AI Scours 1.2 Million Natural Products to Find Fungal Compound That Starves Glioblastoma of Cholesterol

Glioblastoma, the most aggressive and lethal primary brain cancer in adults, has stubbornly resisted every systemic therapy introduced over the past decade. Now, a research team from the Naval Medicine Center of PLA at Naval Medical University and the School of Life Sciences at Henan University in China reports a strikingly different route of attack: rather than going after mutations or signaling pathways directly, they targeted the tumor’s profound addiction to cholesterol, and they let artificial intelligence do the searching. Writing in the journal Molecular Diversity, Qi Li, Chunxue Zhang, Tiantian Hu, Zhenzhen Zhang, and Haigang Wu describe an integrated AI-driven screening pipeline that sifted through roughly 1.2 million natural products and surfaced a humble fungal sterol, 5,6-epoxyergosterol, as a potent and selective candidate against glioma cells.

The biological rationale behind the study rests on one of glioblastoma’s most distinctive metabolic vulnerabilities. Unlike many tissues that synthesize their own cholesterol, glioblastoma cells are heavily co-dependent on exogenous cholesterol to fuel oncogenic signaling and the relentless membrane biogenesis required for rapid proliferation. This dependency places the liver X receptor beta (LXRβ), the principal transcriptional regulator of cholesterol efflux in the central nervous system, at the center of a promising therapeutic strategy. Activating LXRβ in tumor cells forces them to export cholesterol, depleting intracellular pools and effectively starving the cancer of a raw material it cannot easily replace.

The catch has always been selectivity. LXR receptors come in two flavors, LXRα and LXRβ, and indiscriminate activation of both has historically caused hepatotoxicity through excessive lipogenesis in the liver, where LXRα dominates. Early synthetic agonists such as T-0901317 activated both isoforms and were never suitable for chronic cancer therapy. On top of that, any drug aimed at a brain tumor must cross the blood-brain barrier, a formidable filter that excludes most large or polar molecules. Developing an LXRβ-selective, brain-penetrant agonist has therefore remained an unsolved challenge in neuro-oncology, and it is precisely the problem the Chinese team set out to crack with machine learning rather than traditional medicinal chemistry.

Their solution is a multi-layered computational pipeline that combines several complementary artificial intelligence approaches. At its core sits a machine learning-based quantitative structure-activity relationship (QSAR) model trained to predict LXRβ binding affinity, working alongside a directed message passing neural network (D-MPNN), a graph-based deep learning architecture that learns molecular representations directly from chemical structure. A deep learning-based drug-target interaction (DTI) predictor adds a third independent estimate of whether a given compound is likely to engage the receptor. By running a curated library of approximately 1.2 million natural products through this ensemble, the team could rank candidates not just on raw predicted potency but on the probability that the predictions would hold up experimentally.

Crucially, the pipeline did not stop at affinity. The researchers applied sequential filters for LXRβ/LXRα selectivity, ensuring that hits would preferentially activate the beta isoform and spare the liver from unwanted LXRα-driven lipogenesis. An integrated pharmacokinetic scoring step then assessed whether candidates possessed the physicochemical properties needed to reach the brain. Surviving compounds were subjected to molecular docking and MM-GBSA binding free energy calculations, which model the physical fit and energetics of each ligand inside the LXRβ binding pocket. This layered funnel, from millions of compounds down to a handful of high-confidence candidates, exemplifies how modern AI screening can compress what was once a decade-long campaign into a focused computational exercise.

Four candidates emerged from the computational gauntlet, and the team moved to the laboratory to test them. Using CCK-8 cytotoxicity assays across five glioblastoma cell lines and six normal cell models, the researchers evaluated both anti-tumor potency and therapeutic window. One compound stood out: 5,6-epoxyergosterol, an oxidized derivative of ergosterol, the fungal counterpart of cholesterol. The molecule showed potent cytotoxicity against glioma cells while sparing normal cells, exactly the selectivity profile the computational filters were designed to enforce. That a compound derived from fungal sterols, a chemical class evolutionarily tuned to interact with sterol-sensing proteins, would emerge as an LXRβ agonist is a satisfying convergence of natural product chemistry and computational prediction.

To understand how 5,6-epoxyergosterol engages its target at atomic resolution, the team ran a 200-nanosecond molecular dynamics simulation of the LXRβ-ligand complex. The simulation demonstrated remarkably stable binding throughout the trajectory, with the compound anchored by dominant hydrogen-bonding contacts to two key residues, His435 and Trp443, in the receptor’s ligand-binding domain. The free energy landscape computed from the simulation was consistent with a single dominant agonist-bound conformation, suggesting that the compound locks the receptor into an active state rather than sampling multiple binding modes. This kind of conformational stability is what medicinal chemists look for when distinguishing genuine agonists from transient binders, and it provides a structural hypothesis for how the fungal sterol activates cholesterol efflux genes.

The broader significance of the work extends beyond a single compound. The study draws on rich genomic datasets to justify its target: single-cell RNA sequencing data from 24 IDH-wildtype glioblastoma tumors comprising 7,550 cells, bulk RNA-seq data from 173 patients in the TCGA-GBM cohort, and spatial transcriptomics from the Ivy Glioblastoma Atlas Project. By grounding the computational campaign in human tumor data, the researchers ensured that the cholesterol dependency they were exploiting is not an artifact of cell culture but a feature of real disease. The approach also positions fungal natural products, an underexplored corner of chemical space compared with plant-derived compounds, as a rich source of central nervous system drug leads.

For a field that has seen no new approved systemic therapy in more than ten years, the prospect of a cholesterol-starvation strategy delivered by a brain-penetrant, LXRβ-selective natural product is genuinely exciting. The standard of care remains surgical resection followed by radiotherapy and temozolomide chemotherapy, with median survival measured in months, and recent immunotherapy and targeted therapy efforts have largely failed to move the needle. Metabolic vulnerabilities like the LXR-cholesterol axis offer a way to attack the tumor’s fundamental biochemistry, an approach that may be harder for heterogeneous tumors to escape than single-pathway inhibition, since every glioblastoma cell needs membranes.

Considerable work remains before 5,6-epoxyergosterol or its analogs reach the clinic. The current evidence rests on computational prediction, in vitro cytotoxicity, and simulation; animal pharmacokinetics, brain exposure studies, and formal selectivity profiling against the full nuclear receptor family will be essential next steps. Nevertheless, the study delivers a validated AI-driven framework for natural product drug discovery in neuro-oncology, one that other laboratories can adapt to different targets and compound libraries. If the fungal ergosterol derivatives identified here continue to perform as they move toward preclinical development, the marriage of machine learning and mycology may prove to be one of the more unexpected alliances in the fight against brain cancer.

Subject of Research: AI-guided discovery of fungal ergosterol derivatives as selective LXRβ agonists targeting cholesterol dependency in glioblastoma

Article Title: Artificial intelligence-guided discovery of fungal ergosterol derivatives as selective LXRβ agonists targeting the cholesterol dependency of glioblastoma

Article References: Li, Q., Zhang, C., Hu, T., Zhang, Z., & Wu, H. (2026). Artificial intelligence-guided discovery of fungal ergosterol derivatives as selective LXRβ agonists targeting the cholesterol dependency of glioblastoma. Molecular Diversity. https://doi.org/10.1007/s11030-026-11728-7

Image Credits: AI Generated

DOI: 10.1007/s11030-026-11728-7

Keywords: glioblastoma, LXRβ, cholesterol metabolism, artificial intelligence, virtual screening, natural products, ergosterol, blood-brain barrier, molecular dynamics, drug discovery, neuro-oncology, machine learning