Computational Molecular Biology 2025, Vol.15, No.5, 254-264 http://bioscipublisher.com/index.php/cmb 260 6.1 Data quality and heterogeneity The problem with AD data is not only the technical differences, but more intractable is its inherent biological complexity. The disease stage, genetic background and whether there are comorbidities of different patients may all affect the pattern of the data. In addition, due to the generally small sample size and limited longitudinal data, the accuracy of statistical analysis will also decline accordingly. Although multi-omics integration sounds more comprehensive, when data from different scales and with different sources of error are placed together, the problem of heterogeneity becomes even more prominent. Some methods proposed in recent years, such as fuzzy hypergraph models or contrast learning, can reduce noise to a certain extent and capture more complex correlations. However, the extent of improvement ultimately achieved is still limited by the quality of the original data (Bi et al., 2025; Koksalmis et al., 2025). 6.2 Methodological constraints Whether a network can be successfully inferred largely depends on the selected algorithm and its parameters. However, many methods are based on some default assumptions, such as the stability of the network topology or the interaction mode between genes, and these assumptions may not hold true in multifactorial diseases like AD. Although deep learning or graph neural networks are powerful, the problem of poor interpretability has always existed. Meanwhile, overfitting is also common. Once the data scale is too small or the samples are biased, the model's performance on external data often declines rapidly. Furthermore, integrating heterogeneous data itself brings computational and statistical challenges. How to balance sensitivity and specificity simultaneously remains a difficulty for current methods (Khatami et al., 2020; Raza et al., 2025). 6.3 Biological interpretation Although the Internet can be built, it is actually not easy to truly explain and interpret these results into biologically significant content. The pathological mechanism of AD has not been fully understood yet. Coupled with the complex structure and diverse cell types of the brain, many things cannot be accurately presented by a static model. The identified modules or genes often lack clear functional annotations and are difficult to directly correspond to clinical manifestations, which also makes the experimental verification progress relatively slow. Furthermore, the development of AD does not progress linearly. Different cells will undergo different changes at different stages, and these differences can be easily "flattened" in a static network. Current research is attempting to enhance the interpretability of models through more detailed annotation methods or by more closely pairing network results with clinical phenotypes. Even so, there is still a considerable gap between computational prediction and actual biological understanding that needs to be gradually narrowed (Huang et al., 2025). 7 Future Directions in Network Reconstruction for AD In recent years, the progress of Alzheimer's disease (AD) network reconstruction has increasingly relied on data of a finer scale. Especially single-cell and spatial omics, they can present the state and tissue structure of cells from different perspectives, allowing researchers to see details that were difficult to observe in the past. Single-cell multi-omics can record transcriptional, epigenetic and protein information at the single-cell level, while spatial omics fills in the gap in the location and neighborhood relationships of cells in brain tissue. To handle these complex "irregular" data, computational methods such as graph neural networks (GNNS) are increasingly adopted, and they have shown considerable potential in reconstructing the gene regulation and signal networks related to AD (Efremova and Teichmann, 2020). 7.1 Integration of single-cell and spatial omics In future research, the combination of single-cell omics and spatial omics is almost regarded as an inevitable trend, because the pathology of AD does not occur uniformly throughout the brain but is highly dependent on the vulnerability of specific brain regions. Combining the molecular state and spatial position for analysis can more directly reveal which cells and which regions undergo changes first during the disease process. In recent years, deep learning frameworks such as contrastive learning and domain adaptive models have been helping to align different datasets and reduce batch differences. Methods such as GraphCellNet and NicheTrans demonstrate how the combination of the two types of data can more clearly depict the spatial features in AD, from spatial domains
RkJQdWJsaXNoZXIy MjQ4ODYzNA==