NeuroPathX: Explainable Imaging-Genetics

• NeuroPathX is an explainable deep learning framework that integrates ROI-level MRI features and pathway-level genetic data for both disease classification and discovery of pathway–brain associations.
• It employs an early-fusion architecture with cross-attention to capture complex interactions between imaging ROIs and biological pathways, enhancing interpretability.
• Experimental validation on ASD and Alzheimer’s shows high accuracy and specificity while producing cohort-level maps that link genetic pathways to brain regions.

NeuroPathX is an explainable deep learning framework for imaging-genetics in neurological disorders that uses an early fusion strategy, powered by cross-attention, to capture interactions between structural variations in the brain derived from MRI and established biological pathways derived from genetics data. It is designed to serve two coupled objectives: binary disease classification and the discovery of interpretable pathway–brain associations. The framework was validated on autism spectrum disorder and Alzheimer’s disease, and its reported outputs include both predictive performance and cohort-level maps linking biological pathways to structural brain regions (Wang et al., 22 Aug 2025).

1. Scope and scientific setting

NeuroPathX is situated in imaging-genetics, where the central problem is to relate genetic variation to neuroanatomical phenotypes in disorders such as autism spectrum disorder and Alzheimer’s disease. The method is motivated by two limitations identified in prior work. First, traditional linear association models between SNPs and imaging phenotypes may miss nonlinear interactions, distributed polygenic effects, and the fact that genes act in interacting biological pathways rather than in isolation. Second, black-box multimodal deep learning methods may perform prediction but often do not directly model interactions between specific brain regions and specific genetic units, and early-fusion transformer-style approaches may still fail to enforce consistency in learned interactions across subjects or may overfit when using large genetic encoders on small datasets (Wang et al., 22 Aug 2025).

The central design choice in NeuroPathX is to move genetics from the SNP or gene level to the pathway level and to retain MRI at the region-of-interest level. This produces a model whose latent interaction object is an attention matrix between pathways and ROIs rather than a single fused latent code. A plausible implication is that the framework is intended less as a generic multimodal classifier than as a structured hypothesis-generation system for pathway–ROI association discovery.

The paper frames NeuroPathX as both predictive and explanatory. Classification is binary—patient versus control for ASD and AD versus control for ADNI—but the same cross-attention mechanism is also interpreted as a map of disorder-related pathway–brain associations. The interpretability claim is therefore not post hoc; it is built into the architecture and regularized explicitly through sparsity and pathway-similarity losses (Wang et al., 22 Aug 2025).

2. Data representation and problem formulation

NeuroPathX takes two modalities per subject: structural MRI features and SNP-derived genetic features aggregated into biological pathways. For MRI, the input consists of ROI-level structural measurements extracted from T1-weighted MRI using FreeSurfer. Each ROI contributes four structural features: volume size, surface area, average cortical thickness, and standard deviation of cortical thickness. If there are $N_{\mathcal I}$ ROIs and $d=4$ features per ROI, the imaging input for subject $n$ is represented as

$$\mathbf{i}^n \in \mathbb{R}^{d \times N_{\mathcal I}}.$$

The paper uses the Brainnetome atlas with $N_{\mathcal I}=210$ for ACE and the Desikan–Killiany atlas with $N_{\mathcal I}=68$ for ADNI. Imaging features are normalized using mean and standard deviation computed on the training set (Wang et al., 22 Aug 2025).

The genetics pipeline begins with a subject-level SNP vector

$\mathbf{x}^n \in \mathbb{R}^{N_G}.$

SNPs are then mapped to genes using GWAS effect sizes $\beta_j$: $$\mathbf{g}_i^n = \sum_{j \in \Omega_i} \beta_j \mathbf{x}_j^n,$$ where $\Omega_i$ is the set of SNP indices assigned to gene $d=4$0. Genes are subsequently aggregated into KEGG canonical pathways: $d=4$1 where $d=4$2 is the set of genes in pathway $d=4$3. This yields the pathway-level subject representation

$d=4$4

After removing KEGG cancer pathways, the final pathway vocabulary comprises $d=4$5 pathways, with 3,907 genes for ACE and 3,943 genes for ADNI (Wang et al., 22 Aug 2025).

This formulation substantially reduces the dimensionality of the genetics modality while imposing biological structure. The paper’s stated motivation is that pathway-level genetics is better aligned with known biology and less prone to overfitting than raw SNP- or gene-level encoders. This suggests that the model’s interpretability depends not only on attention, but also on the prior biological granularity at which the genetics modality is represented.

3. Architecture and pathway-guided cross-attention

The architecture uses early fusion. Rather than first compressing each modality into a single latent vector and then combining them, NeuroPathX learns interactions directly between pathway units from genetics and ROI units from imaging. This is implemented through small modality-specific encoders. The genetics pathway vector is encoded into

$d=4$6

while imaging is encoded into separate key and value representations,

$d=4$7

The paper emphasizes that $d=4$8 and $d=4$9 are not identical, which departs from the most basic attention parameterizations and gives the model more flexibility in how pathway queries score ROIs versus how imaging information is aggregated (Wang et al., 22 Aug 2025).

The learnable projection matrices are

$n$0

The core fusion layer is the Pathway-guided Cross Attention, or PathAttn: $n$1 with

$n$2

The output $n$3 is called the Imaging-Guided pathway Concepts. The manuscript notes some typographic inconsistencies in dimensions and symbols, but its intended semantics are explicit: pathway embeddings provide the query-like signal, imaging ROI embeddings provide key and value representations, and the resulting attention matrix $n$4 encodes subject-specific pathway–ROI interactions (Wang et al., 22 Aug 2025).

A notable modeling decision is the use of ReLU followed by SoftSign rather than softmax. The paper argues that softmax forces attention weights to compete and sum to one, whereas biologically some pathways may simply be irrelevant and should receive uniformly low attention. The scaling constant is written as

$n$5

although the notation around $n$6 is described as slightly inconsistent. The final classifier is an MLP with dropout, trained with weighted cross-entropy (Wang et al., 22 Aug 2025).

4. Optimization, regularization, and explanation mechanism

NeuroPathX augments its classification loss with two regularizers defined directly over the attention matrix. The first is a sparsity loss: $n$7 where $n$8 is the target sparsity level. This encourages most pathway–ROI attention entries to be near zero. The paper states that, for computational reasons, this term is approximated by applying it to two randomly selected subjects from each batch (Wang et al., 22 Aug 2025).

The second regularizer is the pathway similarity loss. For each subject $n$9, the pathway influence vector is obtained by summing attention across ROIs: $$\mathbf{i}^n \in \mathbb{R}^{d \times N_{\mathcal I}}.$$0 At each epoch, $$\mathbf{i}^n \in \mathbb{R}^{d \times N_{\mathcal I}}.$$1 patient/control pairs are sampled, and the loss is

$$\mathbf{i}^n \in \mathbb{R}^{d \times N_{\mathcal I}}.$$2

The paper’s interpretation is that there exists a shared set of disorder-related pathways across the cohort, even if the detailed ROI-level effects differ. The $$\mathbf{i}^n \in \mathbb{R}^{d \times N_{\mathcal I}}.$$3 penalty is said to promote sparsity while also encouraging pathway-level consistency across cohorts (Wang et al., 22 Aug 2025).

The full objective is written as

$$\mathbf{i}^n \in \mathbb{R}^{d \times N_{\mathcal I}}.$$4

with

$$\mathbf{i}^n \in \mathbb{R}^{d \times N_{\mathcal I}}.$$5

The paper notes that the classification term is written without the usual leading minus sign, so the expression should be understood as the weighted cross-entropy term as presented in the manuscript.

Explanation extraction is based on the learned attention matrices themselves. Subject-specific explanations are, in principle, available through $$\mathbf{i}^n \in \mathbb{R}^{d \times N_{\mathcal I}}.$$6, but the reported biological analyses are cohort-level. Mean attention matrices are computed separately for patients and controls across test folds, summed across ROIs to obtain pathway influence vectors, and then used to extract the intersection of the top seven pathways for each group. For those pathways, four imaging ROIs with the highest attention values are identified. The paper does not present formal causal validation of attention-based explanations; instead, it treats interpretability as biological plausibility and stability of learned pathway–ROI associations (Wang et al., 22 Aug 2025).

5. Experimental evaluation

The framework was evaluated with 10-fold cross-validation on two datasets. ACE contained $$\mathbf{i}^n \in \mathbb{R}^{d \times N_{\mathcal I}}.$$7 subjects after removing related individuals, comprising 98 ASD cases and 67 controls. ADNI contained $$\mathbf{i}^n \in \mathbb{R}^{d \times N_{\mathcal I}}.$$8 subjects, comprising 168 AD cases and 270 controls. SNP preprocessing retained only autosomal SNPs, imposed genotyping rate $$\mathbf{i}^n \in \mathbb{R}^{d \times N_{\mathcal I}}.$$9, Hardy–Weinberg equilibrium $N_{\mathcal I}=210$0, minor allele frequency $N_{\mathcal I}=210$1, and hard call threshold $N_{\mathcal I}=210$2. LD clumping used $N_{\mathcal I}=210$3 and a physical distance threshold of 50 KB, leaving 44,102 SNPs for ACE and 45,074 SNPs for ADNI. Training used AdamW with learning rate $N_{\mathcal I}=210$4, weight decay $N_{\mathcal I}=210$5, $N_{\mathcal I}=210$6, $N_{\mathcal I}=210$7, batch size 128, and $N_{\mathcal I}=210$8 patient/control pairs sampled per epoch; the latent dimensions were $N_{\mathcal I}=210$9, $N_{\mathcal I}=68$0, $N_{\mathcal I}=68$1, and $N_{\mathcal I}=68$2, with 1300 epochs for ACE and 100 epochs for ADNI (Wang et al., 22 Aug 2025).

The comparison set included CCA, G-MIND, and SurvPATH, along with two ablations: NeuroPathX without $N_{\mathcal I}=68$3 and NeuroPathX without $N_{\mathcal I}=68$4. The reported metrics were Accuracy, Sensitivity, Specificity, and AUC.

Dataset	NeuroPathX result	Notable comparison
ACE	Accuracy $N_{\mathcal I}=68$5, Specificity $N_{\mathcal I}=68$6, AUC $N_{\mathcal I}=68$7	Best accuracy, specificity, and AUC
ADNI	Accuracy $N_{\mathcal I}=68$8, Specificity $N_{\mathcal I}=68$9, AUC $\mathbf{x}^n \in \mathbb{R}^{N_G}.$0	Best accuracy and specificity; SurvPATH AUC $\mathbf{x}^n \in \mathbb{R}^{N_G}.$1

On ACE, NeuroPathX outperformed all baselines on accuracy, specificity, and AUC. On ADNI, it achieved the best accuracy and specificity, while SurvPATH attained a slightly higher AUC and sensitivity. The paper therefore presents NeuroPathX as the most balanced model overall rather than the dominant method on every individual metric (Wang et al., 22 Aug 2025).

The ablations support both custom losses. Removing pathway similarity loss reduced ACE accuracy from $\mathbf{x}^n \in \mathbb{R}^{N_G}.$2 to $\mathbf{x}^n \in \mathbb{R}^{N_G}.$3 and ADNI accuracy from $\mathbf{x}^n \in \mathbb{R}^{N_G}.$4 to $\mathbf{x}^n \in \mathbb{R}^{N_G}.$5. Removing sparsity loss reduced ACE accuracy to $\mathbf{x}^n \in \mathbb{R}^{N_G}.$6 and ADNI accuracy to $\mathbf{x}^n \in \mathbb{R}^{N_G}.$7. The paper interprets the pathway similarity loss as especially important. At the same time, it notes that NeuroPathX’s sensitivity is somewhat lower, possibly because the encoding and classifier networks are deliberately small.

6. Biological interpretation, limitations, and nomenclature

The reported biological findings are deliberately high-level. For ASD, NeuroPathX highlights pathways involved in perception, neuroinflammation, and cellular communication, interacting with brain regions involved in sensory processing, perception, emotion, and cognition. The paper states that these findings align with prior ASD literature and also remarks that, in ACE, each pathway interacts with a more diverse set of ROIs, while patient and control groups tend to select overlapping brain regions. For AD, the framework identifies pathways related to neuroinflammation and neurodegeneration linked to regions involved in memory, language, and social functions. Compared with ASD, the ADNI results show pathways linked to more consistent ROIs and more obvious divergence between patient and control ROI patterns (Wang et al., 22 Aug 2025).

Several limitations are explicit. Sample sizes are modest, especially for ACE. The genetics modality depends on SNP-to-gene and gene-to-pathway mapping, which is database-dependent and imperfect. The framework uses only structural MRI and SNP-derived pathway features; it does not integrate functional imaging, transcriptomics, longitudinal imaging, or broader omics. The paper also states that attention is an interpretability proxy rather than a causal proof, and that the manuscript contains some notation and architecture ambiguities whose exact semantics may depend partly on the released code. Generalizability beyond the two reported datasets remains unproven, and the model’s lower sensitivity is acknowledged as a possible consequence of compact encoders and classifier design (Wang et al., 22 Aug 2025).

The exact-name method is NeuroPathX. This distinguishes it from similarly named systems in adjacent neuroimaging literature, including NeuroPath, a twin-branch Transformer for structural–functional connectome coupling (Wei et al., 2024), and NeuroPathNet, a dynamic path-trajectory framework for inter-community functional connectivity analysis (Guo et al., 28 Oct 2025). A plausible implication is that the shared “NeuroPath” prefix denotes a family of path- or pathway-oriented modeling ideas across different domains, but the NeuroPathX paper itself defines a specific imaging-genetics model centered on pathway-guided cross-attention and attention-regularized explanation discovery.

NeuroPathX is therefore best understood as an interpretable early-fusion imaging-genetics architecture whose defining object is a pathway–ROI attention matrix. Its main technical contribution is to make that matrix both predictive and analyzable: predictive through weighted cross-entropy optimization and early-fusion multimodal interaction learning, and analyzable through sparsity and cohort-level pathway-consistency constraints. Within the reported experiments, this design yields competitive disease classification while producing biologically plausible pathway–brain association maps in autism spectrum disorder and Alzheimer’s disease (Wang et al., 22 Aug 2025).