Graph AI generates neurological hypotheses validated in molecular, organoid, and clinical systems

Abstract: Neurological diseases are the leading global cause of disability, yet most lack disease-modifying treatments. We present PROTON, a heterogeneous graph transformer that generates testable hypotheses across molecular, organoid, and clinical systems. To evaluate PROTON, we apply it to Parkinson's disease (PD), bipolar disorder (BD), and Alzheimer's disease (AD). In PD, PROTON linked genetic risk loci to genes essential for dopaminergic neuron survival and predicted pesticides toxic to patient-derived neurons, including the insecticide endosulfan, which ranked within the top 1.29% of predictions. In silico screens performed by PROTON reproduced six genome-wide $α$-synuclein experiments, including a split-ubiquitin yeast two-hybrid system (normalized enrichment score [NES] = 2.30, FDR-adjusted $p < 1 \times 10^{-4}$), an ascorbate peroxidase proximity labeling assay (NES = 2.16, FDR $< 1 \times 10^{-4}$), and a high-depth targeted exome sequencing study in 496 synucleinopathy patients (NES = 2.13, FDR $< 1 \times 10^{-4}$). In BD, PROTON predicted calcitriol as a candidate drug that reversed proteomic alterations observed in cortical organoids derived from BD patients. In AD, we evaluated PROTON predictions in health records from $n = 610,524$ patients at Mass General Brigham, confirming that five PROTON-predicted drugs were associated with reduced seven-year dementia risk (minimum hazard ratio = 0.63, 95% CI: 0.53-0.75, $p < 1 \times 10^{-7}$). PROTON generated neurological hypotheses that were evaluated across molecular, organoid, and clinical systems, defining a path for AI-driven discovery in neurological disease.

• The paper presents Proton, a heterogeneous graph transformer that integrates multimodal data to generate and stratify neurological hypotheses across molecular, organoid, and clinical systems.
• It validates predictions with strong metrics (AUROC 0.9145) and significant enrichment in PD experimental screens, effectively linking genetic risk and environmental signals.
• The framework employs disease-centric data splits and rare relation upweighting to enable robust drug repurposing and cross-scale validation in neurological disorders.

Graph Transformer-Driven Hypothesis Generation and Cross-System Validation in Neurological Disease

Introduction

The paper "Graph AI generates neurological hypotheses validated in molecular, organoid, and clinical systems" (2512.13724) presents Proton, a heterogeneous graph transformer model designed to generate and stratify biomedical hypotheses across molecular, organoid, and real-world clinical contexts for neurological disorders such as Parkinson’s disease (PD), bipolar disorder (BD), and Alzheimer’s disease (AD). Proton leverages an integrative, multimodal knowledge graph (NeuroKG), encoding relationships among genes, proteins, cell types, brain regions, phenotypes, drugs, and environmental exposures, contextualized with single-nucleus RNA-seq data from the adult human brain. The aim is to computationally generate biologically and clinically relevant hypotheses and prospectively validate them through rigorous experimentation.

Figure 1: Overview of Proton’s architecture—graph transformer pre-trained on a multimodal KG—used to generate hypotheses and guide discovery across PD, BD, and AD.

Proton Architecture and NeuroKG Construction

Proton is instantiated as a 578M-parameter Heterogeneous Graph Transformer (HGT) trained on NeuroKG—a 147,020 node, 7,366,745 edge knowledge graph integrating 36 biomedical datasets and ontologies across biological scales. NeuroKG uniquely encodes molecular, cellular, anatomical, clinical, and environmental relationships specific to the human brain. The model’s attention mechanism stratifies complex networks, enabling cross-modal and cross-level reasoning while upweighting rare relations to counter node degree bias. Training is performed using a self-supervised link prediction objective validated with AUROC (0.9145) and high accuracy (82.23%).

Notably, the graph is enriched by contextualization with single-nucleus RNA-seq data encompassing 3.7 million cells (both patient and control tissue), conferring specificity for cell-type/disease interactions and supporting network inference at cellular resolution.

Cross-Scale Validation in PD: Integrating Molecular and Environmental Signals

In Silico Recapitulation of Experimental Data

Proton’s predictions on PD were evaluated against six genome-wide experimental $\alpha$-synuclein screens, spanning MYTH system interactome, APEX2 proximity labeling, and targeted exome sequencing. Gene set enrichment analysis (GSEA) of Proton’s combined PD $+$ $\alpha$-synuclein screen demonstrated strong, significant enrichment for experimental hits, outperforming graph random walk and baseline algorithms across all assays (e.g., MYTH NES=2.30, FDR=$<$1e-4).

Figure 2: Proton enrichment for genes/proteins linked to PD and pesticides associated with toxicity using multi-modal data.

Integrating Genetics and Functional Data

Proton robustly connects GWAS-identified PD risk loci to genes essential for dopaminergic neuron survival, as established in an unbiased whole-genome CRISPR screen. Comparative analysis across 11 diseases revealed strongest median ranking for PD GWAS hits (median-median rank=9,487.5, Kruskal-Wallis p$<$0.001). These associations hold even after controlling for GWAS list-size differences, indicating disease-specific functional integration. Further, the framework identifies autophagy-related essential genes bridging rare and common variant signals in PD modules, corroborating results from NERINE rare variant burden tests.

Environmental Toxicology Prediction

Fine-tuned Proton models prioritize environmental signals, ranking pesticides that confer PD risk with high fidelity (e.g., endosulfan at top 1.29%). Predictions directly correlate with compounds shown experimentally to be toxic to patient-derived dopaminergic neurons, validating Proton’s utility in environmental risk forecasting from sparse datasets.

Drug Repurposing Framework for BD and AD

Disease-Centric Data Splitting for Generalizability

To rigorously evaluate the model’s ability to repurpose drugs, Proton employs disease-centric data splits which remove all drug-disease relationships for the target disease (and related diseases) from the training set, enforcing strict independence and minimizing leakage. The model is retrained per split and demonstrates robust recall for known FDA-approved, off-label, or clinically supported drugs ($>$80% at top 10%, $>$93% at top 20%).

BD: Organoid Model Evaluation of AI-Nominated Drug Candidates

Proton predicts candidate drugs for BD, with high recall for known drugs even in split settings. Among top candidates, calcitriol (active vitamin D) is selected based on LLM and expert review. Treatment of BD patient-derived cortical organoids with calcitriol reverses disease-associated proteomic signatures in 4/5 lines and normalizes clusters to control expression, indicating restoration of altered cellular networks—a mode of action partly distinct from lithium.

Figure 3: Calcitriol reverses BD-associated proteomic changes in patient-derived organoid models.

AD: Large-Scale EHR-Based Risk Reduction Analysis

Model-driven drug repurposing predictions for AD are validated retrospectively in EHR data from 610,524 patients. Proton-nominated drugs (e.g., aflibercept, dapagliflozin, valsartan, spironolactone, metoprolol) exhibit significant associations with reduced 7-year risk of ADRD (min HR=0.63, 95% CI=[0.53–0.75], p$<$1e-7), suggesting candidate neuroprotective agents for clinical re-evaluation.

Figure 4: Proton-predicted drugs with evidence of risk reduction for ADRD in large-scale EHR analysis.

Implications and Critical Assessment

Proton demonstrates that attention-based heterogeneous graph transformers built on multimodal, brain-specific KGs can synthesize molecular, cellular, and clinical relationships, generating hypotheses that withstand rigorous, prospective experimental validation in real-world systems. The strong numerical enrichment in cross-validated molecular screens, direct prediction of toxic pesticides, and clinical correlation of drug efficacy/repurposing indicate practical tractability for AI-driven biological research. Importantly, the split protocol and node/edge upweighting strategies address information leakage and sparsity issues central to robust biomedical AI.

A significant limitation is incomplete coverage of rare or understudied conditions due to bias in available datasets and KG topology—models trained on extant networks may overweight well-studied, highly connected nodes, though this bias is mitigated by rare relation optimization. Interpretation and causal inference will require longitudinal retraining and continuous expansion of KGs as new evidence emerges.

Future Directions

The demonstrated cross-scale congruence in three major neurological disorders suggests feasibility for expansion to other brain diseases, provided enhancements in KG completeness and diversity. Development of real-time updating pipelines, integration of detailed variant-level data, and benchmarking against additional experimental systems will be instrumental. Further, coupling similar models with generative experimental design or closed-loop hypothesis-experiment cycles could accelerate precision therapy discovery.

Conclusion

Proton validates the utility of graph transformer AI models for holistic, scalable hypothesis generation in neurology, producing predictions validated across molecular, organoid, and patient data. This approach supports robust, testable target identification, risk prediction, and interventional mapping—informing both mechanistic understanding and translational strategy in neurological disease research.