The research team, led by Defang Liu, Jiao Yan, Ting Luo and colleagues at the General Hospital of Western Theater Command in Chengdu, China, began with a rich resource: the Gene Expression Omnibus dataset GSE46480, containing peripheral blood mononuclear cell samples from 98 healthy volunteers after acute high-altitude exposure and 98 matched sea-level controls, profiled on the Affymetrix Human Genome U133 Plus 2.0 array. Applying rigorous differential expression analysis with the limma framework, they identified 1,103 differentially expressed genes, of which a striking 1,063 were upregulated and only 40 downregulated at altitude. Gene Ontology and KEGG enrichment analyses revealed that these genes clustered around ribonucleoprotein complex biogenesis, cytoplasmic translation, ribosomal structure and electron transfer activity, alongside a neurodegeneration-related pathway, painting a coherent picture of hypoxia disrupting protein synthesis and energy handling in immune cells.

On the intervention side, the team turned to the drug–food homology tradition rooted in the ancient Chinese medical text Huang Di Nei Jing, which holds that food and medicine share a common origin. They compiled the complete list of 106 substances officially recognized by the Chinese National Health Commission, then mined the Traditional Chinese Medicine Systems Pharmacology database for their chemical constituents. Applying conventional filters of drug-likeness of at least 0.18 and oral bioavailability of at least 30 percent yielded 610 candidate active components. The critical step came next: rather than relying on static databases, the researchers used GraphBAN, a graph neural network trained on the BindingDB, BioSNAP and KIBA datasets, to predict compound–protein interaction pairs with probability scores above 0.7. Intersecting predictions from the three pretrained models produced 754 high-confidence pairs linking 95 candidate compounds to 113 candidate genes.

Not every predicted binder deserves attention, however, so the team subjected the 95 compounds to a two-stage gauntlet on the ADMETlab3 platform. First came physicochemical assessment against Lipinski’s Rule of Five, evaluating molecular weight, hydrogen bond donors and acceptors, and lipophilicity, plus topological polar surface area as a proxy for membrane permeability; 27 compounds passed without violations. Then came toxicology: predictions for hERG cardiac channel inhibition, hepatotoxicity, drug-induced liver injury, carcinogenicity and acute oral toxicity, with compounds retained only if every toxicity probability fell below 0.5. Just 8 components survived this screen, and 6 cleared both hurdles, a high elimination rate the authors say underscores the value of multistage safety filtering before any candidate is proposed for human use.

To distill the gene list further, the researchers deployed a three-algorithm machine learning pipeline. A random forest model ranked genes by importance, the Boruta algorithm confirmed features with statistically robust relevance, and LASSO regression with 10-fold cross-validation selected a third set. The intersection of all three feature sets yielded just five genes: DNAJC19, MRFAP1L1, TBPL1, RPS21 and NABP1. Strikingly, all five traced back through the interaction network to a single drug–food homologous substance, Panacis Quinquefolii Radix, better known as American ginseng, and to two of its polyacetylene constituents designated PQ-2 and polyacetylene PQ-2, a pair of epimers whose stereochemical differences subtly shifted their predicted lipophilicity and toxicity profiles.

External validation added nuance to the story. In two independent datasets comparing samples at moderate and high altitude, expression trends were consistent with the discovery cohort, but only RPS21 reached significance in one, while a random-effects meta-analysis across both external datasets confirmed significant downregulation of MRFAP1L1, TBPL1 and NABP1 with essentially no inter-study heterogeneity. DNAJC19 showed a consistent downward trend of marginal significance, and RPS21 proved heterogeneous across cohorts, suggesting its expression may be swayed by study-specific factors such as altitude gradient, exposure duration and genetic background. The authors are candid that this inconsistency is a limitation, not a footnote.

The candidate genes were not random stragglers; they converged on a coherent functional architecture. Gene set enrichment analysis showed all five participating in 32 core pathways spanning ribosome function, oxidative phosphorylation, the spliceosome, protein export, DNA repair and apoptosis, with pathways such as apoptosis upregulated and axon guidance downregulated in the altitude group. The authors propose a division of labor: DNAJC19, a mitochondrial inner membrane chaperone, anchors energy metabolism; RPS21, a small ribosomal subunit protein, governs translation; TBPL1, part of the transcription factor IID complex, links to pulmonary fibroblast biology; MRFAP1L1 connects to chromatin regulation and genome stability; and NABP1 ties into RNA processing and complement–coagulation pathways. A nomogram built from the five genes distinguished altitude samples from controls with an area under the curve of 0.95 and excellent calibration, hinting at diagnostic potential.

Immune infiltration analysis deepened the physiological picture. Comparing 28 immune cell types between groups revealed a Th1/Tc1-skewed response at altitude: gamma delta T cells, effector memory and activated CD4 T cells expanded, while central memory CD8 T cells, natural killer T cells and CD56dim natural killer cells contracted. All five candidate genes showed significant correlations with multiple differential immune cells, with correlations reaching as high as 0.905 between MRFAP1L1 and activated CD4 T cells. The pattern suggests a statistical link between these genes and excessive inflammatory activation, weakened immune memory and cytotoxic injury, precisely the immune imbalance that characterizes high-altitude pathogenesis.

Finally, the team tested whether the two ginseng polyacetylenes could actually engage their predicted targets. Molecular docking with AlphaFold-predicted protein structures showed binding energies below −5.0 kcal/mol for DNAJC19, NABP1 and TBPL1, ranging from −5.2 to −6.2 kcal/mol, while RPS21 and MRFAP1L1 bound more weakly, likely reflecting incompatible pocket geometries. Hundred-nanosecond molecular dynamics simulations in GROMACS then confirmed stable complexes for the strong binders, with root mean square deviations holding near 0.15 to 0.30 nanometers, and MM-PBSA calculations yielded strongly negative binding free energies across all complexes, the strongest at −198.2 kcal/mol for MRFAP1L1 with polyacetylene PQ-2. The authors also emphasize the safety caveats: low predicted hERG inhibition and hepatotoxicity probabilities for both PQ-2 epimers support druggability, though these remain computational estimates.

The researchers are appropriately measured about what their findings mean. Every result here is in silico: the gene–immune correlations are statistical associations from bulk transcriptomic data inferred with the ssGSEA algorithm, the binding predictions rely on modeled protein structures, and no cell, animal or clinical experiment has yet tested whether American ginseng polyacetylenes actually mitigate altitude illness. Still, the convergence of transcriptomics, machine learning, network pharmacology and simulation on a single well-known edible herb offers a genuinely novel framework, and previous hints that PQ-2 and its epimer have never been reported as antihypoxic agents makes them intriguing leads. If future experimental work confirms the gene–compound axis identified here, the remedy for thin mountain air could someday be as familiar as the produce aisle, a possibility that makes this study one of the more provocative computational explorations in environmental physiology this year.

Subject of Research: Computational identification of high-altitude illness genes and drug–food homologous compound targets

Article Title: Transcriptomic identification of high-altitude illness candidate genes and physiological mechanisms: a multidimensional bioinformatics study with 106 drug–food homologous substance target prediction

Article References: Transcriptomic identification of high-altitude illness candidate genes and physiological mechanisms: a multidimensional bioinformatics study with 106 drug–food homologous substance target prediction. (n.d.). https://doi.org/10.3389/fphys.2026.1752620

Image Credits: AI Generated

DOI: 10.3389/fphys.2026.1752620

Keywords: high-altitude illness, bioinformatics, transcriptomics, machine learning, drug–food homology, Panacis Quinquefolii Radix, molecular docking, molecular dynamics, hypoxia, immune infiltration, ADMET, biomarkers