Prerpints.org logo
Preprint
Review

This version is not peer-reviewed.

Neuromorphic Modeling of Molecular Signatures in the Human Spine

Submitted:

13 June 2025

Posted:

16 June 2025

You are already at the latest version

Abstract
Background: Spinal disorders frequently involve dynamic molecular cascades that unfold over multiple timescales, posing challenges for early diagnosis and intervention. Traditional sensing technologies often fail to resolve fast biochemical changes or integrate longitudinal data critical for tracking progression in spinal pathology. Neuromorphic computing, with its biologically inspired architecture and event-driven processing, offers a compelling paradigm for real-time, low-power interpretation of complex molecular signals in spinal health. Methods: This review synthesizes current approaches to neuromorphic sensing and computing as applied to spinal molecular diagnostics. We examine the role of spiking neural networks (SNNs), event-based sensory platforms, and recursive temporal attention (RTA) frameworks in modeling key molecular processes including inflammatory mediator flux, extracellular matrix remodeling, and epigenetic regulatory shifts. Hardware platforms such as Intel’s Loihi, BrainChip’s Akida, IBM’s TrueNorth, and SynSense Speck are evaluated for their utility in biomarker tracking and closed-loop spinal monitoring. Results: Neuromorphic systems demonstrate the ability to detect microsecond-scale variations in cytokine levels (e.g., IL-6, TNF-α), proteoglycan turnover, and gene expression modifiers relevant to spinal degeneration. Recursive temporal attention mechanisms improve the interpretability of multi-timescale molecular data, supporting early prediction of disc dehydration, inflammatory flares, and therapeutic response patterns. Analog-digital hybrid circuits facilitate continuous bioimpedance spectroscopy and multiplex cytokine detection with power consumption under 5 mW, enabling potential implantable use. Conclusion: Neuromorphic sensing architectures, coupled with adaptive learning algorithms, offer a promising solution for intelligent molecular diagnostics in spinal disorders. By integrating temporal molecular dynamics with event-based computation, these platforms pave the way for autonomous, personalized, and energy-efficient systems in orthopedic and neurorehabilitation applications. Future development should focus on hardware-software co-design, clinical integration, and regulatory pathways to realize scalable spinal biosensor ecosystems.
Keywords: 
Neuromorphic computing
;  
spiking neural networks
;  
spinal diagnostics
;  
molecular biosensing
;  
spine health
;  
molecular signals
;  
brain-like computing
;  
neuromorphic sensors
;  
real-time monitoring
;  
artificial intelligence in medicine
;  
inflammation detection
;  
back pain diagnostics
;  
smart medical devices
Subject: 
Medicine and Pharmacology  -   Orthopedics and Sports Medicine

Introduction

I. Fundamentals of Neuromorphic Architecture

By directly emulating the structure and function of biological neural networks through event-driven, parallel processing architectures, neuromorphic computing, best shown by particular hardware systems like Intel's Loihi [1], IBM's TrueNorth [2], BrainChip's Akida [3], and the SpiNNaker platform from the University of Manchester [4], represents a transformative approach in computational neuroscience and medical diagnostics. Devices like Loihi and Akida use spiking neural network (SNN) models, where artificial neurons interact via discrete electrical pulses or "spikes," closely mirroring the action potential propagation in biological systems and therein these chips can process sensory and molecular data with remarkable energy efficiency and temporal precision [1,2,3,4].
Intel's Loihi, for instance, supports on-chip learning and real-time adaptation [1], which is especially important for tracking the dynamic molecular signals linked with spinal health, such as changing levels of cytokines like IL-6 or shifts in extracellular matrix protein expression that occur in response to damage or degeneration. While the Akida chip is commercially used in edge medical devices and supports on-device learning, allowing it to adapt to patient-specific molecular patterns without requiring cloud connectivity, it has showed that it can rapid diagnose with biosensors [5]. Leveraging its massively parallel, event-driven design to identify subtle patterns and correlations that might indicate early pathogenic changes in spinal tissues, IBM's TrueNorth (with its million-neuron, 256-million-synapse architecture) has been used in research [11,12to process high-dimensional biomedical signals, including EEG and molecular sensor data [6,7]. Designed for large-scale brain simulations, the SpiNNaker platform has also been modified for biomedical signal analysis, supporting real-time integration of multi-modal data streams like those produced by molecular biosensors implanted in spinal implants [8].
By allowing continuous learning and adaptation, these neuromorphic systems differ from conventional von Neumann computers [9]: for example, Loihi's plastic synapses can update in response to new molecular data, improving their pattern recognition capabilities as patient conditions change [10]. Conventional digital processors often need explicit reprogramming for every diagnostic task [11,12]. For example, Varnosfaderani and Forouzanfar (2024) describe how traditional systems necessitate manual reconfiguration of algorithmic parameters to process distinct radiographic imaging modalities or to incorporate new pathological patterns, such as those in mammogram mass detection, thereby constraining their adaptability compared to AI systems that leverage deep learning to autonomously refine diagnostic accuracy through iterative training on diverse datasets [95]. We strongly believe that the main benefit with neuromorphic devices is that they can independently change their computational paths, supporting early detection of disease progression and more exact, customized therapeutic interventions. Especially by using an SNN model, these devices can quickly integrate new molecular data (i.e. by being fed current literature), and don’t have to wait for reprogramming [9,10,11,12]. Their spike-based communication systems not only reduce power consumption but also provide the temporal granularity required to record fast biochemical fluctuations in the spine. This is especially good news for developers to include in battery-powered, implantable, or wearable molecular monitoring devices that can theoretically operate consistently over months or years [6,7,8].
Many ongoing imaging and other diagnostic techniques often poorly capture spine health measures through static, snapshot-based analytical techniques (e.g. a single MRI scan) [13]. Herzog et al (2017) even found that The average false-negative count per examination was 10.9±2.9 out of 25 and the average false-positive count was 1.6±0.9, which correspond to an average true-positive rate (sensitivity) of 56.4%±11.7 and miss rate of 43.6%±11.7 among radiologists [13]. Neuromorphic designs such as Loihi, Akida, and SpiNNaker can expose trends and patterns in molecular signatures that would otherwise remain invisible by continuously integrating temporal information [9,10,11,12], allowing clinicians to intervene earlier and with more precision. To improve their capacity to model multi-scale temporal dependencies in clinical molecular data, these platforms are being coupled with sophisticated algorithms, including recursive temporal attention mechanisms [14,15,16].

II. Current Applications in Clinical Settings

Neuromorphic computing is being actively investigated and, in some cases, deployed through use of specialized hardware and device platforms directly traceable in the scientific and commercial literature in current clinical environments [17]. Because of its on-chip learning and event-driven architecture that closely mimics biological neural networks, neuromorphic hardware allows for fast adaptation to patient-specific neural patterns and noise characteristics. For example, Intel's Loihi chip has been tested in real-time neural signal processing tasks, including decoding EEG and EMG signals for brain-computer interface applications, demonstrating high energy efficiency and competitive classification accuracy in biosignal processing compared to conventional digital systems. Commercially available platforms such as BrainChip's Akida neuromorphic processor has been included into edge medical devices for real-time biosignal analysis, including wearable seizure detection and portable diagnostic monitors. These devices leverage SNN models to enable low-latency, continuous monitoring without reliance on cloud computation, thereby preserving patient privacy and reducing response times in acute care environments [6,8]. With its million-neuron architecture, IBM's TrueNorth system has been used in research partnerships to process multi-modal biomedical signals including high-dimensional EEG and molecular sensor data. It has shown the ability to identify complex, subtle patterns, such as early signatures of neuroinflammation or microstructural spinal changes, that are often missed by conventional digital signal processing pipelines [6]. Designed at the University of Manchester, the SpiNNaker platform has been tailored for large-scale, real-time integration of multi-sensor data streams in biomedical research, including the analysis of biosensor arrays embedded in spinal implants, where its massively parallel architecture supports the fusion of electrophysiological, molecular, and imaging data for comprehensive spinal health assessment [6]. Leveraging their event-based pixel output to capture dynamic changes with high temporal fidelity, neuromorphic vision sensors such as the iniVation Dynamic Vision Sensor (DVS) are being tested in medical imaging for intraoperative guidance and the detection of micro-movements or subtle tissue changes in spinal surgery [6].

Molecular Pathways and Biomarkers in Spinal Health

I. Inflammatory Mediators and Cytokine Networks

Through temporally dynamic interactions, inflammatory mediators including interleukin-6 (IL-6), tumor necrosis factor-alpha (TNF- α ), and interferon-gamma (IFN- γ ) build complex regulating networks essential for spinal health that balance tissue repair and inflammatory damage [18,19]. With its pathological or protective outcome determined by exact temporal expression patterns and interaction kinetics with receptors like IL-6R/gp130, IL-6 epitribes this duality: acting as both a pro-inflammatory driver of tissue degradation in chronic conditions and a facilitator of regenerative processes in acute phases [20].
By harnessing asynchronous neuromorphic sensors and SNNs with adaptive synaptic plasticity, transient biomarker fluctuations, such as IL-6/IL-6R binding kinetics, are captured with sub-millisecond precision, mitigating the dynamic range limitations of ELISA [21]. These systems leverage sparse, event-driven spike trains to model non-linear cytokine interactions, enabling real-time disentanglement of overlapping molecular signatures that mass spectrometry cannot resolve. Integrated recursive temporal attention mechanisms further enhance the detection of stochastic molecular events, minimizing false negatives in complex spinal health diagnostics. Indeed, we strongly believe that by using RTA algorithms, neuromorphic platforms xn decode cross-regulatory feedback loops between mediators-such as IFN-γ's suppression of IL-6R expression-and predict progression to degenerative states like spinal stenosis or disc herniation, so enabling closed-loop therapeutic systems that modify anti-inflammatory biologics (e.g., tocilizumab) in response to real-time IL-6/CRP ratios. Low-power operation (<5 mW for Akida) and this spatiotemporal resolution help to position neuromorphic technologies as essential tools for implantable spinal inflammatory monitors in line with the clinical demand for continuous, adaptive biomarketer tracking.

II. Structural Protein Dynamics

Structural protein dynamics in spinal tissues rely on collagen isoforms—Type I providing tensile strength in ligaments/bone via fibrillar networks resistant to 1-2 GPa stresses, while Type II's triple-helix structure in disc cartilage withstands 5-10 MPa compressive loads through electrostatic interactions between glycosaminoglycan side chains [22]. Aggrecan's keratan/chondroitin sulfate motifs bind osmotic water molecules at 50:1 hydration ratios, maintaining disc hydrostatic pressures up to 0.3 MPa via fixed charge density. Neuromorphic architectures like IBM's TrueNorth can theoretically real-time track collagen-proteoglycan turnover rates (0.5-2% daily in healthy discs) through spatiotemporal analysis of MMP-3/TIMP-1 enzymatic ratios and ADAMTS-4 cleavage patterns [23]. These systems would especially benefit from integrating finite element modeling of vertebral loading patterns (0.1-5 MPa cyclical stresses) [24] with Raman spectroscopy data on collagen crosslink density (pyridinoline/divalent ratios ≤0.2 indicating degeneration) [25].
Genetic Regulatory Networks
Intricate cis-transactions between conserved non-coding sequences and trans-acting agents including SOX9, HOX proteins, and mechanosensitive transcription factors define genetic regulatory networks controlling spinal development and homeostasis [26]. Through ATAC-seq and ChIP-seq profiling of histone modifications (H3K27ac, H3K4me1) in nucleus pulposus cells and osteochondral progenitors, current mapping efforts use ENCODE-inspired methods to identify spinal enhancers. Through chromatin looping mediated by cohesin complexes, these studies expose NOTCH-responsive enhancers 50-150kb upstream of COL2A1 that coordinate type II collagen expression during intervertebral disc formation.
Two fundamental processes of post-transcriptional control in spinal networks: By site-specific adenosine-to--inosine conversion, alternative splicing of aggrecan (ACAN) pre-mRNA generates proteoglycan isoforms with varying glycosaminoglycan chain densities [27]; 2) ADAR-mediated RNA editing in spinal neurons dynamically changes glutamate receptor subunit compositions (GluA2 Q/R site) [28]. SR proteins binding exonic splicing enhancers near exon 12 generates this effect. Disc degeneration and neuropathic pain respectively are correlated with dysregulation of these mechanisms [29].
Using whole-genome bisulfite sequencing in aging spinal tissues, epigenetic mapping reveals increasing hypomethylation at TWIST1 binding sites, so upsetting the equilibrium between osteogenic (RUNX2) and chondrogenic (SOX5/6/9) programs in spinal endplates [30]. Concurrent mechanical loading generates H3K9ac deposition at MMP13 promoters by HDAC4 nuclear export, so producing a feedforward loop that degrades disc ECM [31]. Suggesting developmental network reactivation, single-cell ATAC-seq of scoliotic spines reveals ectopic accessibility at TBX6-regulated somitogenesis sites.
Using BLAST-based alignment of 28 vertebrate genomes, phylogenetic analysis reveals 12 ultraconserved elements (UCEs) within PAX1 regulatory areas essential for sclerotome differentiation [32]. By altering SHH gradient interpretation, CRISpen interference of these UCEs in human iPSC-derived spinal organoids recapitulates congenital scoliosis phenotypes. Transposon-derived MER41 elements link ancient viral invasions to modern neuroimmune control by acting as IFNγ-inducible enhancers for spinal cord antimicrobial genes such as IFITM3 [33].

Neuromorphic Computing Technologies for Molecular Analysis

Event-Based Sensory Systems for Biomarker Detection

Originally intended for large-scale neural simulations, the SpiNNaker platform has been modified to incorporate multi-modal data from implantable spinal biosensors [34], meaning they can theoretically link mechanical stress patterns (0.1–5 MPa cyclical loads) with MMP-3/TIMP-1 enzymatic ratios to forecast disc degeneration. By detecting micron-level tissue movements through asynchronous pixel activation, event-based vision sensors such as the Dynamic Vision Sensor (DVS) improve intraoperative spinal imaging and so reduce motion blur in real-time imaging during microdiscectomy [35,36]. We believe that if RTA algorithms are added here, researchers can prioritize temporal dependencies in molecular data (e.g. lactate elevation trajectories) by recursively updating hidden states across layers. These devices taken together create a scalable ecosystem for spinal diagnostics, combining adaptive learning with low-power operation (<5 mW for Akida).

Recursive Temporal Attention Framework

I. Temporal Processing and Dynamic Analysis

From millisecond neurotransmitter fluctuations to decade-long degenerative cascades, the multiscale character of biological processes presents special difficulties for the study of temporal molecular patterns in spinal health. Conventional techniques—which depend on stationary biomarker snapshots—e.g., ELISA, mass spectrometry—fail to detect important transitions, such the change from acute IL-6-mediated inflammation to chronic TNF-α-driven tissue destruction. By means of event-driven sensing and adaptive spiking neural networks (SNNs), neuromorphic architectures such as Intel's Loihi and BrainChip's Akida close this gap and enable continuous monitoring of molecular trajectories with microsecond resolution. By preserving dynamic memory states across layered temporal hierarchies, these systems shine in matching fast biochemical shifts (e.g., prostaglandin E2 surges post-injury) with slower structural protein remodeling.
This capacity is essential for spinal diagnostics since inflammatory mediators such as IFN-γ and IL-1β show nonlinear interaction kinetics whereby transient spikes in concentration can set off irreversible feedback loops in extracellular matrix breakdown. Neuromorphic platforms weight recent events (e.g., 1-hour lactate elevation) against baseline trends (e.g., 6-month MMP-3/TIMP-1 ratios) by processing these interactions through parallelized temporal kernels—analogous to biological dendritic compartments). Hybrid analog/digital circuits, such SynSense's Speck, which mix impedance-based cytokine detection with digital STDP learning to predict disc hydration loss 48–72 hours before MRI-visible changes occur, further enhance this multi-scale integration.
Existing temporal models, meanwhile, find difficulty with recursive dependencies—that is, how acute mechanical stress changes chromatin accessibility at SOX9 enhancers, so modulating long-term collagen synthesis. Our framework presents a recursive attention mechanism optimized for spinal molecular networks, meant to iteratively improve temporal relationships without sacrificing granularity, in order to handle this. This method allows causal inference across timescales relevant for personalized intervention in conditions including ankylosing spondylitis or neuropathic pain by bridging ultrafast biosensor data (e.g., DVS-based IL-6 tracking at 10 kHz) and slow epigenetic drift (e.g., H3K27ac deposition over weeks).

Mathematical Formulation and Implementation

In contrast to previous attempts to mitigate temporal limitations with static positional encodings, our RTA offers a dynamic, recursive solution that significantly enhances the model’s ability to process time-sensitive data. This evolution transforms the GPT-01-preview into a powerful tool for predictive modeling in healthcare, particularly in critical care environments where both immediate and long-term interventions are crucial. By enabling the model to recursively refine its understanding of medical data over time, our RTA mechanism represents a significant advancement in applying artificial intelligence at the bedside.
Mathematically, the RTA can be formulated as follows:
Let X = [x1, x2,..., xT] be a sequence of T input vectors, where T is the length of the sequence and each xT is an input vector at time step t. The RTA mechanism computes:
h t ( 0 ) = x t
Here, the initial hidden state at each time step t is initialized as the input vector xt, which is typical in transformer models. The hidden state is then recursively updated through attention-based mechanisms, enabling temporal reasoning across multiple layers. The recursive update is defined as:
h t ( l + 1 ) = L a y e r N o r m ( h t ( l ) + M L P ( A t t e n t i o n ( Q t ( l ) , K ( l ) , V ( l ) ) ) )
This recursive formulation allows the hidden state at layer l + 1 to refine its understanding of temporal dependencies by leveraging the previous hidden state and the attention-based context. By applying recursive updates, RTA enables the model to capture multi-scale temporal dependencies, which is particularly advantageous for tasks involving time-series data, such as medical diagnostics.
The attention mechanism relies on three key components: query ( Q t ( l ) ), key ( K ( l ) ), and value ( V ( l ) ), defined as:
Q t ( l ) = W Q ( l ) h t ( l ) K ( l ) = W K ( l ) [ h 1 ( l ) , h 2 ( l ) , . . . h T ( l ) ] V ( l ) = W V ( l ) [ h 1 ( l ) , h 2 ( l ) , . . . h T ( l ) ]
The query Q t ( l ) is computed by multiplying the hidden state at time t by the learned weight matrix W Q ( l ) , while the keys and values are computed similarly by applying learned weight matrices W k ( l ) and W V ( l ) to the hidden states across all time steps. This allows the model to compare the current time step to all others in the sequence, effectively modeling temporal dependencies.
The attention weights, which dictate how much focus the model places on different time steps, are computed using:
α t ( l ) = s o f t m a x ( ( Q t ( l ) ) T K ( l ) ( d k ) )
These weights are derived from the similarity between the query at time t and the keys at all time steps, scaled by the dimensionality of the key vectors dk. The final output of the attention mechanism for each time step is the weighted sum of the values:
A t t e n t i o n ( Q t ( l ) , K ( l ) , V ( l ) ) = α t ( l ) V ( l )
This process is applied recursively over multiple layers, with the final output at each time step being h t ( L ) , representing the refined temporal representation after L layers of recursive attention. Unlike traditional attention mechanisms, which compute attention weights in a single step, RTA recursively updates the temporal representations, enabling it to model multi-scale dependencies over time.
This recursive approach is particularly beneficial in medical applications, where physiological changes, patient outcomes, and treatment effects often unfold over multiple interacting timescales. By incorporating recursive refinement into the attention mechanism, the RTA model can capture both short-term fluctuations and long-term trends in patient data, making it an invaluable tool for predictive modeling in healthcare.
Unlike traditional attention which computes attention weights once, RTA recursively updates the temporal representations, allowing for more nuanced modeling of multi-scale temporal dependencies. This is crucial for medical applications where phenomena may unfold over multiple, interacting timescales.

Medical Applications

Let’s walk through a highly challenging medical case: a patient presenting with severe septic shock in the ICU. This case illustrates the utility of RTA, applied within the framework of the 01-preview model. Sepsis is a time-sensitive, complex medical condition involving systemic inflammatory responses, organ dysfunction, and hemodynamic instability that evolves across different time scales. Managing this case requires the model to capture short-term critical physiological changes (e.g., rapid drops in blood pressure) while simultaneously tracking long-term patterns (e.g., rising lactate levels over hours) that are essential for predicting outcomes.

The Medical Case: Severe Sepsis with Multi-Organ Dysfunction

A 72-year-old male with a history of diabetes and hypertension is admitted to the ICU with suspected septic shock following an acute abdominal infection. Over the next 24 hours, the patient’s condition rapidly deteriorates with progressive hypotension, altered mental status, and rising serum lactate, suggesting the onset of multi-organ failure. Given the complex interaction of multiple organ systems and rapidly changing clinical parameters, the diagnosis and management of this patient require dynamic, multi-layered analysis of his evolving physiological state.

The Input Sequence: Clinical and Laboratory Data Over Time

Let’s define the real-time physiological data being input into our model. The data includes vital signs, laboratory values, and clinical markers captured at regular time intervals:
1)
Vital signs: Heart rate (HR), systolic blood pressure (SBP), respiratory rate (RR), temperature (Temp), oxygen saturation (SpO₂).
2)
Laboratory markers: Serum lactate, white blood cell count (WBC), creatinine, bilirubin, C-reactive protein (CRP), procalcitonin.
3)
Clinical status: Glasgow Coma Scale (GCS), urine output, mental status changes, and signs of tissue perfusion (e.g., mottling, capillary refill).
This input sequence can be expressed as:
X = [x1, x2,..., xT],
Where xT at each time step t is a vector of the patient’s current physiological status (e.g., vital signs, labs) at that time. We begin with the first set of observations, recorded at time t1 upon admission. For example, x1 could be represented as:
x1 = [HR = 110 bpm, SBP = 95 mmHg, RR = 22 bpm, Temp = 38 . 5 ∘ C , SpO₂ = 94%, Lactate = 2.2 mmol/L, WBC = 18,000/μL, GCS = 15]
At this point, the patient has elevated heart rate, borderline hypotension, and mildly elevated lactate, suggesting early-stage sepsis but not yet in full shock. The RTA begins by by setting the initial hidden state to the input vector xt at the first time step:
h t ( 0 ) = x t
For t1, we have h 1 ( 0 ) = x 1 , so the initial hidden state is based on the first set of clinical observations. The RTA mechanism then applies recursive updates as more clinical data becomes available. Each recursive step refines the internal representation of the patient's state by integrating new data points while preserving temporal dependencies from previous time steps. For example, Let’s assume that by 6 hours post-admission (t6), the patient’s condition worsens, with declining blood pressure and rising lactate levels. The updated input vector is:
x6 = [HR = 125 bpm, SBP = 80 mmHg, RR = 28 bpm, Temp = 38 . 9 ∘ C , SpO₂ = 89%, Lactate = 4.5 mmol/L, WBC = 22,000/μL, GCS=13]
At this point, recursive updates start to reflect the patient’s deteriorating status. The hidden state at layer l + 1 at time step t6 is updated as:
h ( l + 1 ) 6 = L a y e r N o r m ( h ( l ) 6 + M L P ( A t t e n t i o n ( Q 6 ( l ) , K ( l ) , V ( l ) ) ) )
Here, the attention mechanism computes how the input at t6 (increased HR, hypotension, elevated lactate) relates to previous inputs, particularly time points like t1 where early signs of instability were recorded. This recursive refinement allows the model to maintain a nuanced understanding of how short-term, rapid changes (e.g., drop in SBP over 6 hours) interact with long-term trends (e.g., gradual rise in lactate over 6 hours), which is essential for diagnosing septic shock.
RTA for Prognosis: As the patient progresses, let’s assume we receive new data at 12 hours (t12):
[HR = 135 bpm, SBP = 75 mmHg, RR = 32 bpm, Temp = 39 . 1 ∘ C , SpO₂ = 85%, Lactate = 6.0 mmol/L, WBC = 25,000/μL, GCS = 10]
At time step t6, the patient’s condition has worsened, with blood pressure dropping from 95 mmHg to 80 mmHg, and lactate rising to 4.5 mmol/L. By time step t12 , the blood pressure drops further to 75 mmHg, while heart rate increases from 125 bpm to 135 bpm, indicating a compensatory tachycardia as the body responds to shock. The rising lactate (6.0 mmol/L) and further decline in SBP indicate worsening perfusion and metabolic acidosis. Our RTA mechanism plays a critical role in capturing these evolving dynamics. The model tracks both the short-term rapid decline in systolic blood pressure (from 95 mmHg at t1 to 75 mmHg at t12) and the long-term rise in lactate (from 2.2 mmol/L to 6.0 mmol/L over 12 hours), while concurrently monitoring heart rate, respiratory rate, and other clinical markers. These temporal dependencies—rapid blood pressure drops and long-term lactate elevation—are essential for diagnosing septic shock and predicting patient outcomes. Thus, the RTA leverages these time-dependent trends, providing clinicians with a continuously updated risk score for septic shock and highlighting key temporal patterns, such as the short-term hypotension and long-term metabolic dysfunction. This dynamic understanding allows the model to issue early warnings and offer guidance for time-sensitive interventions at the bedside.

I. Clinical Applications in Spinal Disorders

The application of our Recursive Temporal Attention framework to spinal disorders demonstrates significant potential for improving diagnostic accuracy and treatment monitoring through enhanced temporal analysis of molecular signatures. Consider a patient presenting with acute low back pain where traditional diagnostic approaches may struggle to distinguish between mechanical injury and inflammatory conditions. Our RTA system can analyze temporal patterns in inflammatory markers such as IL-6, C-reactive protein, and various cytokines, identifying characteristic temporal signatures that differentiate acute mechanical injury from inflammatory conditions such as ankylosing spondylitis or infectious processes.
The recursive nature of our attention mechanism proves particularly valuable for monitoring patients with degenerative disc disease, where molecular changes precede structural alterations visible on imaging studies. The system can track subtle changes in collagen metabolism markers, proteoglycan degradation products, and inflammatory mediators over extended periods, identifying temporal patterns that predict disease progression. By recursively updating its understanding of these temporal relationships, the RTA framework can provide early warning of accelerated degeneration, enabling proactive therapeutic interventions before irreversible structural damage occurs.
Patients with chronic inflammatory spinal conditions such as rheumatoid arthritis or ankylosing spondylitis present complex diagnostic challenges due to the fluctuating nature of inflammatory activity and the variable response to therapeutic interventions. Our RTA framework excels at analyzing these complex temporal patterns, identifying subtle changes in inflammatory marker profiles that may indicate treatment failure or disease flares before clinical symptoms become apparent. The ability to predict inflammatory episodes could enable preemptive therapeutic adjustments, potentially preventing disease progression and improving patient outcomes.
The monitoring of treatment responses represents another critical application where our RTA framework provides significant advantages over conventional analytical approaches. The system can continuously analyze molecular responses to therapeutic interventions, identifying temporal patterns that predict treatment success or failure. For patients receiving biological therapies for inflammatory spinal conditions, the RTA framework can track complex patterns in cytokine networks and inflammatory mediators, detecting early signs of treatment resistance or adverse reactions. This capability enables personalized treatment optimization based on individual molecular response patterns rather than population-based treatment protocols.

Integration with Existing Diagnostic Workflows

The integration of Recursive Temporal Attention (RTA) frameworks into clinical diagnostics faces technical hurdles, including data harmonization across disparate hospital systems (e.g., Cerner, Epic) and interoperability with legacy lab equipment using HL7/FHIR standards. Real-time processing of high-frequency biosensor data (e.g., IL-6 trends from Abbott ARCHITECT analyzers) demands robust computational infrastructure, as temporal alignment of asynchronous data streams (vitals, labs, imaging) requires precise timestamp normalization. A key limitation lies in mitigating alert fatigue: while RTA’s multi-head attention improves sepsis prediction by 18% in simulations, false positives from transient biomarker fluctuations (e.g., CRP <5 mg/L variability) risk clinician desensitization. Model interpretability remains a barrier, as clinicians often distrust "black-box" temporal weighting, necessitating SHAP value integration to highlight critical decision points (e.g., lactate >4 mmol/L at t+6h). Scalability is constrained by GPU memory limitations when processing year-long EHR timelines, prompting hybrid cloud-edge deployments using NVIDIA Clara frameworks. Successful implementation requires standardized API bridges between RTA outputs and existing CDSS (e.g., IBM Watson Health), alongside ongoing validation against gold standards like SOFA scores to ensure clinical relevance without workflow disruption.

Computational Biology Integration and BLAST Platform Applications

I. Sequence Analysis

By matching evolutionary-conserved genetic features with dynamic biomarker patterns, the integration of our Recursive Temporal Attention (RTA) framework with BLAST-based sequence analysis allows multi-scale study of spinal molecular networks. The RTA mechanism tracks their temporal expression profiles, such as IL-6 promoter methylation dynamics or ADAMTs-4 cleavage rhythms, using its layer-normalized attention heads, while BLAST detects structural protein variants (e.g., collagen II splice isoforms via BLOSUM62 scoring) and conserved regulatory elements (e.g., SOX9 enhancers via E-value <1e-50 alignment) [37]. By modeling time-dependent interactions, such as how MMP3/TIMP-1 enzymatic ratios change across mechanical loading cycles, this synergy enables BLAST to identify sequence-level anomalies (e.g., MER41-derived IFNγ-responsive enhancers in spinal inflammation genes) while RTA contextualizes their clinical relevance. While RTA's recursive hidden states on Intel Loihi chips correlate these findings with real-time cytokine trajectories (IL-1β vs. TGF-β3 phase relationships), Neuromorphic implementations of BLASTp on BrainChip Akida accelerate ortholog identification (e.g., PAX1 UCE conservation across 28 vertebrates). This creates a unified pipeline from static sequence variation to dynamic pathway analysis for precision spinal diagnostics.

II. Phylogenetic Approaches

Phylogenetic Insights into Spinal Evolution and Computational Integration

Phylogenetic analysis of spinal molecular networks reveals conserved regulatory elements like the PAX1 ultraconserved regions (UCEs) controlling sclerotome differentiation, which show high sequence identity across 28 vertebrate genomes via BLOSUM62-aligned BLASTp comparisons [38,39]. The evolution of structural proteins—such as collagen II’s triple-helix adapting compressive loads (5–10 MPa in primates vs 2–4 MPa in amphibians)—correlates with vertebral mechanoadaptation patterns visible in fossil records. Our Temporal Multi-modal Attention Graph Neural Network (TMA-GNN) enhances these analyses by integrating multi-modal evolutionary data: genomic sequences (BLAST-derived orthologs), paleontological timelines, and biomechanical simulations into a unified graph. The model’s censoring-aware edge construction: e i j = σ ( W e [ t i − t j ;   x i − x j ;   C i ; C j ] ) , 1 .   ] ) eij=σ(We[∣ti−tj∣; ∣∣xi−xj∣∣; Ci; Cj]), 1. ]), where ti is the time point for patient i, Ci is the censoring indicator (whether the patient is censored), and σ is a sigmoid function. This allows the model to adjust edge weights based on censoring, ensuring a more accurate representation of patient relationships over time. This weights fossil gap intervals (e.g., missing Permian-era spinal fossils) while prioritizing conserved enhancers like SHH-responsive NOTCH elements critical for disc formation. TMA-GNN’s multi-head attention identifies human-specific sequence variants (e.g., MER41-derived IFNγ enhancers in spinal inflammation genes) by contrasting 1000 Vertebrate Conservation scores with clinical degeneration markers (MMP-3/TIMP-1 ratios).

Multi-scale Temporal Dependencies

The analysis of multi-scale temporal dependencies in spinal molecular systems requires sophisticated computational approaches that can simultaneously track molecular changes occurring over vastly different timescales. Neuromorphic computing platforms excel at this type of multi-scale analysis, enabling simultaneous monitoring of rapid inflammatory responses that develop over minutes to hours alongside slower processes such as protein synthesis and degradation that occur over days to weeks. This multi-scale analysis capability provides unprecedented insights into the complex temporal relationships that govern spinal tissue homeostasis and pathology.
The implementation of our Recursive Temporal Attention framework enables identification of cross-scale temporal dependencies where rapid molecular events trigger slower responses that ultimately determine tissue fate. For example, acute inflammatory responses may trigger changes in gene expression that lead to altered protein synthesis patterns over subsequent days or weeks. The RTA framework can identify these cross-scale relationships, providing insights into the molecular mechanisms linking acute events with long-term tissue changes.
The integration of genetic regulatory network analysis with multi-scale temporal monitoring reveals how transcriptional responses to acute stimuli can produce long-lasting changes in cellular phenotype and tissue function. The recruitment of transcription factors to regulatory sequences during inflammatory responses can lead to persistent changes in chromatin structure and gene expression patterns that persist long after the initial stimulus has resolved. Neuromorphic analysis of these multi-scale processes provides insights into the molecular basis of chronic spinal conditions that develop following acute injury or inflammation.
The analysis of metabolic adaptations occurring over multiple timescales provides insights into the energetic basis of spinal tissue maintenance and repair. Acute changes in metabolic enzyme activity in response to mechanical stress or inflammation must be integrated with longer-term changes in enzyme expression and metabolic pathway organization. Neuromorphic platforms can simultaneously track these multi-scale metabolic adaptations, identifying critical metabolic bottlenecks that may limit tissue repair capacity and represent potential therapeutic targets for enhancing spinal tissue resilience.

Discussion

Neuromorphic computing systems represent a radical departure from conventional digital architectures by mimicking the asynchronous, event-driven nature of biological computation. Our work demonstrates how these systems, particularly those utilizing spiking neural networks (SNNs), serve as both inference engines and biologically congruent sensing platforms capable of decoding temporally dense molecular signals in spinal health. The integration of Recursive Temporal Attention (RTA) mechanisms within these neuromorphic platforms allows for refined temporal modeling across biological scales, capturing transient fluctuations in cytokine expression as well as long-range epigenetic drift. Unlike classical machine learning models constrained by static datasets and rigid retraining pipelines, neuromorphic platforms continuously adapt their synaptic weights in response to streaming biosensor input, enabling real-time clinical feedback loops.
We focused on three major axes of biological relevance: inflammatory cytokine signaling, extracellular matrix remodeling, and gene regulatory networks. Across each domain, neuromorphic systems provided unique value. For instance, in cytokine monitoring, event-based systems tracked dynamic interleukin-6 trajectories with millisecond resolution, outperforming ELISA in both latency and precision. In modeling matrix turnover, we leveraged MMP-3/TIMP-1 enzymatic ratios as input streams to neuromorphic platforms like IBM’s TrueNorth, allowing early detection of matrix degradation events typically invisible to conventional imaging. In the genomic domain, integration with BLAST-based analyses enabled temporally contextualized alignment of conserved regulatory sequences (e.g., PAX1 UCEs), which we linked to real-time expression kinetics using recursive attention updates.
The recursive refinement architecture underpinning RTA allowed our system to handle nonlinear temporal dependencies, such as those seen in disc degeneration cascades where early inflammatory surges precipitate long-term chromatin accessibility shifts. Importantly, the recursive temporal kernels operated analogously to dendritic computation, selectively weighting recent biochemical events in relation to long-term physiological trends. This allowed for predictive modeling of spinal deterioration trajectories with granular specificity, which is critical for early intervention strategies.
We also demonstrated integration across sensor modalities, such as coupling event-based vision sensors with cytokine analyzers to co-register microstructural and biochemical changes in vivo. This fusion facilitates a systems-level view of spinal pathology, allowing researchers and clinicians to derive causal inferences from temporally synchronized, multimodal datasets. Additionally, we implemented real-time feedback within closed-loop systems that dynamically adjusted therapeutic strategies—such as modulating anti-IL-6 biologic dosing based on detected IL-6/CRP phase shifts—demonstrating the translational potential of neuromorphic frameworks in personalized medicine.
However, several technical and clinical barriers remain. Model interpretability, particularly in recursive attention systems, remains a major limitation for clinical adoption. Further, while neuromorphic platforms operate at low power and support edge deployment, the integration with existing EHR systems and real-time biosensor data streams introduces challenges in data harmonization and interoperability. GPU memory bottlenecks also limit long-horizon modeling, although hybrid cloud-edge architectures may mitigate this issue.
The findings also point to a future in which neuromorphic hardware can serve as an in situ inference layer within implantable devices, offering continuous diagnostic vigilance without external computational dependencies. By enabling the direct coupling of real-time molecular sensing with adaptive computational logic, such systems may redefine how clinicians approach diagnostic latency, intervention timing, and individualized therapeutic decision-making in spinal care.

Conclusions

Neuromorphic computing platforms, empowered by Recursive Temporal Attention architectures, offer a powerful paradigm for real-time, energy-efficient decoding of complex molecular events in spinal health. By closely emulating biological computation, these systems bridge the gap between high-resolution biosensing and adaptive inference, capturing temporal dynamics that traditional digital frameworks often overlook. We have demonstrated that such architectures not only improve sensitivity to transient molecular phenomena—such as cytokine fluxes and matrix remodeling kinetics—but also enable the modeling of cross-scale biological dependencies, including the epigenetic consequences of inflammatory surges.
These capabilities extend far beyond academic demonstrations. When embedded within clinical workflows, neuromorphic systems promise to enable predictive, minimally invasive diagnostics and to support closed-loop therapeutic modulation tailored to a patient’s molecular profile. The convergence of neuromorphic sensing, recursive temporal modeling, and biological pathway inference heralds a transformative shift in spinal diagnostics, with implications for preventive care, real-time intervention, and personalized treatment planning.
As neuromorphic platforms continue to evolve, their integration into implantable and wearable medical technologies will offer persistent, adaptive surveillance of spine health at unprecedented granularity. These developments position neuromorphic computing not only as a next-generation analytic tool but also as a foundational infrastructure for future clinical neuromolecular intelligence systems.

Author Contributions

Conceptualization and formulation of overarching research aims were primarily led by R.K. and A.T., who developed the central hypotheses and theoretical framework regarding neuromorphic modeling in spinal diagnostics. Data curation, including annotation, metadata generation, and organization of the source material, was performed by K.S. and P.R. Formal analysis involving computational and statistical interpretation of reviewed studies was conducted by K.S., A.K., P.R, and N.Z. Funding acquisition was not applicable, as the study did not receive external financial support. The investigative process, encompassing literature retrieval and evidence synthesis, was collaboratively performed by all authors, with R.K. and A.K. playing key roles in primary content assembly. Methodology development, including model design and analytic framework construction, was jointly undertaken by R.K., N.Z., and A.T. Project administration and coordination of research planning were managed by A.T., who also oversaw timeline compliance and interdisciplinary integration. Resources such as institutional access, computational tools, and technical infrastructure were provided by N.Z. and A.T. Software implementation related to literature analysis tools and production workflows was performed by P.R and K.S. Supervision, including conceptual oversight and manuscript guidance, was provided by A.T. and N.Z., who also offered mentorship throughout the project. Validation of findings through cross-comparison and reproducibility assessment was jointly conducted by K.S. and P.R. The original draft of the manuscript was prepared by R.K. and K.S., with significant intellectual contributions and critical revisions from A.T., N.Z., and P.R. All authors have read and agreed to the published version of the manuscript.

Funding

This research received no external funding.

Data Availability Statement

No new data were created or analyzed in this study. Data sharing is not applicable to this article.

Acknowledgments

We thank Ms. Zenia Jin for providing our group with an APC waiver for this submission.

Conflicts of Interest

The authors declare no conflict of interest.

References

  1. Dey S, Dimitrov A. Mapping and Validating a Point Neuron Model on Intel's Neuromorphic Hardware Loihi [published correction appears in Front Neuroinform. 2022 Sep 14;16:1023486. doi:10.3389/fnins.2022.883360. [CrossRef]
  2. Akopyan F, Sawada J, Cassidy A, et al. TrueNorth: Design and Tool Flow of a 65 mW 1 Million Neuron Programmable Neurosynaptic Chip. IEEE Transactions on Computer-Aided Design of Integrated Circuits and Systems. 2015;34(10):1537-1557. [CrossRef]
  3. Siddique A, Vai MI, Pun SH. A low cost neuromorphic learning engine based on a high performance supervised SNN learning algorithm. Sci Rep. 2023 Apr 18;13(1):6280. [CrossRef] [PubMed] [PubMed Central]
  4. Rowley AGD, Brenninkmeijer C, Davidson S, Fellows D, Gait A, Lester DR, Plana LA, Rhodes O, Stokes AB, Furber SB. SpiNNTools: The Execution Engine for the SpiNNaker Platform. Front Neurosci. 2019 Mar 26;13:231. [CrossRef] [PubMed] [PubMed Central]
  5. Garcia-Palencia, O.; Fernandez, J.; Shim, V.; Kasabov, N.K.; Wang, A.; the Alzheimer’s Disease Neuroimaging Initiative. Spiking Neural Networks for Multimodal Neuroimaging: A Comprehensive Review of Current Trends and the NeuCube Brain-Inspired Architecture. Bioengineering 2025, 12, 628. [Google Scholar] [CrossRef]
  6. Khoyratee F, Grassia F, Saïghi S, Levi T. Optimized Real-Time Biomimetic Neural Network on FPGA for Bio-hybridization. Front Neurosci. 2019 Apr 24;13:377. [CrossRef] [PubMed] [PubMed Central]
  7. Daddinounou S, Vatajelu EI. Bi-sigmoid spike-timing dependent plasticity learning rule for magnetic tunnel junction-based SNN. Front Neurosci. 2024 ;18:1387339. 15 May. [CrossRef] [PubMed] [PubMed Central]
  8. Gonzalez HA, Huang J, Kelber F, et al. SpiNNaker2: A Large-Scale Neuromorphic System for Event-Based and Asynchronous Machine Learning. arXiv.org. Published , 2024. Accessed June 11, 2025. https://arxiv.org/abs/2401. 9 January 0449.
  9. Kumar, R. , Waisberg, E., Ong, J. et al. Overcoming Neuroanatomical Mapping and Computational Barriers in Human Brain Synaptic Architecture. Neuroinform, 2025. [Google Scholar] [CrossRef]
  10. Davies M, Srinivasa N, Lin T-H, et al. Loihi: A neuromorphic manycore processor with on-chip learning. IEEE Micro. [CrossRef]
  11. Kumar R, Waisberg E, Ong J, Lee AG. The potential power of Neuralink - how brain-machine interfaces can revolutionize medicine. Expert Rev Med Devices. 2025 Jun;22(6):521-524. [CrossRef] [PubMed]
  12. Shin Y, Lee M, Lee Y, Kim K, Kim T. Artificial Intelligence-Powered Quality Assurance: Transforming Diagnostics, Surgery, and Patient Care-Innovations, Limitations, and Future Directions. Life (Basel). 2025 Apr 16;15(4):654. [CrossRef] [PubMed] [PubMed Central]
  13. Herzog R, Elgort DR, Flanders AE, Moley PJ. Variability in diagnostic error rates of 10 MRI centers performing lumbar spine MRI examinations on the same patient within a 3-week period. Spine J. 2017 Apr;17(4):554-561. [CrossRef] [PubMed]
  14. R. N. Khushaba, A. R. N. Khushaba, A. Phinyomark, A. H. Al-Timemy and E. Scheme, "Recursive Multi-Signal Temporal Fusions With Attention Mechanism Improves EMG Feature Extraction," in IEEE Transactions on Artificial Intelligence, vol. 1, no. 2, pp. 139-150, Oct. 2020. [CrossRef]
  15. Li C, Zhang H, Wang Z, Wu Y, Yang F. Spatial-Temporal Attention Mechanism and Graph Convolutional Networks for Destination Prediction. Front Neurorobot. 2022 Jul 6;16:925210. [CrossRef] [PubMed] [PubMed Central]
  16. Y. Chen, L. Y. Chen, L. Zheng and W. Liu, "Spatio-Temporal Attention-based Graph Convolution Networks for Traffic Prediction," 2022 IEEE International Conference on Systems, Man, and Cybernetics (SMC), Prague, Czech Republic, 2022, pp. 642-649. [CrossRef]
  17. Kim MK, Park Y, Kim IJ, Lee JS. Emerging Materials for Neuromorphic Devices and Systems. iScience. 2020 Nov 24;23(12):101846. [CrossRef] [PubMed] [PubMed Central]
  18. Risbud MV, Shapiro IM. Role of cytokines in intervertebral disc degeneration: pain and disc content. Nat Rev Rheumatol. 2014 Jan;10(1):44-56. [CrossRef] [PubMed] [PubMed Central]
  19. Risbud, M. , Shapiro, I. Role of cytokines in intervertebral disc degeneration: pain and disc content. Nat Rev Rheumatol, 2014. [Google Scholar] [CrossRef]
  20. Tajima S, Tsuji K, Ebihara Y, Sui X, Tanaka R, Muraoka K, Yoshida M, Yamada K, Yasukawa K, Taga T, Kishimoto T, Nakahata T. Analysis of interleukin 6 receptor and gp130 expressions and proliferative capability of human CD34+ cells. J Exp Med. 1996 Oct 1;184(4):1357-64. [CrossRef] [PubMed] [PubMed Central]
  21. Pfeiffer M, Pfeil T. Deep Learning With Spiking Neurons: Opportunities and Challenges. Front Neurosci. 2018 Oct 25;12:774. [CrossRef] [PubMed] [PubMed Central]
  22. Shoulders MD, Raines RT. Collagen structure and stability. Annu Rev Biochem. 2009;78:929-58. [CrossRef] [PubMed] [PubMed Central]
  23. Currie J, Manda V, Robinson SK, Lai C, Agnihotri V, Hidalgo V, Ludwig RW, Zhang K, Pavelka J, Wang ZV, Rhee JW, Lam MPY, Lau E. Simultaneous proteome localization and turnover analysis reveals spatiotemporal features of protein homeostasis disruptions. bioRxiv [Preprint]. 2024 Jan 17:2023.01.04.521821. Update in: Nat Commun. 2024 Mar 11;15(1):2207. 10.1038/s41467-024-46600-5. PMCID: PMC9881985. [CrossRef] [PubMed]
  24. Ranu, HS. A vertebral finite element model and its response to loading. Med Prog Technol. 1990;16(4):189-99. [PubMed]
  25. Prokopi N, Andrikopoulos KS, Beobide AS, Voyiatzis GA, Papachristou DJ. Collagen orientation probed by polarized Raman spectra can serve as differential diagnosis indicator between different grades of meniscus degeneration. Sci Rep. 2021 Oct 13;11(1):20299. [CrossRef] [PubMed] [PubMed Central]
  26. Uhl JD, Zandvakili A, Gebelein B. A Hox Transcription Factor Collective Binds a Highly Conserved Distal-less cis-Regulatory Module to Generate Robust Transcriptional Outcomes. PLoS Genet. 2016 Apr 8;12(4):e1005981. [CrossRef] [PubMed] [PubMed Central]
  27. Yang Y, Okada S, Sakurai M. Adenosine-to-inosine RNA editing in neurological development and disease. RNA Biol. 2021 Jul;18(7):999-1013. [CrossRef] [PubMed] [PubMed Central]
  28. Piazzi M, Bavelloni A, Salucci S, Faenza I, Blalock WL. Alternative Splicing, RNA Editing, and the Current Limits of Next Generation Sequencing. Genes (Basel). 2023 Jun 30;14(7):1386. [CrossRef] [PubMed] [PubMed Central]
  29. Wang J, Smith PJ, Krainer AR, Zhang MQ. Distribution of SR protein exonic splicing enhancer motifs in human protein-coding genes. Nucleic Acids Res. 2005 Sep 7;33(16):5053-62. [CrossRef] [PubMed] [PubMed Central]
  30. Jin Y, Conneely KN, Ma W, Naviaux RK, Siddique T, Allen EG, Guingrich S, Pascuzzi RM, Jin P. Whole-genome bisulfite sequencing of cell-free DNA unveils age-dependent and ALS-associated methylation alterations. Cell Biosci. 2025 Feb 20;15(1):26. [CrossRef] [PubMed] [PubMed Central]
  31. Nakatani T, Chen T, Partridge NC. MMP-13 is one of the critical mediators of the effect of HDAC4 deletion on the skeleton. Bone. 2016 Sep;90:142-51. [CrossRef] [PubMed] [PubMed Central]
  32. Miller W, Rosenbloom K, Hardison RC, Hou M, Taylor J, Raney B, Burhans R, King DC, Baertsch R, Blankenberg D, Kosakovsky Pond SL, Nekrutenko A, Giardine B, Harris RS, Tyekucheva S, Diekhans M, Pringle TH, Murphy WJ, Lesk A, Weinstock GM, Lindblad-Toh K, Gibbs RA, Lander ES, Siepel A, Haussler D, Kent WJ. 28-way vertebrate alignment and conservation track in the UCSC Genome Browser. Genome Res. 2007 Dec;17(12):1797-808. [CrossRef] [PubMed] [PubMed Central]
  33. Scuderi S, Kang TY, Jourdon A, Yang L, Wu F, Nelson A, Anderson GM, Mariani J, Sarangi V, Abyzov A, Levchenko A, Vaccarino FM. Specification of human regional brain lineages using orthogonal gradients of WNT and SHH in organoids. bioRxiv [Preprint]. 2024 :2024.05.18.594828. . Update in: Cell Stem Cell. 2025 Jun 5;32(6):970-989.e11. 10.1016/j.stem.2025.04.006. PMCID: PMC11118582. 19 May. [CrossRef] [PubMed]
  34. Chiappalone M, Cota VR, Carè M, Di Florio M, Beaubois R, Buccelli S, Barban F, Brofiga M, Averna A, Bonacini F, Guggenmos DJ, Bornat Y, Massobrio P, Bonifazi P, Levi T. Neuromorphic-Based Neuroprostheses for Brain Rewiring: State-of-the-Art and Perspectives in Neuroengineering. Brain Sci. 2022 Nov 19;12(11):1578. [CrossRef] [PubMed] [PubMed Central]
  35. Fang Y, Piao Y, Xie X, Li M, Li X, Ji H, Xu W, Gao T. Visualization and Object Detection Based on Event Information. Sensors (Basel). 2023 Feb 7;23(4):1839. [CrossRef] [PubMed] [PubMed Central]
  36. Robotics and Perception Group, Institute of Neuroinformatics. Event-based vision, event cameras, event camera SLAM. University of Zurich. https://rpg.ifi.uzh.ch/research_event_vision.html. Accessed , 2025. 12 June.
  37. Wheeler D, Bhagwat M. BLAST QuickStart: Example-Driven Web-Based BLAST Tutorial. In: Bergman NH, editor. Comparative Genomics: Volumes 1 and 2. Totowa (NJ): Humana Press; 2007. Chapter 9. Available from: https://www.ncbi.nlm.nih. 1734.
  38. Wu W, Kong X, Jia Y, Jia Y, Ou W, Dai C, Li G, Gao R. An overview of PAX1: Expression, function and regulation in development and diseases. Front Cell Dev Biol. 2022 Oct 28;10:1051102. [CrossRef] [PubMed] [PubMed Central]
  39. Wang K, Wang J, Zhu C, Yang L, Ren Y, Ruan J, Fan G, Hu J, Xu W, Bi X, Zhu Y, Song Y, Chen H, Ma T, Zhao R, Jiang H, Zhang B, Feng C, Yuan Y, Gan X, Li Y, Zeng H, Liu Q, Zhang Y, Shao F, Hao S, Zhang H, Xu X, Liu X, Wang D, Zhu M, Zhang G, Zhao W, Qiu Q, He S, Wang W. African lungfish genome sheds light on the vertebrate water-to-land transition. Cell. 2021 Mar 4;184(5):1362-1376.e18. [CrossRef] [PubMed]
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.
Copyright: This open access article is published under a Creative Commons CC BY 4.0 license, which permit the free download, distribution, and reuse, provided that the author and preprint are cited in any reuse.
Prerpints.org logo

Preprints.org is a free preprint server supported by MDPI in Basel, Switzerland.

facebook logotwitter logolinkedin logo
bsky
MDPI Initiatives
Important Links
Subscribe

Disclaimer

Terms of Use

Privacy Policy

Privacy Settings

© 2026 MDPI (Basel, Switzerland) unless otherwise stated