Abstract
# Fractal Correction Engine as a Geometric Classifier for Amyloid Misfolding: Machine Learning on Frenet-Serret Curvature Features Achieves AUC = 0.996 Across Seven Disease Families **Authors:** Adam L McEvoy **Date:** March 2026 **Keywords:** fractal correction engine, amyloid classifier, protein misfolding, logistic regression, ROC analysis, leave-one-out cross-validation, Frenet-Serret curvature, backbone geometry, Alzheimer's disease, Parkinson's disease, prion disease, ALS, type 2 diabetes, fractal dimension, Hurst exponent **Repository:** Fractal Correction Engine — Protein Folding Dynamics **Software:** `protein_fce.py`, `expanded_amyloid_analysis.py`, `fce_classifier.py` --- ## Abstract I present a geometric classifier for amyloid protein misfolding based on the Fractal Correction Engine (FCE), a mathematical framework that extracts multi-scale fractal structure from any trajectory or waveform using $\pi$-weighted curvature analysis and the zeta series $\sum n^{-3/2}$. The protein backbone $\text{C}\alpha$ trace is treated as a three-dimensional space curve, and the Frenet-Serret apparatus computes local curvature $\kappa$ and torsion $\tau$ at each residue. The FCE extracts twelve geometric features — including mean curvature, fractal dimension $D_f$, Hurst exponent $H$, secondary structure fractions, contact density, golden ratio scaling score, and $\pi$-autocorrelation period — from each protein structure. An L2-regularized logistic regression classifier trained on these features achieves an area under the receiver operating characteristic curve (AUC) of $0.996$ under rigorous leave-one-out cross-validation (LOOCV) on a database of 22 protein structures (11 well-folded controls, 11 amyloid fibrils spanning seven disease families). The classifier achieves perfect classification — 100% accuracy, 100% sensitivity, 100% specificity, F1 = 1.000 — correctly identifying every amyloid structure and every normal structure when held out from training. Feature importance analysis reveals the top three discriminating features are intra-chain contact density ($|w| = 2.01$), helix fraction ($|w| = 1.02$), and mean backbone curvature ($|w| = 0.93$). Even the single feature of helix fraction alone achieves AUC = 0.992. The optimal three-feature subset (mean curvature, helix fraction, contact density) achieves AUC = 1.000 — perfect discrimination. The classifier requires only static $\text{C}\alpha$ coordinates, runs in under one second per protein, and provides a standalone screening tool for evaluating any PDB structure for amyloid characteristics. --- ## 1. Introduction ### 1.1 The Amyloid Misfolding Problem Protein misfolding diseases — collectively termed proteopathies — represent one of the greatest challenges in modern medicine. Alzheimer's disease, Parkinson's disease, amyotrophic lateral sclerosis (ALS), prion diseases, and type 2 diabetes all involve the pathological conversion of normally soluble proteins into insoluble amyloid fibrils characterized by cross-$\beta$ sheet architecture [1]. The World Health Organization estimates that Alzheimer's disease alone affects over 55 million people worldwide, with numbers projected to triple by 2050 [2]. Despite decades of research, rapid computational methods for identifying amyloid-prone conformations from protein structure remain underdeveloped. ### 1.2 The Geometric Perspective At the most fundamental level, a protein backbone is a curve in three-dimensional space. Each $\text{C}\alpha$ atom defines a point along this curve, and the sequence of these points traces a path whose local geometry — bending and twisting — encodes the entire secondary and tertiary structure of the protein. This observation places protein structure squarely within the domain of differential geometry, where the Frenet-Serret formulas provide the natural mathematical framework for describing space curves. The key insight of this work is that the geometric transformation from a well-folded protein to an amyloid fibril produces systematic, quantifiable changes in backbone curvature, torsion, and self-similarity that can be detected, measured, and classified using purely geometric features. ### 1.3 The Fractal Correction Engine The Fractal Correction Engine (FCE) is a universal mathematical framework that operates on any orbit, wave, wavelength, or waveform by using $\pi$ and local curvature to extract a fractal path identical to the observed path, enabling forward and backward trajectory prediction and wave interference mapping [3]. The FCE has been applied successfully to quantum decoherence, gravitational wave analysis, fluid dynamics (Navier-Stokes), antimatter containment, and over 120 other physical systems. In each case, the core mechanism is the same: compute local curvature, decompose the signal across multiple scales to extract self-similar (fractal) structure, and apply a $\pi$-weighted correction using the zeta-like series $\sum n^{-3/2}$. ### 1.4 From Detection to Classification In a previous study [4], we demonstrated that FCE-derived geometric features — mean curvature, fractal dimension, Hurst exponent, and a composite misfolding score — statistically separate normal proteins from amyloid conformations with $p = 7.1 \times 10^{-5}$ across seven disease families. In this work, we formalize these descriptive statistics into a rigorous predictive classifier: 1. We extract 12 FCE-derived geometric features from each protein structure2. We train an L2-regularized logistic regression classifier3. We validate using leave-one-out cross-validation (LOOCV)4. We compute ROC curves, AUC, confusion matrices, precision, recall, and F15. We systematically evaluate which feature subsets carry the most discriminative information6. We provide a standalone screening tool that classifies any PDB structure --- ## 2. The Fractal Correction Engine: Theory and Mathematics ### 2.1 Core Formula The Fractal Correction Engine applies the following correction to any input signal $\psi(x)$: $$\psi_{\text{corrected}}(x) = \psi(x) + \alpha \cdot \pi \cdot r(x) \cdot \sum_{n=1}^{N} \frac{1}{n^{3/2}} \cdot D_n(x)$$ where: - $\psi(x)$ is the raw input signal (curvature profile, torsion profile, contact map, energy landscape, or any other waveform)- $\alpha$ is the correction strength parameter (typically $\alpha = 0.1$)- $\pi$ is the geometric curvature constant, entering naturally through the relationship between curvature and circular geometry ($\kappa = 1/R$, circumference $= 2\pi R$)- $r(x) = 1/\kappa(x)$ is the local radius of curvature at position $x$- $n^{-3/2}$ is the fractal power-law weighting across scales, whose infinite sum converges to the Riemann zeta value $\zeta(3/2) \approx 2.612$- $D_n(x)$ are detail coefficients at scale $n$, extracted via multi-scale decomposition- $N$ is the number of decomposition levels (typically $N = 5$) The key mathematical insight is the convergence of the zeta series: $$\sum_{n=1}^{\infty} \frac{1}{n^{3/2}} = \zeta\left(\frac{3}{2}\right) \approx 2.612375348685488$$ This series weights finer scales more heavily than coarser ones — the characteristic signature of fractal (scale-free) structure — while still converging to a finite value, ensuring the correction is bounded. ### 2.2 Local Curvature Computation For a one-dimensional signal $f(x)$, the local curvature is the standard differential geometry formula for a plane curve: $$\kappa(x) = \frac{|f''(x)|}{(1 + f'(x)^2)^{3/2}}$$ This formula connects directly to $\pi$: the curvature $\kappa$ is the reciprocal of the radius $R$ of the osculating circle at each point, and the circumference of that circle is $2\pi R$. The constant $\pi$ is not inserted artificially — it emerges naturally from the geometry of curves. A circle of radius $R$ has constant curvature $\kappa = 1/R$, and one full revolution traces an arc length of $2\pi R = 2\pi/\kappa$. For a two-dimensional field $f(x, y)$ — such as a contact map or energy surface — the mean curvature is computed via the Hessian: $$H = \frac{(1 + f_y^2)f_{xx} - 2f_x f_y f_{xy} + (1 + f_x^2)f_{yy}}{2(1 + f_x^2 + f_y^2)^{3/2}}$$ ### 2.3 Multi-Scale Fractal Decomposition The detail coefficients $D_n(x)$ are extracted via difference-of-Gaussians filtering, a standard multi-resolution analysis technique related to wavelets and the Laplacian pyramid: $$D_n(x) = G_{\sigma_{n-1}} * \psi(x) - G_{\sigma_n} * \psi(x)$$ where $G_\sigma$ denotes Gaussian convolution with kernel: $$G_\sigma(x) = \frac{1}{\sigma\sqrt{2\pi}} \exp\left(-\frac{x^2}{2\sigma^2}\right)$$ and standard deviation $\sigma_n = n \cdot \sigma_{\text{base}}$. Each $D_n$ captures self-similar structure at scale $n$: $D_1$ extracts the finest detail, $D_2$ captures structure at twice the base scale, and so on. The $1/n^{3/2}$ weighting ensures that fine-scale detail contributes more strongly than coarse-scale trends — encoding the fractal hypothesis that the signal contains self-similar structure across scales. The total FCE correction at any point can be written as: $$\Delta\psi(x) = \alpha \cdot \pi \cdot r(x) \cdot \sum_{n=1}^{N} \frac{1}{n^{3/2}} \cdot \left[G_{\sigma_{n-1}} * \psi(x) - G_{\sigma_n} * \psi(x)\right]$$ ### 2.4 Fractal Dimension Estimation The box-counting fractal dimension $D_f$ characterizes the space-filling complexity of the signal: $$D_f = -\lim_{\epsilon \to 0} \frac{\log N(\epsilon)}{\log \epsilon}$$ where $N(\epsilon)$ is the number of boxes of side length $\epsilon$ required to cover the graph of the signal. In practice, we compute $N(\epsilon)$ at $K = 8$ logarithmically spaced box sizes and estimate $D_f$ as the negative slope of the linear regression: $$\log N(\epsilon_k) = -D_f \cdot \log \epsilon_k + c, \qquad k = 1, \ldots, K$$ A smooth, differentiable curve has $D_f = 1.0$; a space-filling curve has $D_f \to 2.0$; fractal signals occupy intermediate values. For protein curvature profiles, we observe $D_f \in [0.50, 1.30]$. ### 2.5 H