Computational modeling is crucial for understanding how neuronal features influence brain function. While the impact of subcellular features like dendrites on single-neuron computation is well-studied, their role in network-level operations remains largely unexplored. This gap is due to limitations in existing modeling tools. Current SNNs, while efficient, are often simplistic, neglecting essential dendritic properties. Conversely, detailed biophysical models, while accurate, are computationally expensive and unsuitable for large-network simulations. This paper introduces Dendrify, a new tool designed to incorporate dendritic features into SNNs efficiently and realistically. Dendrites, the branched extensions of neurons, receive most synaptic inputs and possess significant computational capabilities including acting as semi-independent thresholding units generating dendritic spikes (dSpikes), operating on multiple timescales, enabling coincidence detection and nonlinear processing, and impacting synaptic input integration and plasticity. The arrangement of synapses along dendrites significantly influences local and somatic responses, with inhibitory pathway location relative to excitatory inputs being crucial. Furthermore, dendritic morphology and passive properties shape electrotonic properties, influencing the amplitude and kinetics of synaptic currents. The complexity of dendritic processing limits the computational power of dendrite-ignorant SNNs. Biophysical models with detailed morphology are ideal for single-cell studies, but their computational cost makes them impractical for large-scale network simulations. A middle ground involves simplified models capturing essential electrophysiological characteristics. Studies show that incorporating dendritic mechanisms improves network model performance, such as associative learning, input discrimination, information binding, memory capacity, and reducing trainable parameters in artificial neural networks (ANNs). This work addresses the challenge of incorporating dendritic properties into SNNs by developing Dendrify.

The literature extensively documents the importance of dendrites in neuronal computation. Studies using detailed biophysical models have shown that dendrites are not simply passive integrators but active computational units. They are capable of performing complex operations such as coincidence detection, nonlinear signal processing, and input segregation. However, these detailed models are computationally expensive and difficult to scale to large networks. Conversely, simpler point neuron models, widely used in SNNs, ignore these crucial dendritic features, limiting their biological realism and potentially their computational capabilities. Several studies have demonstrated the benefits of incorporating simplified dendritic mechanisms into network models, showing improvements in learning, memory, and pattern separation. However, a unifying framework and readily accessible tools for incorporating these dendritic features into SNNs have been lacking.

Dendrify is a free, open-source Python package built on the Brian 2 simulator. It automates the generation of reduced compartmental neuron models with simplified, biologically relevant dendritic and synaptic properties. Users can easily add dendrites and dendritic mechanisms to SNNs using simple commands. Dendrify's internal library supports various neuronal mechanisms, and users can also provide custom model equations. A novel phenomenological approach to modeling dSpikes is significantly more efficient and mathematically tractable than the Hodgkin-Huxley formalism. The paper presents a step-by-step guide for designing reduced compartmental models, combining a theoretical framework with a tool for adding dendrites to simple, phenomenological neuronal models in a standardized and concise way. The paper demonstrates Dendrify's features through four modeling paradigms of increasing complexity: 1) A basic compartmental model with passive dendrites, showing signal attenuation and segregation. 2) A reduced compartmental model with active dendrites, demonstrating branch-specific integration rules, supralinear summation of correlated inputs, and backpropagating action potentials (BPAPs). 3) A simplified model of a CA1 pyramidal neuron validated against experimental data, replicating various passive and active properties. 4) A network of CA1 neurons, assessing the role of dendritic Na spikes in coincidence input detection. The scalability and low computational cost of Dendrify are also demonstrated by comparing execution times for single-cell and network models of varying complexity and size.

Dendrify successfully generates reduced compartmental models replicating key dendritic features. Even simple passive dendritic models demonstrate signal attenuation and compartmentalization. The incorporation of voltage-gated ion channels (VGICs) using a phenomenological approach allows for the simulation of active dendritic properties such as dendritic spikes (dSpikes) and backpropagating action potentials (BPAPs). The model shows branch-specific integration rules and supralinear summation of correlated inputs, mimicking biological behavior. A simplified CA1 pyramidal neuron model, built using Dendrify, accurately reproduces several experimental observations regarding passive and active properties, including distance-dependent attenuation of EPSPs and coincidence detection. A network simulation of CA1 neurons demonstrates that dendritic Na spikes significantly increase the efficacy of coincidence detection, leading to a substantial increase in the firing rate and temporal precision of the somatic output. The scalability analysis shows that Dendrify offers a good balance between biological accuracy and computational efficiency. Simulations of networks with thousands of neurons remain feasible, making it suitable for large-scale network simulations. The runtime analysis demonstrated that Dendrify's performance scales reasonably with increasing network size and complexity, even on standard laptops and iPads.

Dendrify provides a significant advancement in the field of SNN modeling by efficiently incorporating biologically realistic dendritic features. The ability to simulate dendritic compartments with active mechanisms, such as dSpikes and BPAPs, allows for a more accurate representation of neuronal function in SNNs. The results demonstrate the importance of considering dendritic processing for understanding network-level dynamics and behavior. The observed improvements in coincidence detection and other dendritic functions highlight the potential for enhancing the computational capabilities of SNNs through the inclusion of dendritic properties. Dendrify's scalability and ease of use make it a valuable tool for both researchers and educators interested in exploring the roles of dendrites in brain function and developing bio-inspired neuromorphic computing systems. The demonstrated ability to reproduce complex, experimentally observed phenomena, such as coincidence detection in CA1 pyramidal neurons, strongly suggests the usefulness of Dendrify for addressing fundamental questions in neuroscience and advancing the field of artificial intelligence.

Dendrify offers a powerful and accessible framework for incorporating dendritic properties into SNNs. It balances biological accuracy with computational efficiency, enabling large-scale simulations previously impractical with detailed biophysical models. The results demonstrate the significant impact of dendritic computations on network-level operations. Future research directions include expanding the range of dendritic mechanisms supported by Dendrify, incorporating more sophisticated synaptic plasticity rules, and exploring the use of Dendrify in neuromorphic hardware platforms. Dendrify's open-source nature facilitates community contributions and continued development, enhancing its potential as a key tool for both neuroscience research and neuromorphic engineering.

Dendrify’s reduced compartmental models have limitations compared to detailed biophysical models. The spatial resolution is lower, and certain phenomena, such as depolarization block and complex changes in backpropagation efficiency, may not be fully replicated using the current standard models. The current version supports only a limited set of VGICs. Synaptic plasticity rules must be implemented manually. However, Dendrify is an ongoing project, and future updates will address some of these limitations.