Deep Learning and Brain Organoids Uncover Azole Antifungals as Repurposable Drug Candidates for Leigh Syndrome
核心洞察
A deep learning framework trained on non-cancer cell perturbation data outperformed conventional gene set enrichment methods, achieving 9.2% recall of known hits within the top 1st percentile of ranked drugs for Leigh syndrome (搜索).
Parallel screening in a SURF1 (搜索)-deficient yeast model and validation in patient-derived midbrain organoids identified sertaconazole and talarozole as promising repurposable candidates.
Both azole compounds rescued key disease phenotypes in Leigh midbrain organoids, including abnormal growth rate, elevated lactate release, and impaired neuronal calcium response to metabolic stress.
A multi-pronged drug discovery effort combining deep learning-based computational screening, yeast survival assays, and patient-derived brain organoid models has identified two azole antifungal compounds—sertaconazole and talarozole—as potential repurposable treatments for Leigh syndrome (搜索), an incurable pediatric neurometabolic disorder. The findings, published in Nature Communications, demonstrate how integrating in silico prediction with complex human 3D models can accelerate therapeutic discovery for rare neurodevelopmental diseases.
Leigh syndrome (搜索), also known as infantile subacute necrotizing encephalomyelopathy, is a severe mitochondrial disorder caused by inherited gene variants that impair the oxidative phosphorylation (OXPHOS) machinery. The disease primarily affects the basal ganglia and midbrain, leading to motor impairment, intellectual disability, lactic acidosis, and early mortality. Mutations in SURF1 (搜索), which encodes an assembly factor for cytochrome c oxidase (COX), represent one of the most frequent genetic causes. Currently, no curative treatments exist.
A Deep Learning Framework Tailored for Non-Cancer Drug Repurposing
The research team developed a multi-stage deep learning (DL) framework based on ChemPert, a database of transcriptional perturbations in non-cancer cells. Unlike widely used tools such as CMap, which rely predominantly on cancer cell line signatures, this framework was trained on 24,134 transcriptional signatures from 264 unique non-cancer cell types exposed to 2,905 perturbagens.
The framework operated over a curated molecular interaction graph comprising 20,455 protein nodes assembled from protein–protein interactions, signaling pathways, transcriptional regulatory networks, and ligand–receptor relationships. It learned to predict protein-level perturbations—activation, inhibition, or no effect—required to drive a transcriptional state transition from radial glia (RG) identity to intermediate progenitor cells (IPCs) or neurons, the differentiation trajectory found to be impaired in Leigh cerebral organoids.
Using a strictly independent hold-out set of 47 test datasets, the DL framework achieved 9.2% average recall of true positive hits within the top 1st percentile of ranked drugs, substantially outperforming gene set enrichment analysis (GSEA)-based methods (best case: 2.1%) and random ranking (1%). At the 2nd percentile cutoff, recall reached 16.3% compared to 6.1% for the best GSEA result.
The framework employed a variational autoencoder (VAE) for denoising and regularization, followed by an ensemble of three feedforward neural networks with majority voting. Final drug rankings were generated using a Bayesian enrichment score (BES) integrating hypergeometric enrichment between predicted targets and known drug targets, empirical robustness calculated from randomized input matrices, and normalized rank weighting. Talarozole achieved one of the highest BES values in the screen (BES = 9.034277; 38th out of 5,692 drugs, approximately 99.3rd percentile).
Parallel Yeast Screening Converges on Azole Compounds
Concurrently, the researchers conducted a rescue screen in a yeast model lacking SHY1, the SURF1 (搜索) homologue, using a library of 2,250 repurposable drugs. The assay measured survival of wild-type and ΔSHY yeast strains upon nutrient removal. Among the top 2% of rescuing compounds, four belonged to the azole class, which emerged as one of the most strongly affected categories.
Combining both screening strategies yielded nine candidate repurposable compounds—five from the DL predictions and four from the yeast screen.
Validation in Human Induced Neurons and Midbrain Organoids
Using a high-content analysis pipeline to quantify neuromorphogenesis in NGN2-induced neurons (iNs) derived from Leigh patient neural progenitor cells (NPCs), talarozole demonstrated the most pronounced effect among DL-predicted drugs, increasing neuronal numbers two-fold in a concentration-dependent manner at 1 µM and 10 µM. Sertaconazole, identified from the yeast screen, produced a 1.5-fold increase in neuronal count and enhanced mean neurite length at 10 µM. Both compounds showed toxicity only at 50 µM.
The team then generated midbrain organoids (MOs) from Leigh patient-derived NPCs carrying the SURF1 (搜索) c.769 G>A (p.G257R) variant. These organoids recapitulated key disease features: reduced expression of axonal (SMI312), dendritic (MAP2), and dopaminergic (TH) markers; significantly larger organoid size consistent with upregulated proliferation genes; elevated lactate release; and an aberrant calcium response to acute metabolic stress induced by glucose starvation with glycolysis and OXPHOS inhibition.
At non-toxic concentrations (0.1 µM sertaconazole, 1 µM talarozole), both azoles significantly reduced the abnormal growth rate of Leigh MOs and decreased lactate release by approximately 20%. Talarozole treatment increased the number of TH-positive neurons and resulted in a higher number of spontaneously active cells, while also producing a two-fold increase in calcium signal in response to metabolic stress.
Divergent Mechanisms: Bioenergetics Versus Lipid Metabolism
Single-cell RNA sequencing of compound-treated MOs revealed that both azoles primarily affected the neuronal population, upregulating genes associated with lipid and cholesterol metabolism (STARD4, FABP3, MVD, SCD, HMGCR, HMGCS1) and neuronal generation (TUBB3, ELAVL3).
Targeted metabolomics in Leigh NPCs showed that talarozole reversed the impaired AMP/ATP ratio—which was reduced by more than two-fold compared to controls—while sertaconazole's effect did not reach statistical significance. Both compounds normalized the TCA cycle intermediate succinyl-CoA, and talarozole additionally normalized oxaloacetate levels.
Lipidomics revealed divergent effects: sertaconazole principally modulated glycerophospholipids (phosphatidylethanolamine and phosphatidylglycerol) and restored hexosylceramide levels, which were reduced by 50% in Leigh NPCs. Both azoles significantly modulated cholesterol levels, and membrane-bound cholesterol—found to be significantly reduced in Leigh NPCs—was increased by both treatments, with sertaconazole showing a more prominent effect.
Pathway Engagement: Retinoic Acid and PPARγ (搜索)
Molecular docking studies using both GOLD and the machine learning algorithm DiffDock refined by SMINA confirmed that both azoles bind to CYP26A1 (搜索) and CYP26B1 (搜索), the main targets of talarozole in the retinoic acid (RA) pathway. Talarozole exhibited the highest binding affinity among all tested azoles. Luciferase reporter assays demonstrated that talarozole restored CYP26A1 promoter activity in Leigh NPCs, while sertaconazole was ineffective.
Both compounds also showed binding affinity for PPARγ (搜索), a target recently identified as therapeutic in a Leigh mouse model caused by Ndufs4 depletion. Talarozole treatment led to a significant 1.5-fold increase in PPARγ activity in Leigh NPCs, while sertaconazole had no significant impact.
The authors note that talarozole, originally developed for acne and psoriasis and now under investigation for osteoarthritis, appeared as the most effective compound overall. "Talarozole currently represents a single prospectively validated hit and should therefore be viewed as proof-of-concept example, with further testing of additional top-ranked candidates needed to generalize the framework's predictive performance," the researchers stated.
The study underscores the potential of combining computational drug discovery with patient-specific brain organoid models to identify treatments for rare neurodevelopmental disorders, while also supporting the view that Leigh syndrome (搜索) pathogenesis involves impaired neurodevelopment in addition to early-onset neurodegeneration.