AI-Generated Parkinson’s Disease Hypothesis Advances to Animal Testing, With CHK2 Emerging as a New Target

Generative AI can rapidly list disease-associated genes, but the real challenge is to propose mechanisms that can be disproven or confirmed in the laboratory. A research team published the XunZi system in *Nature Biomedical Engineering*, enabling the model not only to rank candidate targets but also to attempt to explain how they participate in disease. Its proposed Parkinson’s disease target, CHK2, subsequently received support from cell and mouse experiments.

XunZi consists of two interconnected components. The multi-omics module, XunZi-M, integrates molecular and biological data to screen for relevant genes based on disease characteristics. The reasoning module, XunZi-R, which is based on Mistral-7B, synthesizes biomedical knowledge to generate experimentally testable gene–disease mechanisms. The research team said the system’s training materials encompassed 24.4 million biomedical publications and 613.6 TB of data from multiple sources, covering 21,008 human genes and 5,850 diseases.

In its Parkinson’s disease analysis, XunZi pointed to the cell-cycle checkpoint kinase CHK2 and the immune-signaling kinase IRAK4, finding that both were abnormally activated in multiple disease models. The study then focused further on CHK2: reducing its function genetically or inhibiting its activity with drugs alleviated the loss of dopaminergic neurons in the substantia nigra and motor deficits in mice. This moved the research a critical step beyond computer-generated rankings, establishing a testable chain of “candidate target–mechanism of action–animal phenotype.”

The associated proteomic and phosphoproteomic data came from two commonly used models: mice in which lesions were induced using the neurotoxin MPTP, and mice inoculated with α-synuclein preformed fibrils. The researchers analyzed substantia nigra and cerebellar tissues to identify changes in protein expression and phosphorylation across different brain regions. Each model simulates some features of Parkinson’s disease. Using them together can reduce bias caused by relying on a single model, but still cannot fully reproduce the long-term and heterogeneous progression of the human disease.

The research team has released the code, model checkpoints, evaluation scripts, and demonstration data. It has also provided a Parkinson’s disease-specific kinase model and examples of prompts related to CHK2 and IRAK4, helping other researchers examine the inference process. However, the public version primarily supports inference demonstrations, and the complete training data are far larger than the included files. Whether external teams can independently reproduce the target rankings and mechanistic reasoning remains an important test of the credibility of systems of this kind.

A more immediate limitation is that the current efficacy evidence remains at the preclinical stage. Neuroprotection and motor improvements in mice cannot be directly extrapolated to efficacy in humans. Whether long-term CHK2 inhibition affects the DNA damage response, whether a drug can safely enter the brain, and which types of patients might be suitable all remain to be studied. The available summary also mainly supports abnormal IRAK4 activation and does not present intervention validation at the same level as for CHK2. XunZi is therefore more like a “hypothesis engine” capable of proposing experimental directions than an AI physician already proven suitable for clinical decision-making.