Auditory Transduction of Protein Misfolding for Lightweight Alzheimer’s Detection
Desk rejected at ICLR 2027, so left out of the paper lists and the map.
Abstract
Early identification of neurodegenerative disorders such as Alzheimer’s disease remains a critical clinical challenge, heavily constrained by the extreme computational burden of analyzing subtle, early-stage structural changes in amyloid-β (Aβ) protein misfolding. Existing geometric deep-learning archi- tectures, such as 3D Graph Neural Networks and Equivariant GNNs, operate directly on three-dimensional atomic coordinates and require quadratic spatial attention overhead, making high- throughput population screening computationally intractable. To solve this scalability bottleneck, we propose Diagnostic Sonifica- tion, a novel cross-modal machine learning framework that con- verts protein structural dynamics into a continuous microtonal auditory representation prior to classification. Given molecular- dynamics trajectories of Aβ1−42, a cross-modal Transformer encoder extracts latent topological features from spatial atomic coordinates and backbone dihedral angles (ϕ,ψ). A neural decoder maps these latent representations to continuous audio signals structured by microtonal scales, where toxic oligomeric misfolding manifests as distinct patterns of harmonic acoustic dissonance. A Short-Time Fourier Transform (STFT) generates two-dimensional time-frequency spectrograms, which are subse- quently processed by a lightweight 2D Convolutional Neural Net- work (CNN) to detect anomalous structural folding. Preliminary experiments on simulated Aβ1−42 monomeric and oligomeric trajectories achieve an accuracy of 94.6% and an F1-score of 0.94. Crucially, our audio transduction framework requires only 4.3 GFLOPs and 35.8 ms inference latency, yielding a 4.3× reduction in computational complexity and 15% lower latency compared to 3D-GCN baselines. These results demonstrate that converting spatial geometry into time-frequency acoustics pre- serves diagnostically vital structural information while drastically lowering computational overhead, offering a scalable foundation for decentralized Alzheimer’s screening.