Research map / Architectures and efficiency / RNNs and normalization

Unravelling Causal Genetic Biomarkers of Alzheimer’s Disease via Neuron to Gene-token Backtracking in Neural Architecture: A Groundbreaking Reverse-Gene-Finder Approach

AAAI 2025 · Architectures and efficiency › RNNs and normalization · approach: Mechanistic circuit interpretability · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Fine-tunes Geneformer on ROSMAP single-cell microglia data for early-AD vs non-AD classification, then uses causal tracing on masked known AD genes and an attention-weighted backtracking from most causal neurons to gene tokens to nominate novel candidate AD genes.

Problem

Identifying novel causal genetic biomarkers of Alzheimer's disease from black-box genomic foundation models, since GWAS and correlation-based methods struggle to isolate causal genes.

Approach

Adapts LLM causal tracing (ROME-style indirect effect) to a gene-token transformer, selects most causal neurons under masking of known AD genes, and backtracks through self-attention weights to score input gene tokens, mapping top tokens to candidate causal genes.

Key results

• Fine-tuned Geneformer reaches AUC 74.67% ± 6.74% over five-fold CV, with sensitivity 62.68% and specificity 73.22%.
• Backtracking yields top-10 candidate causal genes (e.g., PLXDC2, MT-CO3, DOCK4), all absent from the masked known AD gene list; literature review is used to support them.
• KEGG enrichment of identified MCGs finds Alzheimer's disease pathway significant (FDR < 0.01) along with several neurodegeneration and metabolic pathways.

Surprising findings

• The top candidate genes were supported only by post-hoc literature cross-referencing, with no experimental or held-out validation of causality.

Limitations

• Causality is defined only as statistical 'most probable' effect; no experimental validation of candidate genes.
• Modest classifier performance with high variance (AUC 74.67% ± 6.74%) on a small single-cell dataset of 351 subjects.
• Analysis restricted to one microglial cluster from prefrontal cortex and early-AD stage.

Experimental setup

Datasets: ROSMAP single-cell microglial transcriptomics (Sun et al. 2023) AlzGene top AD-associated genes

Benchmarks: ROSMAP microglial single-cell AD classification

Metrics: AUC sensitivity specificity F1-score

Builds on: Geneformer causal tracing (ROME) rank value encoding

Closest papers by meaning

• Learning Multi-Modal Biomarker Representations via Globally Aligned Longitudinal Enrichments (AAAI 2020 · similarity 0.84)
• Integrating Static and Dynamic Data for Improved Prediction of Cognitive Declines Using Augmented Genotype-Phenotype Representations (AAAI 2021 · similarity 0.84)
• Interpretable Causal Representation Learning for Biological Data in the Pathway Space (ICLR 2025 · similarity 0.82)
• DECT: Harnessing LLM-assisted Fine-Grained Linguistic Knowledge and Label-Switched and Label-Preserved Data Generation for Diagnosis of Alzheimer’s Disease (AAAI 2025 · similarity 0.81)
• CASL-VAE: Learning Structured Latent Variables from Unpaired Data for Semi-supervised Clustering and Paired Sample Generation (NeurIPS 2026 · similarity 0.81)
• Generating Hypotheses of Dynamic Causal Graphs in Neuroscience: Leveraging Generative Factor Models of Observed Time Series (ICML 2025 · similarity 0.80)