-based ensemble, while for near-saturated habitats it uses a linear ensemble. This adaptive design reflects the reality that the genomic signatures of salt adaptation may differ between moderate and extreme salinity regimes, and that prediction methods optimized for one end of the salinity gradient may not perform well at the other. According to Jun Zhou, the study’s corresponding author from the College of Biotechnology and Pharmaceutical Engineering at Nanjing Tech University, the goal was reliability grounded in biology. Crucially, he notes, the tool’s predictions are transparent and traceable back to specific genomic features, which means each result can be tied to real evolutionary signals rather than emerging as an opaque output of a black-box model.

Salinity adaptation is a systems-level trait shaped by the entire genome, not by a single gene, Zhou explains. By reading the amino acid signature of a whole genome, SuSha allows researchers to infer how a microbe that has never been cultured would cope with salt, across the full range from near-freshwater conditions to saturation. This capability has implications that extend well beyond answering an academic question about microbial ecology. Salt-loving microorganisms are a rich source of salt-tolerant enzymes, proteins that remain active and stable in high-salt conditions where conventional enzymes denature or lose function. Such enzymes are valuable in industrial biotechnology, where high-salinity processes are common, and in the treatment of saline wastewater, an increasingly important challenge as industries seek to manage brines generated by desalination, chemical manufacturing, and resource extraction.

The practical promise of SuSha lies in how it reshapes the search for these biotechnological assets. Traditionally, identifying a microbe with useful salt-tolerant enzymes has required laborious screening, growing candidate organisms under varying conditions and testing their enzymatic output, a process Zhou characterizes as blind trial-and-error. With a tool that can scan enormous genomic datasets and pinpoint the most promising organisms first, the search becomes targeted rather than exploratory. Researchers can prioritize which of the thousands of uncultured genomes in public databases deserve the effort of developing new cultivation methods or synthetic biology approaches, dramatically accelerating the pipeline from raw sequence data to industrial application. In an era when genomic databases are doubling at a staggering pace, tools that convert raw sequence information into ecological and functional insight are becoming essential infrastructure for microbiology.

The broader significance of the work lies in what it suggests about the future of studying unculturable life. Most of Earth’s microbial diversity remains inaccessible to classical microbiology, and metagenome-assembled genomes continue to accumulate faster than researchers can characterize them. Frameworks like SuSha demonstrate that meaningful physiological traits can be extracted from genomic sequence alone, provided the underlying biology, in this case the compositional signature that salt stress imprints on proteins, is well understood and translated into appropriate computational features. By grounding its predictions in traceable genomic evidence rather than opaque correlations, SuSha offers a template for building similar tools to predict other hard-to-measure traits, from temperature preference to pH tolerance. For the microbes of the Dead Sea, the solar salt pans, and the countless hypersaline niches in between, the era of invisibility may finally be drawing to a close, one genome at a time.

Subject of Research: Machine learning prediction of microbial salinity adaptation from whole-genome amino acid composition

Article Title: New AI tool predicts how microbes survive in Earth’s saltiest habitats

Article References: New AI tool predicts how microbes survive in Earth’s saltiest habitats. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: halophiles, extremophiles, machine learning, metagenomics, salt tolerance, microbiology, genomics, amino acid composition, ensemble learning, hypersaline environments, biotechnology, Engineering Microbiology