Evaluating the Role of AI Assistants in Accelerating Neurodegenerative Disease Research: Opportunities and Translational Limitations Authors: Xiyue Yu, Qianqian Yao Published: 2026 ABSTRACT AbstractNeurodegenerative diseases including Alzheimer’s disease and Parkinson’s disease remain among the most challenging disorders to study, diagnose and treat. Despite rising prevalence with population aging, disease-modifying therapies remain scarce and research progress is hindered by biological complexity, patient heterogeneity, and incomplete experimental systems. Artificial intelligence (AI) has emerged as a transformative approach given the high-dimensional and multimodal data generated in this field. Traditional machine learning and deep learning have advanced imaging biomarker detection, disease trajectory prediction, drug target prioritization, and clinical data mining. More recently, foundation models and large language models (LLMs) have expanded AI from task-specific prediction tools to versatile assistants supporting literature retrieval, summarization, coding, data interpretation, and hypothesis generation. Although AI assistants promise to accelerate research workflows, their outputs are prone to dataset biases, poor interpretability, distribution shift, and hallucinations, which are particularly problematic in neurodegenerative research given subtle phenotypic variations, imperfect labeling, and protracted disease courses. This review evaluates the current applications of AI in neurodegenerative disease research, including drug discovery, biomarker identification, and multi-omics integration. We then discuss the transition from analytical AI models to general-purpose AI assistants and their potential to streamline scientific workflows. Critical limitations including bias, interpretability, reproducibility, and LLM hallucinations are highlighted, alongside ethical, regulatory and practical challenges. We argue that the most sustainable near-term model is human-AI collaboration rather than fully autonomous research, with its primary focus placed on research rather than clinical practice. Meaningful acceleration requires rigorous validation, transparent usage, and expert oversight to preserve scientific rigor and translational relevance.Keywords: neurodegenerative diseases, Alzheimer’s disease, Parkinson’s disease, artificial intelligence Introduction Neurodegenerative diseases such as Alzheimer’s disease (AD) and Parkinson’s disease (PD) are one of the most difficult diseases to study, diagnose and treat. With the aging of the population, the incidence of these diseases continues to rise, but disease modification therapy is still scarce, and the research and development cycles are costly.1–3 The slow progress of research in this field reflects the complexity of underlying biology, heterogeneity among patients, imperfect experimental models, and the need to integrate evidence spanning genetics, imaging, pathology, bedside phenotypic assessment, and therapeutic testing. For these reasons, the study of neurodegenerative diseases holds unique appeal for artificial intelligence (AI): this field generates high-dimensional and multimodal data, yet converting such information into actionable knowledge remains a persistent challenge. Over the past decade, AI has evolved from a specialized computational tool into a broad scientific infrastructure. Machine learning (ML) and deep learning (DL) methods have been applied to detect imaging biomarkers, predict disease trajectories, prioritize drug targets, and identify hidden patterns in clinical records.4–6 More recently, the emergence of foundation models and large language models (LLMs) has broadened the scope of AI from narrow prediction systems to versatile assistants capable of supporting literature retrieval, summarization, coding, data interpretation, and hypothesis generation.7–9 Nevertheless, caution is warranted about the prospect of accelerating the research process. Greater output speed does not equate to better scientific outcomes. AI systems may amplify biases embedded in training data, conceal their reasoning within opaque model architectures, generate statements that appear plausible but are factually incorrect, and fail to generalise beyond the distributions on which they were trained.10–13 These limitations are particularly consequential in the field of neurodegenerative diseases, where subtle phenotypic variation, incomplete labelling, and protracted disease courses make validation especially demanding. The key question, therefore, is not simply whether AI assistants can generate biomedical text or predictive outputs, but whether they can improve the quality, speed, and translational relevance of research without introducing new sources of error. This review evaluates the role of AI assistants in accelerating neurodegenerative disease research. It first surveys domains in which AI has already made substantive contributions, including drug discovery, biomarker research, and multimodal disease classification. The review then examines how AI assistants built on foundation models and LLMs can support broader scientific workflows. The central argument is that AI assistants can accelerate certain aspects of the research process, but their real value is shaped by data quality, model interpretability, reproducibility, scientific reliability, and translational effectiveness. For these reasons, the most productive near-term model is likely human-AI collaboration rather than AI-driven autonomous research. Current Applications of AI in Neurodegenerative Disease Research Distinguishing between AI as an analytical method and AI as an assistant is a useful entry point. Before the rise of chat-based systems, AI had played an important role in the study of neurodegenerative diseases through some more clearly defined tasks. One of the main applications is drug discovery and target prioritization. Machine learning methods can combine genomic, chemical and phenotype data to identify candidate targets, predict compound properties and optimize the screening process.14 At the level of basic biology, AlphaFold has revolutionized the structural prediction method by achieving structural prediction of many proteins close to experimental accuracy, thus reducing a major bottleneck in structural biology and bringing downstream benefits to target identification and molecular design.15 Although AlphaFold is not a tool specifically for neurodegenerative diseases, it shows how AI shortens a slow research step that is applicable to various disease areas. The second main application is the discovery of biomarkers. Neurodegenerative diseases usually develop several years before clinical diagnosis, so there is an urgent need for early, sensitive and scalable biomarkers. Therefore, data sets rich in images have become fertile ground for AI. Deep learning systems can extract subtle spatial patterns that may be difficult for human readers to detect from MRI or PET scans, and have been widely used in a wider range of medical image analysis.6 In dementia research, AI models have been used to use multimodal data to classify cognitive disorders and distinguish the causes of dementia. For example, Xue et al reported an AI framework that integrates demographic variables, clinical history, neuropsychological evaluation, functional measurement and multimodal neural imaging to distinguish the causes of dementia, and achieves high overall performance on geographically distributed data sets.16 The importance of this study lies not only in its accuracy index, but also in its attempt to run under incomplete data conditions similar to the real clinical environment. Similarly, Marshall et al explored low-cost ocular or imaging-derived biomarkers for AD, reflecting a shift toward scalable digital or imaging phenotyping.17 The study of Parkinson’s disease (PD) has produced similar examples. The research of Yang et al shows that AI systems can identify Parkinson’s disease and assess its severity through nocturnal breathing signals, which suggests a non-invasive digital biomarker that may be helpful for diagnosis and clinical trial monitoring.18 The significance of this work lies in its potential for longitudinal monitoring: if disease progression can be measured more frequently and non-invasively, interventional studies can be conducted more rapidly and economically. Winchester et al also emphasized the potential of AI in the discovery of biomarkers of Alzheimer’s disease, especially in the case of integration rather than independent interpretation of heterogeneous data sources.19 Beyond imaging-based biomarkers, AI and machine learning techniques are increasingly being applied to the detection of molecular and biochemical biomarkers for neurodegenerative diseases. For example, Wang et al developed an ultrasensitive graphene field-effect transistor biosensor capable of detecting five Alzheimer’s disease plasma biomarkers at femtomolar levels, and used machine learning to assemble these into a composite diagnostic panel that achieved an area under the curve exceeding 0.94 in staging AD progression.20 Similarly, Gong et al developed a carbon nanotube field-effect transistor biosensor using a dopamine self-polymerization approach for oriented antibody immobilization, enabling highly sensitive detection of p-tau 217, and coupled this with a convolutional neural network model to achieve 100% diagnostic accuracy in clinical serum samples from AD patients.21 These developments suggest that the role of AI in biomarker research extends well beyond neuroimaging, encompassing the full spectrum from molecular diagnostics to digital phenotyping. The third application is multi-omics integration. Neurodegenerative diseases involve complex interactions spanning genetics, epigenetics, transcriptomics, proteomics, and environmental factors. AI is well suited to these data-rich but conceptually fragmented domains. By identifying patterns across layers of biological data, AI models can identify disease subtypes, risk signatures, or therapeutic hypotheses that would be difficult to detect using conventional univariate approaches.22,23 However, this prospect is inseparable from the limitations of the existing data set. Cohort sizes in this field are frequently modest, sample selection is often biased towards high-resource clinical centres, and geographical diversity remains limited.12,24 Because diagnosis is typically based on clinical criteria rather than pathological confirmation, outcome labels can be noisy and inconsistent.19 These data quality issues directly constrain the reliability of AI models trained on such cohorts. In summary, these examples show that AI is already helping to study neurodegenerative diseases. It can accelerate target discovery, improve the feature extraction of images and signals, and help integrate complex data sets. But these are still mostly examples of task-specific AI, not AI assistants in a broad sense. From Analytical Models to AI Assistants AI assistants are different from early biomedical AI systems because they aim to support multiple stages of the workflow rather than solving a single prediction problem. Foundation models are pre-trained on large and diverse corpora, requiring relatively little additional data for adaptation to downstream tasks, while LLM-based assistants provide a conversational interface that makes complex analytical tools accessible to non-specialists.7,8 This is particularly relevant for neurodegenerative disease research, which is inherently interdisciplinary. A single project may require researchers to navigate genetics, neuropathology, neuroimaging, clinical neurology, and therapeutic development simultaneously. In principle, AI assistants can lower the friction associated with such interdisciplinary integration. At the most basic level, AI assistants can accelerate literature retrieval. They can summarize articles, cluster studies by topic, extract key arguments, and surface connections between concepts that would otherwise require extensive manual reading.9 This is a meaningful advantage in a field characterised by rapidly expanding literature. In addition, LLMs can assist with writing code, organising analysis plans, and converting unstructured notes into searchable formats. They can also support hypothesis generation by identifying links between pathways, genes, or phenotypes that have not yet been fully explored. In neurodegenerative disease research, where many projects involve exploratory synthesis prior to targeted validation, these capabilities have clear practical value. AI assistants also have strong application prospects in multimodal research and design. Moor et al believe that general medical AI can eventually flexibly interpret the combination of images, electronic health records, laboratory data, genomics and free text.8 This is closely related to neurodegenerative diseases, because clinically significant signals are rarely limited to a single mode. AI assistants with cross-modal coordination can help researchers shift from isolated analysis to more integrated disease models. However, the leap from assistance to reliable acceleration should not be exaggerated. AI assistants may save time, but they also increase the burden of verification. Summaries may omit important methodological caveats. Hypothesis generation may favour statistically frequent associations in training data rather than biologically meaningful ones. Coding assistance may produce plausible-looking scripts that nonetheless require substantial manual review. The net effect on research speed is therefore conditional: AI assistants can only genuinely accelerate discovery if the time saved is not consumed by downstream error correction, verification, or interpretive clarification. Furthermore, there remains a meaningful cognitive gap between retrieval and genuine understanding AI. Therefore, the relevant measure is not whether AI assistants can generate smooth outputs, but whether they can retain the uncertainty, nuances and methodological distinctions necessary for reliable science. Critical Evaluation: Where Acceleration Breaks Down The strongest argument against blind optimism is that AI systems inherit the limitations of the data and evaluation environments in which they are built. In neurodegenerative disease research, data quality is a central concern. Clinical datasets commonly over-represent specific populations, specialist centres, or particular disease stages; imaging acquisition protocols vary substantially across institutions; outcome labels are frequently uncertain; and longitudinal follow-up is often incomplete.10,12,19,24 If AI assistants are built on these imperfect foundations, they risk amplifying errors as readily as they amplify insights. Obermeyer et al shows that when biased proxy variables are used in model development, clinically deployed algorithms will encode and amplify unfairness.10 Although their research is not specifically aimed at neurodegenerative diseases, their lessons are directly related: if AI systems are trained on datasets that distort disease burden or healthcare access patterns, they risk systematically misleading researchers about which questions are most scientifically and clinically important. The second limitation is explainability. The black box model is usually defended on the grounds that predictive performance is the main goal. However, in the research environment, explanations are not optional. The science of neurodegenerative diseases does not only try to classify patients, but also tries to understand their mechanisms and determine interventions. Rudin believes that for high-risk fields, models that are interpretable in themselves should be prioritized as much as possible, rather than relying on the explanation of the opaque system after the fact.11 Doshi-Velez and Kim also pointed out that interpretability must be rigorously evaluated according to practical application.25 For neurodegenerative disease research, this means that a model lacking interpretable reasoning may be scientifically insufficient even when predictive performance is high, because the field requires mechanistic inference and intervention design, not merely classification. AI assistants compound this problem by presenting outputs in fluent, authoritative-sounding natural language, which can make uncertainty appear smaller than it actually is. Third, reproducibility and generalisation remain major obstacles. Many AI research reports say that the model performs well on the internal data set, but when the model encounters different people, collection schemes or missing data patterns, its performance will decrease. Finlayson et al demonstrated that dataset shift is a primary driver of clinical AI,12 while Beam and Kohane warned that the model may become vulnerable when deployed outside the training environment.24 Neurodegenerative diseases are especially susceptible to these issues because diagnostic categories continue to evolve, biological subtypes are heterogeneous, and gold-standard pathological labels are rarely available at scale. A model that performs well within one research consortium may fail to transfer to another, and an AI assistant that does not flag these limitations when summarising such literature may instil unwarranted confidence. Fourth, LLM-based assistants introduce the risk of hallucinations and artificially coherent but factually unsupported outputs. Bender et al demonstrated that LLMs can generate fluent text that lacks grounding in factual evidence, a problem that is especially consequential in scientific and medical contexts.26 In the research workflow, hallucinations are not merely an inconvenience; they represent a source of silent contamination. Assistants may fabricate references, misattribute results, or obscure meaningful differences between studies. Gao et al shows that ChatGPT can generate seemingly credible but actually fabricated abstracts of medical research,27 while Clusmann et al emphasized that without adequate supervision, LLMs may spread misinformation and promote academic misconduct.9 In neurodegenerative disease research, this risk is particularly salient because subtle errors at the boundary of a reader’s expertise may go undetected without close scrutiny. Ethical, Regulatory and Practical Considerations The application of AI assistants in neurodegenerative disease research is subject to important ethical and regulatory constraints. Many valuable datasets contain sensitive clinical narratives, longitudinal records, identifiable neuroimaging data, or genetic information. Access to such data is rightly restricted, yet this creates a tension: the datasets most relevant to clinical translation are often the hardest to use for developing inclusive and robust AI systems.28 Privacy-preserving techniques such as federated learning and differential privacy may help, but they introduce their own technical trade-offs and cannot substitute for thoughtful data governance frameworks. Bias and fairness are equally important considerations. If AI systems are trained predominantly on data from high-resource healthcare settings, they risk entrenching global inequalities in evidence generation. Pfohl et al argued that LLM-based health tools can produce a range of equity harms, a concern that applies equally to research assistants and clinical tools.29 Systematic over-representation of mainstream publishing platforms, English-language literature, or cohorts from well-resourced institutions may subtly skew what counts as established evidence. The use of LLMs in scientific writing and reasoning raises further questions about authorship, accountability, and research integrity. Ong et al outlined a bioethical framework for the use of LLMs in medicine, grounded in the principles of beneficence, non-maleficence, autonomy, and justice.30 In the research context, a related concern is accountability: who bears responsibility when an AI-generated summary contains errors, fabricates citations, or overlooks methodological flaws? Current guidance makes clear that AI tools cannot be listed as authors, but this does not resolve the practical challenge of ensuring verification. Gao et al and Mishra et al both pointed out that while researchers recognise the value of LLMs, they are also acutely aware of the associated integrity risks.27,31 Accordingly, the most prudent practice is one of openness, transparency, and traceability. Where AI assistants are used for literature review, protocol drafting, or analytical scripting, such use should be declared explicitly and accompanied by rigorous human review. Regulatory frameworks for AI in biomedical research are still evolving; in the interim, researchers bear direct responsibility for the accuracy and integrity of any AI-assisted outputs. Future Directions: What Meaningful Acceleration Would Look Like The most credible future of AI assistants in neurodegenerative disease research is not autonomous scientific discovery, but carefully designed human-AI collaboration. Meaningful acceleration means reducing low-value friction while reinforcing rather than bypassing scientific judgement. In practice, assistants are most valuable when used to support systematic literature search, extract structured information from papers, identify missing data fields, summarise cohort characteristics, and generate reproducible analysis code templates that operate under expert supervision. Several technical directions hold promise for improving the reliability of AI assistants. Retrieval-augmented generation systems can reduce hallucinations by grounding responses in identifiable source documents rather than relying solely on parametric memory. Domain adaptation and model fine-tuning can improve performance on biomedical tasks, although specialisation should not be conflated with validation. In neurodegenerative disease research, multimodal foundation models may become increasingly useful if text, imaging and structured clinical data require integrated interpretation.8 Meanwhile, interpretability should be treated as a design goal rather than a post-hoc consideration, especially in studies aimed at clarifying mechanisms or stratification. A more rigorous evaluation framework is equally necessary. Research should move beyond benchmark accuracy and examine whether AI assistants genuinely reduce the time to insight, improve methodological quality, enhance reproducibility, and increase the success rate of downstream experiments. Head-to-head comparisons of researchers working with and without AI assistants on realistic tasks would be more informative than leaderboard comparisons alone. Hager et al offer a useful example of how evaluation can be brought closer to real deployment conditions.13 The research field also needs a similar transformation: AI assistants should be evaluated within realistic scientific workflows that include uncertainty communication and source verification as explicit performance criteria. Specifically to the study of neurodegenerative diseases, the most promising short-term application scenario may be the scenario in which AI assistants serve as an integrative scaffold. These scenarios include synthesising evidence across imaging and omics studies, and systematically mapping the biomarker literature. AI can be used for disease staging, or to support transparent pre-screening of candidates for experimental follow-up research. In these applications, AI does not replace domain expertise; rather, it extends researchers’ capacity to engage with complex, high-dimensional problems. Conclusion AI assistants hold significant potential to accelerate neurodegenerative disease research, particularly in compressing literature, supporting multimodal data integration, and reducing the burden of repetitive analytical tasks. Established AI applications in neuroimaging, biomarker discovery, digital phenotyping, and drug development demonstrate that computational methods can already accelerate specific steps in the research pipeline. LLM-based assistants and foundation models extend this potential to broader scientific workflows. However, acceleration is not unconditionally beneficial. In neurodegenerative disease research, the value of faster AI systems depends on whether they can maintain scientific rigour and translational relevance. Challenges including data bias, limited interpretability, distributional shift, hallucinations, and insufficient task-specific evaluation mean that apparent efficiency gains may ultimately generate downstream errors or waste. The most defensible conclusion is therefore neither that AI assistants are transformative nor that they are merely overhyped. They are genuinely useful tools whose value is determined by careful task selection, transparent deployment, and rigorous human oversight. Funding Statement This study was funded by Shandong Traditional Chinese Medicine Science and Technology Project (Q-2022107). Abbreviations AD, Alzheimer’s disease; AI, Artificial intelligence; DL, Deep learning; LLM, Large language model; ML, Machine learning; MRI, Magnetic resonance imaging; PET, Positron emission tomography; PD, Parkinson’s disease. Author Contributions All authors made a significant contribution to the work reported, whether that is in the conception, study design, execution, acquisition of data, analysis and interpretation, or in all these areas; took part in drafting, revising or critically reviewing the article; gave final approval of the version to be published; have agreed on the journal to which the article has been submitted; and agree to be accountable for all aspects of the work. Disclosure The authors declare that they have no competing interests. References