pith. sign in

arxiv: 2606.25185 · v1 · pith:GI7BVQ7Pnew · submitted 2026-06-23 · 💻 cs.LG · math-ph· math.MP

Neural operator-based digital twins for modeling amyloid-β and tau propagation and treatment optimization in Alzheimer's disease

Xiaofeng Xu , Tingting Dan , Zifan Zhou , Bin Li , Guorong Wu , Wenrui Hao This is my paper

Pith reviewed 2026-06-25 23:40 UTC · model grok-4.3

classification 💻 cs.LG math-phmath.MP
keywords Alzheimer's diseasedigital twinsneural operatorsamyloid-betatau proteinoptimal controlPET imagingreaction-diffusion
0
0 comments X

The pith

Neural operator learning builds patient-specific digital twins that predict amyloid-β and tau spread and optimize treatments

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper develops a data-driven framework to construct digital twins for Alzheimer's disease that model the propagation of amyloid-β and tau proteins on the cortical surface using reaction-diffusion dynamics. It combines neural operator learning with reduced-order representations to infer the unknown governing equations directly from sparse longitudinal PET imaging data. This yields predictive accuracies of 87% for amyloid-β and 81% for tau. The learned dynamics then support formulation of a PDE-constrained optimal control problem to design personalized therapeutic strategies. A sympathetic reader would care because the method turns observational clinical data into forecasts and treatment suggestions for individual patients.

Core claim

The central claim is that neural operator learning on reduced-order representations can infer the unknown nonlinear aggregation mechanisms of amyloid-β and tau from clinical PET observations, enabling accurate patient-specific forecasts of biomarker evolution on the cortical surface together with the solution of optimal control problems to regulate that evolution through interventions.

What carries the argument

Neural operator learning on reduced-order representations that learns the evolution operator for the biomarker concentration fields directly from data.

Load-bearing premise

The unknown nonlinear aggregation mechanisms of the proteins can be recovered by learning operators from sparse noisy longitudinal PET scans using reduced-order representations.

What would settle it

A test set of future PET scans from the same patients where the predicted biomarker distributions deviate substantially from the observed 87% and 81% accuracy levels.

Figures

Figures reproduced from arXiv: 2606.25185 by Bin Li, Guorong Wu, Tingting Dan, Wenrui Hao, Xiaofeng Xu, Zifan Zhou.

Figure 1
Figure 1. Figure 1: Digital twin framework for personalized modeling and treatment of Alzheimer’s disease. [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Population-level digital twin simulations of amyloid- [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Patient-level performance of the digital twin model for amyloid- [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Comparison of optimal dosing strategies and treatment outcomes under constant and time-dependent penalty weights for amyloid-β and tau. (A) Constant penalty α = 4 × 105 . Left: optimal control C(t) for anti-amyloid-β and anti-tau treatment. Right: evolution of the reduction percentage of the mean amyloid-β and tau concentration over time. (B) Time-dependent penalty α(t) = α1 + α2e −t/τ with α1 = 105 , α2 =… view at source ↗
Figure 5
Figure 5. Figure 5: Patient-specific treatment effects predicted by the learned digital twin for amyloid-β (top, 111 subjects) and tau (bottom, 25 subjects). Left: Final biomarker concentration with treatment versus without treatment for individual subjects in the left and right hemispheres. The dashed line indicates equality; points below the line represent treatment-induced reduction. Right: Distribution of percentage reduc… view at source ↗
Figure 6
Figure 6. Figure 6: Virtual reality platform for interactive visualization of the brain digital twin. [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Neural operator learning framework for cortical biomarker dynamics. [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗
Figure 9
Figure 9. Figure 9: Spectral projection-error analysis for left-hemisphere amyloid and tau data. Based on this error analysis, we restrict our experiments to P ∈ {1500, 2048, 4096}, which provides a balance between approximation accuracy and computational ef￾ficiency. Furthermore, we filter out subjects with large projection errors. In particular, we remove subjects whose projection error exceeds 0.1 for amyloid-β and 0.15 fo… view at source ↗
Figure 8
Figure 8. Figure 8: shows the distribution of longitudinal scans per subject for amyloid-β and tau after preprocessing. The amyloid-β cohort is larger (124 subjects) and more densely sampled, with many subjects having three or more scans. In contrast, the tau cohort (62 subjects) is dominated by subjects with only two scans, indicating more limited temporal sampling. In both datasets, the irregular and sparse longitudinal sam… view at source ↗
Figure 10
Figure 10. Figure 10: Cumulative spectral energy fractions of amyloid-β (top) and tau (bottom) on the left (left column) and right (right column) hemispheres at t = 0, 5, 10, 20 [PITH_FULL_IMAGE:figures/full_fig_p011_10.png] view at source ↗
Figure 12
Figure 12. Figure 12: Nonlinear reaction term in the population-level digital twin simulation of amyloid-β. The learned nonlinear term is evaluated along the simulated trajectory and visualized on the cortical surface [PITH_FULL_IMAGE:figures/full_fig_p012_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Nonlinear reaction term in the population-level digital twin simulation of tau. The learned nonlinear term is evaluated along the simulated trajectory and visualized on the cortical surface. 5. Coupling Amyloid and Tau We consider a coupled reaction–diffusion system describing the spatiotemporal dynamics of amyloid-beta uA(x, t) and tau uT (x, t). The amyloid-beta dynamics are assumed to be autonomous, wh… view at source ↗
Figure 14
Figure 14. Figure 14: reports the comparison under constant α in a single row of three panels: discrete-time objective values along the optimization run (left), optimized dosing trajectories C(t) (center), and spatially averaged amyloid burden under each method (right). The forward–backward sweep based on the adjoint formulation attains the lowest objective value and the fastest convergence in these panels [PITH_FULL_IMAGE:fi… view at source ↗
Figure 15
Figure 15. Figure 15: Numerical comparison of the three solvers under the decaying penalty α(t) = α1 +α2e −t/τ with (α1, α2, τ) = (105 , 1.6×106 , 1.5). (Left) objective value vs. optimization iteration; (center) optimized dosing intensity C(t) vs. time; (right) spatially averaged amyloid burden vs. time. Curve styles match [PITH_FULL_IMAGE:figures/full_fig_p014_15.png] view at source ↗
read the original abstract

Accurately predicting the spatiotemporal evolution of amyloid-$\beta$ and tau proteins at the individual level is critical for improving the diagnosis and treatment of Alzheimer's disease. We consider the problem of constructing patient-specific digital twins that model the propagation of these biomarkers on the cortical surface using reaction--diffusion dynamics. A major challenge is that the underlying nonlinear aggregation mechanisms are unknown and must be inferred from sparse, noisy, and heterogeneous longitudinal PET imaging data. To address this, we develop a data-driven framework that learns biomarker dynamics directly from clinical observations. The approach combines operator learning with reduced-order representations to infer governing equations of disease progression from data. Using this framework, we achieve predictive accuracies of 87\% for amyloid-$\beta$ and 81\% for tau. Building on the learned dynamics, we further formulate a PDE-constrained optimal control problem to design personalized therapeutic strategies that regulate pathological protein propagation. By integrating data-driven dynamical modeling with treatment optimization, the proposed digital twin framework provides an interpretable and predictive platform for understanding disease progression and enabling precision interventions in neurodegenerative disorders.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 0 minor

Summary. The manuscript proposes a neural operator-based framework combined with reduced-order representations to infer unknown nonlinear reaction-diffusion dynamics governing amyloid-β and tau propagation directly from sparse longitudinal PET imaging data. It reports predictive accuracies of 87% for amyloid-β and 81% for tau on these data and formulates a PDE-constrained optimal control problem to derive personalized therapeutic strategies that regulate protein propagation.

Significance. If the learned operators recover patient-invariant, generalizable dynamics rather than cohort-specific artifacts, the work would integrate data-driven operator learning with optimal control to produce interpretable digital twins for Alzheimer's progression modeling and precision intervention design.

major comments (3)
  1. [Abstract] Abstract: the reported predictive accuracies of 87% for amyloid-β and 81% for tau are stated without defining the accuracy metric, providing baselines, describing train/test splits or cross-validation procedure, or reporting error bars; this directly affects the central claim that the framework achieves predictive performance.
  2. [Abstract] Abstract: the framework learns the governing nonlinear aggregation mechanisms from the same longitudinal PET observations subsequently used to report accuracies, with no independent test sets or out-of-sample mechanistic validation described; this creates a circularity risk that the 'predictions' reflect in-sample fit rather than recovered dynamics.
  3. [Abstract] Abstract: the claim that operator learning on reduced-order representations can infer unknown nonlinear mechanisms from sparse, noisy, heterogeneous PET data requires evidence that the reduced-order step preserves critical spatial modes and that the learned operator generalizes under distribution shift; no such identifiability or robustness analysis is supplied.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive comments on our manuscript. We address each major comment below and commit to revisions that will clarify the abstract, validation procedures, and supporting analyses to strengthen the central claims.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the reported predictive accuracies of 87% for amyloid-β and 81% for tau are stated without defining the accuracy metric, providing baselines, describing train/test splits or cross-validation procedure, or reporting error bars; this directly affects the central claim that the framework achieves predictive performance.

    Authors: We agree that the abstract requires additional detail to support the reported accuracies. In the revised manuscript we will expand the abstract to explicitly define the accuracy metric (normalized mean squared error on regional protein concentrations), include baseline comparisons (e.g., linear autoregressive models and fixed-form reaction-diffusion PDEs), describe the patient-wise temporal hold-out splits with 5-fold cross-validation, and report standard deviations across folds as error bars. These clarifications will also appear in a dedicated methods paragraph. revision: yes

  2. Referee: [Abstract] Abstract: the framework learns the governing nonlinear aggregation mechanisms from the same longitudinal PET observations subsequently used to report accuracies, with no independent test sets or out-of-sample mechanistic validation described; this creates a circularity risk that the 'predictions' reflect in-sample fit rather than recovered dynamics.

    Authors: This concern is valid given the current abstract wording. The operator is trained on longitudinal sequences and evaluated via forward prediction on later time points held out per patient; however, to eliminate any appearance of circularity we will revise the abstract to state the temporal hold-out protocol explicitly and add a short description of cross-cohort testing. If the existing experiments do not fully separate training and test distributions, we will conduct additional out-of-sample runs for the revision. revision: partial

  3. Referee: [Abstract] Abstract: the claim that operator learning on reduced-order representations can infer unknown nonlinear mechanisms from sparse, noisy, heterogeneous PET data requires evidence that the reduced-order step preserves critical spatial modes and that the learned operator generalizes under distribution shift; no such identifiability or robustness analysis is supplied.

    Authors: We acknowledge that the manuscript would be strengthened by explicit supporting analysis. In the revision we will add a paragraph (and supplementary figures) quantifying spatial-mode preservation via reconstruction error of the reduced-order basis and will include a distribution-shift experiment training on one imaging cohort and testing on another. These results will be summarized in the revised abstract. revision: yes

Circularity Check

1 steps flagged

Reported predictive accuracies reduce to in-sample fit quality on the same PET data used for operator learning

specific steps
  1. fitted input called prediction [Abstract]
    "Using this framework, we achieve predictive accuracies of 87% for amyloid-β and 81% for tau. Building on the learned dynamics, we further formulate a PDE-constrained optimal control problem to design personalized therapeutic strategies that regulate pathological protein propagation."

    The framework infers the governing equations 'directly from clinical observations' (the same sparse longitudinal PET data). The reported accuracies are therefore the in-sample reconstruction error of the learned operator on those observations; no held-out subjects or external mechanistic validation is cited, so the 'prediction' step is statistically forced by the fitting procedure itself.

full rationale

The paper's central results are the 87%/81% accuracies and the subsequent PDE control, both obtained after learning the unknown nonlinear dynamics directly from the longitudinal PET observations via neural operator learning on reduced-order representations. No independent test cohort, out-of-sample mechanistic benchmark, or external validation is described in the provided text; the 'predictions' are therefore the model's reconstruction performance on the identical sparse, noisy inputs from which the operator was inferred. This matches the fitted-input-called-prediction pattern. The optimal-control step inherits the same learned operator and therefore inherits the same reduction. No self-citation chain or self-definitional equations are visible in the abstract, so the circularity is partial rather than total.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Abstract-only review yields minimal ledger entries; the framework rests on the domain assumption that reaction-diffusion PDEs on the cortical surface capture the dominant dynamics and that operator learning can recover unknown nonlinear terms from the given data.

axioms (1)
  • domain assumption Biomarker propagation on the cortical surface is governed by reaction-diffusion dynamics whose nonlinear aggregation terms are unknown a priori.
    Stated in the abstract as the modeling premise that the framework addresses.

pith-pipeline@v0.9.1-grok · 5741 in / 1203 out tokens · 26448 ms · 2026-06-25T23:40:04.661549+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

33 extracted references · 6 canonical work pages

  1. [2]

    The proposed framework combines forward prediction of biomarker evolution with inverse optimization of therapeutic interventions

    Discussion We develop a data-driven digital twin framework that integrates reaction–diffusion modeling on cortical surfaces, operator learning from longitudinal neuroimaging data, PDE-constrained optimal control for treatment design, and immersive visualization for interactive exploration of disease progression. The proposed framework combines forward pre...

    work page doi:10.1073/pnas.xxxxxxxxxx5 2026

  2. [4]

    SI Appendix

    Data Availability The source code and processed data necessary to reproduce the findings of this study will be made publicly available athttps://github.com/georgexxu/ neural-operator-based-brain-digital-twins/upon publica- tion. SI Appendix

  3. [5]

    The amyloid-βcohort is larger (124 subjects) and more densely sampled, with many subjects having three or more scans

    Data preprocessing and spectral representation Figure 8 shows the distribution of longitudinal scans per subject for amyloid-βand tau after preprocessing. The amyloid-βcohort is larger (124 subjects) and more densely sampled, with many subjects having three or more scans. In contrast, the tau cohort (62 subjects) is dominated by subjects with only two sca...

    2048

  4. [6]

    The prediction accuracies for the left and right hemi- spheres are summarized in Table 2 and Table 3, respec- tively

    Supplementary experiments for amyloid-βand tau data To assess the effect of spectral truncation, we test multiple choices of the number of Laplacian eigenfunctions, char- acterized by the truncation level P∈{1500, 2048, 4096}. The prediction accuracies for the left and right hemi- spheres are summarized in Table 2 and Table 3, respec- tively. NN architect...

    work page doi:10.1073/pnas.xxxxxxxxxx9 2048

  5. [7]

    Let u(x,t ) denote the biomarker field and{ϕk}the Laplace–Beltrami eigenmodes, which are orthonormal under the mass-matrix inner product

    Spectral analysis of simulated cortical biomarker fields To quantify the spatial complexity of the simulated cortical biomarker distributions, we analyze their spectral decomposition in the Laplace–Beltrami eigenbasis of the cortical surface. Let u(x,t ) denote the biomarker field and{ϕk}the Laplace–Beltrami eigenmodes, which are orthonormal under the mas...

    work page doi:10.1073/pnas.xxxxxxxxxx11 2026

  6. [8]

    Evolution of the nonlinear reaction term For the population-level digital twin simulation, we visualize the evolution of the nonlinear reaction term on the cortical surface for amyloid-βand tau in Figures 12 and 13. Fig. 12.Nonlinear reaction term in the population-level digital twin simulation of amyloid-β. The learned nonlinear term is evaluated along t...

  7. [9]

    Coupling Amyloid and Tau We consider a coupled reaction–diffusion system describing the spatiotemporal dynamics of amyloid-betauA(x,t ) and tau uT (x,t ). The amyloid-beta dynamics are assumed to be autonomous, whereas the evolution of tau depends on both tau and amyloid-beta, inducing a unidirectional coupling from amyloid-beta to tau. The governing equa...

  8. [11]

    Two other approaches for optimal control 7.1. Neural network parameterized control.We parametrize the control using a shallow neural network ˜Cθ(t) = m∑ i=1 ciσ(t+bi), θ:={(ci,bi)}m i=1.[33] Since the dosing intensity is required to be nonnegative, we introduce a smooth softplus transformation and define PNASJune 25, 2026 LaTex v2025 Vol. XXX No. XX eXXXX...

    work page doi:10.1073/pnas.xxxxxxxxxx13 2026

  9. [13]

    NPJ systems biology applications11, 134 (2025)

    K Rabiei, et al., Data-driven modeling of amyloid-β targeted antibodies for alzheimer’s disease. NPJ systems biology applications11, 134 (2025)

    2025

  10. [14]

    mathematical methods medicine2019, 6216530 (2019)

    JR Petrella, W Hao, A Rao, PM Doraiswamy, Computational causal modeling of the dynamic biomarker cascade in alzheimer’s disease.Comput. mathematical methods medicine2019, 6216530 (2019)

    2019

  11. [15]

    H Zheng, et al., Data-driven causal model discovery and personalized prediction in alzheimer’s disease.NPJ digital medicine5, 137 (2022)

    2022

  12. [16]

    JR Petrella, et al., Personalized computational causal modeling of the alzheimer disease biomarker cascade.The journal prevention Alzheimer’s disease11, 435–444 (2024)

    2024

  13. [17]

    J Wang, Y Mao, X Liu, W Hao, ADN Initiative, Learning patient-specific spatial biomarker dynamics via operator learning for alzheimer’s disease progression.npj Syst. Biol. Appl. (2026)

    2026

  14. [18]

    C Li, Y Mao, X Liu, W Hao, Data-driven spatiotemporal modeling reveals personalized trajectories of cortical atrophy in alzheimer’s disease.arXiv preprint arXiv:2511.08847(2025)

    arXiv 2025

  15. [19]

    Neuron73, 1204–1215 (2012)

    A Raj, A Kuceyeski, M Weiner, A network diffusion model of disease progression in dementia. Neuron73, 1204–1215 (2012)

    2012

  16. [20]

    A Raj, et al., Network diffusion model of progression predicts longitudinal patterns of atrophy and metabolism in alzheimer’s disease.Cell reports10, 359–369 (2015)

    2015

  17. [21]

    Alzheimer’s Dis

    Z Yu, et al., Uncovering diverse mechanistic spreading pathways in disease progression of alzheimer’s disease.J. Alzheimer’s Dis. Reports7, 855–872 (2023)

    2023

  18. [22]

    Methods Appl

    Z Zhang, Z Zou, E Kuhl, GE Karniadakis, Discovering a reaction–diffusion model for alzheimer’s disease by combining pinns with symbolic regression.Comput. Methods Appl. Mech. Eng.419, 116647 (2024)

    2024

  19. [23]

    M Raissi, P Perdikaris, GE Karniadakis, Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involving nonlinear partial differential equations.J. Comput. physics378, 686–707 (2019)

    2019

  20. [24]

    GE Karniadakis, et al., Physics-informed machine learning.Nat. Rev. Phys.3, 422–440 (2021)

    2021

  21. [25]

    machine intelligence3, 218–229 (2021)

    L Lu, P Jin, G Pang, Z Zhang, GE Karniadakis, Learning nonlinear operators via deeponet based on the universal approximation theorem of operators.Nat. machine intelligence3, 218–229 (2021)

    2021

  22. [26]

    Z Li, et al., Fourier neural operator for parametric partial differential equations.arXiv preprint arXiv:2010.08895(2020)

    Pith/arXiv arXiv 2010

  23. [27]

    S Lee, et al., Optimal control for anti-abeta treatment in alzheimer’s disease using a reaction-diffusion model.arXiv preprint arXiv:2504.07913(2025)

    arXiv 2025

  24. [28]

    W Hao, S Lenhart, JR Petrella, Optimal anti-amyloid-beta therapy for alzheimer’s disease via a personalized mathematical model.PLoS computational biology18, e1010481 (2022)

    2022

  25. [29]

    (Springer Berlin Heidelberg, Berlin, Heidelberg), pp

    K Atkinson, W Han,Differentiation and Integration over the Sphere. (Springer Berlin Heidelberg, Berlin, Heidelberg), pp. 87–130 (2012)

    2012

  26. [30]

    CH Van Dyck, et al., Lecanemab in early alzheimer’s disease.New Engl. J. Medicine388, 9–21 (2023)

    2023

  27. [31]

    JR Sims, et al., Donanemab in early symptomatic alzheimer disease: the trailblazer-alz 2 randomized clinical trial.Jama330, 512–527 (2023)

    2023

  28. [32]

    SG Mueller, et al., The Alzheimer’s disease neuroimaging initiative.Neuroimaging Clin.15, 869–877 (2005)

    2005

  29. [33]

    J Wang, W Hao, Laplacian eigenfunction-based neural operator for learning nonlinear reaction–diffusion dynamics.J. Comput. Phys. p. 114400 (2025). PNASJune 25, 2026 LaTex v2025 Vol. XXX No. XX eXXXXXXXXXX www.pnas.org/cgi/doi/10.1073/pnas.XXXXXXXXXX15 of 16 DRAFT 8 IMPLEMENTATION DETAILS OF VR

    work page doi:10.1073/pnas.xxxxxxxxxx15 2025

  30. [34]

    J Xu, X Xu, Lack of robustness and accuracy of many numerical schemes for phase-field simulations.Math. Model. Methods Appl. Sci.33, 1721–1746 (2023)

    2023

  31. [35]

    computational physics521, 113565 (2025)

    W Hao, S Lee, X Xu, Z Xu, Stability and robustness of time-discretization schemes for the allen-cahn equation via bifurcation and perturbation analysis.J. computational physics521, 113565 (2025)

    2025

  32. [36]

    DYNAMICAL SYSTEMS28, 1669–1691 (2010)

    J Shen, X Y ang, Numerical approximations of allen-cahn and cahn-hilliard equations. DYNAMICAL SYSTEMS28, 1669–1691 (2010)

    2010

  33. [37]

    M ´ecanique353, 1351–1364 (2025)

    JA Hardy, GA Higgins, Alzheimer’s disease: the amyloid cascade hypothesis.Science256, 184–185 (1992). 16 of 16www.pnas.org/cgi/doi/10.1073/pnas.XXXXXXXXXX pnas.org

    work page doi:10.1073/pnas.xxxxxxxxxx 1992