High-throughput digital twin framework for predicting neurite deterioration using MetaFormer attention

TL;DR

The paper addresses the challenge of predicting neurite deterioration in neurodevelopmental disorders by developing a high-throughput digital twin that couples an IGA-based phase-field synthetic data generator with a MetaFormer-based gated spatiotemporal attention model. The approach uses a three-module framework (synthetic data generator, experimental data, and MetaFormer-based predictor) and a combined MSE-perceptual loss to achieve accurate long-horizon predictions, reporting average errors of $1.9641\%$ on synthetic data and $6.0339\%$ on experimental data. Key contributions include the 134 synthetic simulations, a CNN-encoder–decoder with a gSTA-augmented MetaFormer backbone, and an integrated training regimen that enables fast inference (fractions of a second per sequence) and guidance for experimental design. The framework aims to reduce experimental costs and time while advancing understanding of neurite deterioration, with potential extensions to more complex neuronal networks and multimodal data for broader neuroscience applications.

Abstract

Neurodevelopmental disorders (NDDs) cover a variety of conditions, including autism spectrum disorder, attention-deficit/hyperactivity disorder, and epilepsy, which impair the central and peripheral nervous systems. Their high comorbidity and complex etiologies present significant challenges for accurate diagnosis and effective treatments. Conventional clinical and experimental studies are time-intensive, burdening research progress considerably. This paper introduces a high-throughput digital twin framework for modeling neurite deteriorations associated with NDDs, integrating synthetic data generation, experimental images, and machine learning (ML) models. The synthetic data generator utilizes an isogeometric analysis (IGA)-based phase field model to capture diverse neurite deterioration patterns such as neurite retraction, atrophy, and fragmentation while mitigating the limitations of scarce experimental data. The ML model utilizes MetaFormer-based gated spatiotemporal attention architecture with deep temporal layers and provides fast predictions. The framework effectively captures long-range temporal dependencies and intricate morphological transformations with average errors of 1.9641% and 6.0339% for synthetic and experimental neurite deterioration, respectively. Seamlessly integrating simulations, experiments, and ML, the digital twin framework can guide researchers to make informed experimental decisions by predicting potential experimental outcomes, significantly reducing costs and saving valuable time. It can also advance our understanding of neurite deterioration and provide a scalable solution for exploring complex neurological mechanisms, contributing to the development of targeted treatments.

Figures (8)

• Figure 1: Overview of the proposed digital twin framework for predicting neurite deterioration. The framework combines an IGA-based synthetic data generator (red module) and experimental neuron culture images (green module) to adapt the model to different scenarios. The MetaFormer-based ML model (blue module) with gSTA provides accurate neurite deterioration predictions, and the experimental process (purple module) integrates real-world data to guide experimental design and decision-making.
• Figure 2: MetaFormer-based architecture overview. The proposed model predicts future neurite growth frames from input sequences via a spatial encoder, spatiotemporal extractor, and spatial decoder. The encoder compresses spatial dimensions to a latent representation. Repeated multiple times, the spatiotemporal extractor combines depth-wise and dilated convolutions with MLPs. The decoder upsamples features to reconstruct output frames, effectively modeling spatial-temporal dynamics of neurite growth.
• Figure 3: gSTA for learning temporal characteristics tan2022simvp. (A) Depth-wise convolution convolution to collect information from the local reception field. (B) Dilated convolution to obtain relatively distant pixel information. (C) Channel-wise convolution is applied across time dimensions to capture temporal features.
• Figure 4: (A&B) Prediction for single and 2-neuron synthetic neurite deterioration. The input sequence (top row) consists of frames from time steps 0–9, and the ground truth (second row) covers time steps 10–29. The predicted sequence (third row) demonstrates that the model can predict future neurite deterioration. Absolute error maps (fourth row) highlight pixel-wise discrepancies, with MRE percentages shown for sampled frames.
• Figure 5: (A&B) Prediction for 3-neuron synthetic neurite deterioration. The layout follows the same structure as Figure \ref{['fig:synthetic_predictions_1']}.