Connectome Inspired Neural Network

A Comprehensive Guide to Connectome-Based Modeling (2015-2025)
From complete EM reconstructions to predictive mechanistic models bridging structure and function

📑 Table of Contents

📚 Paper Collections by Organism

🔬 Detailed Analysis of Connectome-Based Modeling

🪰 Drosophila Models (Full EM Connectome)

  1. ⭐️ Turaga et al., 2024 - Landmark Study
    Single-neuron prediction (r=0.7-0.9) from FlyWire connectome

  2. Fiete et al., 2025 - Massive Parameter Reduction
    439 neurons → 57 parameters via cell-type constraints

  3. Borst, 2024 - Temporal Filtering
    Conductance-based optic lobe model

  4. Full Brain LIF Model
    140K neurons whole-brain sensorimotor processing

🪱 C. elegans Models (First Complete Connectome)

  1. ⭐️ Zhao et al., 2024 - Most Comprehensive
    Brain-body-environment closed loop with full biophysics

  2. Morrison & Young, 2025 - Premotor Circuit
    Data-driven fit to calcium imaging

  3. Creamer, Leifer & Pillow, 2024 - Theoretical Analysis
    Quantifies connectome insufficiency

🐭 Mouse/Rat Models (Partial EM + Statistical)

  1. MICrONS Consortium - Structure-Function Dataset
    100K neurons with co-registered EM and functional data

  2. Tolias et al., 2025 - Foundation Model
    Transformer-based neural activity prediction

  3. Rajan et al., 2020 - CURBD Method
    Inferring effective connectivity from dynamics

  4. ⭐️ Blue Brain Project, 2015 - Mammalian Landmark
    First mammalian cortical simulation (31K neurons, statistical)

  5. ⭐️ Billeh et al., 2020 - Allen V1 Model
    230K neurons with hybrid neuron models

  6. Potjans & Diesmann, 2014 - Canonical Circuit
    Benchmark cortical microcircuit model

🧬 Theory & Principles

  1. Beiran & Litwin-Kumar - Theoretical Limits
    Connectome alone has prediction floor

  2. General Principles Across Species

  3. The "Connectome Ladder"
    4 levels of modeling abstraction

🔑 Cross-Species Insights

📊 Summary Tables

🎯 Synthesis & Future

📚 Resources

Paper Collections by Organism

Drosophila

Functional correlations are predicted by structural connectivity in the connectome • Correlations are well captured by the number of cells connecting two regions • Mesoscale networks in the Drosophila brain share topological features with cortex • Indirect pathways differentially shape functional correlations across the brain

τdxjA(t)dt=−ℓxjA(t)+σ(∑B∈C∑kw0(1+ZAB)sgnBCjkxkB(t)+bA+uj(t))\tau \frac{d x_j^A(t)}{d t}=-\ell x_j^A(t)+\sigma\left(\sum_{B \in \mathcal{C}} \sum_k w_0\left(1+Z^{A B}\right) s g n^B C_{j k} x_k^B(t)+b^A+u_j(t)\right)

reduces the number of optimized parameters from 4392+439+1=193,161439^{2} +439+1 = 193, 161 to just 72+7+1=577^{2} +7+1 = 57 parameters Total Loss: Linear Consistency Loss, Stability Loss, Minimum Speed Loss, Entropy Loss, L1 and L2 Regularization

Nature Neuroscience volume 27, pages2443–2454 (2024)

Raw imaging data, relevant connectome data, and partially processed visual responses for all 571 ROIs are available on Dryad: https://datadryad.org/dataset/doi:10.5061/dryad.pg4f4qs1j
https://datadryad.org/dataset/doi:10.5061/dryad.bnzs7h4ns https://datadryad.org/dataset/doi:10.5061/dryad.kh18932k1

Human

Perturbation

monkey

Rodent

MICrONS

Functional connectomics spanning multiple areas of mouse visual cortex
The MICrONS Consortium

C. Elegans

Dataset

Theory-Based

cognitive inspired

the following are collected by Ruizhe Zhou

Others

Brain-inspired global-local learning incorporated with neuromorphic computing https://www.nature.com/articles/s41467-021-27653-2

potential model

Basic Architecture

Review

Researcher (TODO)

Detailed Analysis of Connectome-Based Modeling Approaches

Overview

This section provides in-depth analysis of key papers that leverage connectome data to build computational models of neural circuits and whole brains. We focus on three model organisms with complete or near-complete connectomes: Drosophila melanogaster, Caenorhabditis elegans, and Mouse, examining how structural connectivity constrains and predicts neural dynamics and behavior.

🪰 Drosophila Connectome-Based Models

1. Turaga et al., 2024 - Connectome-Constrained Deep Mechanistic Networks ⭐️ Landmark Study

Journal: Nature (2024)
Authors: Srinivas C. Turaga et al.
Link: https://www.nature.com/articles/s41586-024-07939-3

This is arguably the most sophisticated connectome-constrained neural network model to date, achieving single-neuron resolution predictions across the entire fly visual system.

Background & Motivation

The Challenge:

Innovation:

Connectome Data Utilization

Data Sources:

Connectivity Matrix Construction:

Model Architecture

Hierarchical Structure:

Visual Input → Photoreceptors → Lamina → Medulla → 
Lobula/Lobula Plate → Visual Projection Neurons

Neuron Model (Mechanistic Point Neuron):

For each neuron ii:

τidVidt=−Vi+∑jwij⋅f(Vj)+Iiinput\tau_i \frac{dV_i}{dt} = -V_i + \sum_{j} w_{ij} \cdot f(V_j) + I_i^{\text{input}}

Where:

Key Mechanistic Features:

Depth: Effectively 5-8 layers deep (from photoreceptors to output neurons)

Parameter Optimization Strategy

Two-Stage Optimization:

Stage 1: Connectome Initialization

Stage 2: Data-Driven Refinement

Free Parameters (~10,000-100,000 parameters for 60,000 neurons):

Optimization Method:

Training Data:

Computational Cost:

Validation & Results

Predictive Performance:

Key Findings:

  1. Connectome is highly constraining:

  2. Dendritic nonlinearities are essential:

  3. Emergent computations:

  4. Cell type diversity:

Comparison to Previous Approaches

Approach Turaga 2024 Traditional CNNs Detailed Compartmental Models
Biological Connectivity Full connectome Hand-designed Single neuron
Scale 60,000 neurons N/A 1 neuron
Neuron Model Mechanistic point neuron Abstract units Full HH
Prediction Accuracy High Low (wrong neurons) High (1 neuron)
Interpretability High Low High
Computational Cost Moderate Low Very High

Significance & Impact

Scientific Impact:

Technical Impact:

Limitations:

2. Fiete et al., 2025 - Head Direction Circuit with Massive Parameter Reduction

Journal: bioRxiv (2025)
Authors: Ila Fiete et al.
Title: From Synapses to Dynamics: Obtaining Function from Structure in a Connectome Constrained Model

The Parameter Reduction Problem

Traditional Approach:

Connectome-Constrained Approach:

Model Formulation

Connectome-Constrained Dynamics:

τdxjA(t)dt=−ℓxjA(t)+σ(∑B∈C∑kw0(1+ZAB)sgnBCjkxkB(t)+bA+uj(t))\tau \frac{d x_j^A(t)}{d t}=-\ell x_j^A(t)+\sigma\left(\sum_{B \in \mathcal{C}} \sum_k w_0\left(1+Z^{A B}\right) \text{sgn}^B C_{j k} x_k^B(t)+b^A+u_j(t)\right)

Parameter Structure:

Key Insight:

Optimization Strategy

Multi-Objective Loss Function:

Ltotal=Lconsistency+Lstability+Lspeed+Lentropy+Lreg\mathcal{L}_{\text{total}} = \mathcal{L}_{\text{consistency}} + \mathcal{L}_{\text{stability}} + \mathcal{L}_{\text{speed}} + \mathcal{L}_{\text{entropy}} + \mathcal{L}_{\text{reg}}

Where:

  1. Linear Consistency Loss: Activity represents head direction in a linear code
  2. Stability Loss: Bump should persist in absence of input
  3. Minimum Speed Loss: Network should update smoothly with angular velocity input
  4. Entropy Loss: Encourage distributed representations
  5. L1/L2 Regularization: Prevent overfitting, encourage sparse solutions

No Neural Data Required:

Results

Functional Emergence:

Parameter Insights:

Generalizability:

Significance

This work shows that:

  1. Connectomes dramatically reduce parameter space in neural network models
  2. Functional constraints (not neural recordings) can be sufficient for optimization
  3. Cell-type-level parameterization is a powerful middle ground between fully individual and fully shared parameters

3. Borst 2024 - Differential Temporal Filtering in Optic Lobe

Journal: bioRxiv (2024)
Authors: Alexander Borst

Model Approach

Connectome Integration:

Neuron Model:

Temporal Filtering:

Key Findings

  1. Temporal diversity is essential: Different cell types filter visual input at different timescales
  2. Spatial integration: Lateral connections shape receptive field properties
  3. Emergent motion sensitivity: Connectome + temporal parameters → direction selectivity

4. Full Brain LIF Model (Nature 2024)

Journal: Nature (2024)
Title: A Drosophila computational brain model reveals sensorimotor processing

Scale & Ambition

Whole-Brain Model:

Model Type: Leaky Integrate-and-Fire (LIF)

τmdVidt=−(Vi−Vrest)+Rm∑jgij(Vj−Vsyn)\tau_m \frac{dV_i}{dt} = -(V_i - V_{\text{rest}}) + R_m \sum_j g_{ij}(V_j - V_{\text{syn}})

Where:

Connectome Constraints

From FlyWire:

Assumptions & Limitations (acknowledged by authors):

Parameter Optimization

Minimal Free Parameters:

Optimization Method:

Results & Insights

Functional Predictions:

Network Analysis:

Limitations:

Value:

🪱 C. elegans Connectome-Based Models

The C. elegans nervous system (~302 neurons, ~7000 synapses) was the first complete connectome (1986), making it a prime target for whole-organism neural modeling.

1. Zhao et al., 2024 - Integrative Brain-Body-Environment Model ⭐️ Most Comprehensive

Journal: Nature Computational Science (2024)
Title: An integrative data-driven model simulating C. elegans brain, body and environment interactions

This is the most biophysically detailed whole-organism model to date, integrating:

Connectome Data Integration

Structural Connectivity:

Cell Type Information:

Multi-Scale Modeling Framework

1. Neuron Models (Biophysically Detailed):

Morphologically-derived multi-compartmental models:

Compartment dynamics (Hodgkin-Huxley style):

CmdVdt=−Ileak−Ichannels−Isyn+IaxialC_m \frac{dV}{dt} = -I_{\text{leak}} - I_{\text{channels}} - I_{\text{syn}} + I_{\text{axial}}

Where:

2. Synapse Models:

Chemical Synapses (Graded Transmission):

Isyn=gsyn⋅m∞(Vpre)⋅(Vpost−Esyn)I_{\text{syn}} = g_{\text{syn}} \cdot m_{\infty}(V_{\text{pre}}) \cdot (V_{\text{post}} - E_{\text{syn}})

Where m∞(Vpre)m_{\infty}(V_{\text{pre}}) is a sigmoid function of presynaptic voltage

Gap Junctions (Electrical Coupling):

Igap=ggap⋅(Vneighbor−Vself)I_{\text{gap}} = g_{\text{gap}} \cdot (V_{\text{neighbor}} - V_{\text{self}})

3. Muscle Models:

4. Biomechanical Body Model:

5. Environment:

Parameter Optimization Strategy

Challenge: Tens of thousands of parameters across neurons, synapses, muscles

Multi-Stage Hierarchical Optimization:

Stage 1: Single Neuron Parameters

Stage 2: Synaptic Parameters

Stage 3: Neuromuscular Parameters

Stage 4: Whole-System Integration

Total Parameters:

Validation & Results

Neural Dynamics:

Behavior:

Emergent Properties:

Key Insights

  1. Gap junctions are critical: Removing electrical synapses degrades many behaviors
  2. Graded transmission dominates: Most information transfer is analog, not digital (spikes)
  3. Embodiment matters: Body mechanics and environment shape neural activity patterns
  4. Multi-scale coupling: Cannot understand neurons without muscles/body, or vice versa

Significance

Scientific:

Technical:

Limitations:

2. Morrison & Young, 2025 - Data-Driven Premotor Network Model

Journal: arXiv (2025)
Authors: Megan Morrison, Lai-Sang Young
Title: A data-driven biophysical network model reproduces C. elegans premotor neural dynamics

Focus: Forward/Backward Locomotion Circuit

Subset of Connectome:

Why This Circuit:

Model Details

Neuron Model:

Data-Driven Approach:

Parameter Optimization:

Results

Reproduces Key Features:

Predictions:

Advantages:

Limitations:

3. Creamer, Leifer & Pillow, 2024 - Bridging Connectome and Whole-Brain Activity

Journal: bioRxiv (2024)
Authors: Matthew S. Creamer, Andrew M. Leifer, Jonathan W. Pillow

Key Question:

Can we predict whole-brain neural dynamics from connectome alone?

Approach

Linear Dynamical System:

x(t+1)=Wx(t)+u(t)+ϵ(t)\mathbf{x}(t+1) = \mathbf{W} \mathbf{x}(t) + \mathbf{u}(t) + \boldsymbol{\epsilon}(t)

Where:

Connectome Initialization:

Optimization

Data: Whole-brain calcium imaging (Leifer lab)

Method:

Key Findings

  1. Connectome provides strong prior:

  2. Connectome alone is insufficient:

  3. Functional motifs differ from structural:

Interpretation:

🐭 Mouse Visual Cortex Connectome-Based Models

The mouse visual cortex presents unique challenges:

1. MICrONS Consortium, 2025 - Functional Connectomics ⭐️ Game-Changing Dataset

Journal: Nature (2025)
Links:

MICrONS Dataset Overview

Unprecedented Scale:

This is the first mammalian dataset with both structure and function at scale.

Connectome Data Structure

Connectivity Matrix:

Functional Data:

Key Modeling Findings

1. Structural-Functional Connectivity Relationship:

Question: Does structural connectivity predict functional connectivity?

Approach:

Results:

2. General Wiring Rule:

Discovery: Connectivity follows predictable rules based on:

Model:

P(Cij>0∣features)=σ(β0+β1dij+β2Δθij+β3Itype)P(C_{ij} > 0 | \text{features}) = \sigma(\beta_0 + \beta_1 d_{ij} + \beta_2 \Delta\theta_{ij} + \beta_3 \mathbb{I}_{\text{type}})

Where:

Implications:

3. Predictive Models of Neural Responses:

Approach: Use connectome to constrain neural network model (similar to fly work)

Model Architecture:

Visual Input → Linear-Nonlinear encoding → Recurrent Network (connectome-constrained) → Predicted Activity

Connectome Integration:

Results:

2. Tolias et al., 2025 - Foundation Model of Neural Activity

Journal: Nature (2025)
Title: Foundation model of neural activity predicts response to new stimulus types

Beyond Connectome: Data-Driven Foundation Model

Approach:

Model: Transformer-based architecture

Scale:

Results

Generalization:

Comparison to Connectome Models:

3. Rajan et al., 2020 - Data-Constrained RNNs (CURBD)

Journal: bioRxiv (2020)
Authors: Kanaka Rajan et al.
Title: Inferring brain-wide interactions using data-constrained recurrent neural network models

Approach: Reverse-Engineering Brain-Wide Dynamics

Scale: Whole-brain calcium imaging across multiple regions (not single-neuron resolution)

Model: Recurrent Neural Network (RNN)

x(t+1)=f(Wx(t)+Winu(t))\mathbf{x}(t+1) = f(\mathbf{W} \mathbf{x}(t) + \mathbf{W}_{\text{in}} \mathbf{u}(t))

CURBD Method (Connectivity Uncovered via Recurrent-Bayesian Dynamics):

Connectome Relevance

Not directly using synaptic connectome, but:

Key Insight:

4. Blue Brain Project - Neocortical Microcircuit Reconstruction ⭐️ Mammalian Landmark

Journal: Cell (2015)
Authors: Henry Markram et al.
Title: Reconstruction and Simulation of Neocortical Microcircuitry
Link: https://www.cell.com/cell/fulltext/S0092-8674(15)01191-5

This is the first data-driven digital reconstruction of mammalian cortical tissue at cellular resolution, representing a paradigm shift in how we model complex brain circuits.

Background & Vision

The Blue Brain Project (started 2005, EPFL):

Why This Matters:

Scale & Scope

Reconstructed Volume:

This is not a connectome-based model in the traditional sense, but rather a statistically reconstructed model.

The Challenge: No Complete Connectome

Unlike flies or worms, we cannot trace every synapse in mammalian cortex. Instead:

Data-Driven Statistical Reconstruction:

  1. Neuron Positions:

  2. Morphologies:

  3. Connectivity Rules (This is the key innovation):

Touch Detection Algorithm:

For each pair of neurons (i,j)(i,j):

P(synapseij)=f(typei,typej,dij,Voverlap)P(\text{synapse}_{ij}) = f(\text{type}_i, \text{type}_j, d_{ij}, V_{\text{overlap}})

Bouton Density (synapses per connection):

Result:

Neuron Models: Multi-Compartmental Hodgkin-Huxley

For Each of 31,000 Neurons:

Morphology:

Electrical Dynamics (Hodgkin-Huxley):

CmdVdt=−∑channelsIchannel−Isyn+Iaxial+IextC_m \frac{dV}{dt} = -\sum_{\text{channels}} I_{\text{channel}} - I_{\text{syn}} + I_{\text{axial}} + I_{\text{ext}}

Ion Channels (13 types):

Channel Distributions:

Parameter Optimization:

Constraints:

Synapse Models

Detailed Synaptic Dynamics:

AMPA, NMDA, GABA_A, GABA_B receptors:

For AMPA (example):

IAMPA=gAMPA⋅(V−Eexc)⋅∑spikesα(t−tspike)I_{\text{AMPA}} = g_{\text{AMPA}} \cdot (V - E_{\text{exc}}) \cdot \sum_{\text{spikes}} \alpha(t - t_{\text{spike}})

Where α(t)\alpha(t) is a double-exponential function (rise + decay)

Synaptic Plasticity:

Connectome-Like Detail:

Simulation & Validation

Computational Challenge:

In Vivo-Like Simulation:

Spontaneous Activity:

Emergent Properties:

  1. Layer-specific activity patterns:

  2. Cell-type specific recruitment:

  3. Network oscillations:

  4. Propagation of activity:

Validation Against Experiments

Predictions Tested:

  1. Connection probabilities:

  2. Synaptic physiology:

  3. Network responses to stimulation:

Discrepancies:

Key Innovations & Contributions

Methodological:

  1. Statistical connectome generation: When you don't have EM, use touch-detection
  2. Integration framework: Combines morphology, electrophysiology, connectivity, synapse physiology
  3. Scalability: Workflow can be applied to other brain regions

Scientific:

  1. Emergent properties: Shows many cortical features arise from structure + local dynamics
  2. Testable predictions: Generates hypotheses about unmeasured connections and dynamics
  3. In silico experiments: Enables perturbations impossible in vivo (lesion specific cell types, etc.)

Open Science:

Subsequent Developments (2015-2024)

Expansion to Other Regions:

  1. Mouse Whole Neocortex Model (2024):

  2. Hippocampus CA1 (2024) (Romani et al., PLoS Biology):

Refinement of Methods:

  1. Part I: Anatomy (2024) (Reimann et al., eLife):

  2. Part II: Physiology (2024) (Isbister et al., eLife):

New Tools:

  1. Connectome-Manipulator:

Comparison: Blue Brain vs. MICrONS vs. Fly Connectome Models

Aspect Blue Brain (Rat) MICrONS (Mouse) Turaga (Fly)
Connectome Type Statistical (predicted) Partial EM (real) Full EM (real)
Scale 31K neurons, 37M synapses 100K neurons, 500M synapses 60K neurons
Neuron Model Multi-compartmental HH Point (in most models) Mechanistic point
Validation Data Electrophysiology Calcium imaging Calcium imaging
Prediction Accuracy Qualitative agreement Moderate (50-60%) High (70-90%)
Computational Cost Extreme (supercomputer) Moderate Moderate (GPU)
Strength Biophysical detail Structure-function link Single-neuron predictions
Limitation Connectome is inferred Incomplete connectome Simplified neuron model

Significance & Legacy

Scientific Impact:

Technological Impact:

Philosophical Impact:

Critiques & Ongoing Debates:

Current Status:

5. Billeh et al., 2020 - Allen Institute V1 Biophysical Network Model

Journal: bioRxiv → Cell Reports (2020)
Authors: Yazan N. Billeh et al., Allen Institute
Title: Systematic Integration of Structural and Functional Data into Multi-scale Models of Mouse Primary Visual Cortex

This work bridges Blue Brain's statistical reconstruction approach with Allen's rich experimental datasets, creating a data-constrained V1 model with real anatomical connectivity.

Motivation & Approach

Combining Best of Both Worlds:

Scale:

Connectome Data Integration

Unlike Blue Brain's pure statistical approach, this model uses:

  1. Cell Type Atlas (Allen Cell Types Database):

  2. Connectivity Measurements:

  3. Functional Data:

Model Architecture

Neuron Models (Two Levels):

1. Biophysically Detailed (GLIF5 + detailed models):

2. Point Neurons (LIF and GLIF):

Hybrid Strategy Rationale:

Connectivity Construction

Layer-by-Layer Connection Rules:

For each pre-synaptic neuron type → post-synaptic neuron type:

P(connection)=f(distance,layer,typepre,typepost)P(\text{connection}) = f(\text{distance}, \text{layer}, \text{type}_{\text{pre}}, \text{type}_{\text{post}})

Data-Driven Parameters:

Long-Range Connections:

Parameter Optimization

Challenge: 100+ free parameters even with cell-type-level constraints

Multi-Stage Optimization:

Stage 1: Single-Cell Parameters

Stage 2: Synaptic Weights

Optimization Method:

Validation: Comparison to In Vivo Data

Spontaneous Activity:

Evoked Responses (Visual Stimuli):

Network Dynamics:

Key Findings

1. Inhibition is Critical:

2. Recurrent Amplification:

3. Predictions Tested:

4. Layer-Specific Computations:

Comparison to Other Approaches

Feature Billeh 2020 (Allen) Blue Brain 2015 Turaga 2024 (Fly)
Connectome Type Mixed (real + inferred) Statistical Full EM
Scale 230K neurons 31K neurons 60K neurons
Neuron Model Hybrid (GLIF + detailed) Multi-compartmental HH Mechanistic point
Validation Functional data (imaging) Electrophysiology Single-neuron calcium imaging
Strength Large scale + real connectivity Biophysical detail Predictive accuracy
Main Use Circuit perturbations Emergent properties Stimulus-response mapping

Significance

Methodological:

Scientific:

Open Science:

6. Potjans & Diesmann, 2014 - Canonical Cortical Microcircuit Model

Journal: Cerebral Cortex (2014)
Authors: Tobias C. Potjans, Markus Diesmann
Title: The Cell-Type Specific Cortical Microcircuit: Relating Structure and Activity in a Full-Scale Spiking Network Model

Although based on statistical connectivity (not EM), this model is foundational and widely used. It deserves mention because it's been the standard reference for cortical modeling.

The "Canonical" Cortical Circuit

Motivation:

Based On:

Model Structure

Scale:

Neuron Model:

Connectivity:

Connection probability matrix PijP_{ij} (from population jj to population ii):

[L2/3EL2/3IL4EL4IL5EL5IL6EL6IL2/3E0.100.170.030.050.02...L2/3I0.140.24...............]\begin{bmatrix} & L2/3_E & L2/3_I & L4_E & L4_I & L5_E & L5_I & L6_E & L6_I \\ L2/3_E & 0.10 & 0.17 & 0.03 & 0.05 & 0.02 & ... \\ L2/3_I & 0.14 & 0.24 & ... \\ ... & ... & ... & ... \\ \end{bmatrix}

Key Features:

Why This Model is Important

1. Simplicity Meets Biology:

2. Spontaneous Activity:

3. Testable Predictions:

4. Benchmark Model:

Limitations (Acknowledged):

But → It's a starting point, not the final word

🧬 Theory & Principles: What Have We Learned?

This section synthesizes theoretical insights from connectome-based modeling across all organisms.

1. Beiran & Litwin-Kumar, 2024 - Theoretical Limits of Connectome-Constrained Prediction

Journal: bioRxiv (2024)
Authors: Manuel Beiran, Ashok Litwin-Kumar (Columbia University)
Title: Prediction of neural activity in connectome-constrained recurrent networks

The Central Question

Is the connectome sufficient to predict neural dynamics?

Even with perfect knowledge of:

Can we predict neural activity? Or is there irreducible uncertainty?

Theoretical Framework

Connectome-Constrained RNN:

τdxidt=−xi+∑jwijϕ(xj)+Iiext\tau \frac{dx_i}{dt} = -x_i + \sum_j w_{ij} \phi(x_j) + I_i^{\text{ext}}

Where:

Key Unknown: gg (varies across synapses even of same type)

Main Results

1. Degeneracy Problem:

2. Lower Bound on Uncertainty:

For a network with NN neurons and SS synapses:

σprediction2≥f(N,S,σg2)\sigma_{\text{prediction}}^2 \geq f(N, S, \sigma_g^2)

Where σg2\sigma_g^2 is variance in synaptic weights

Implication: Even with connectome, prediction error has a floor

3. What Helps Reduce Uncertainty:

Insights for Connectome Projects

Connectome is Necessary but Not Sufficient:

Practical Recommendations:

  1. Prioritize sparse synaptic weight measurements over complete connectivity
  2. Measure temporal parameters (time constants, delays)
  3. Use functional data to constrain the unknown parameters
  4. Focus on hub neurons and recurrent motifs (highest impact on dynamics)

2. General Principles from Cross-Species Comparisons

What Works Across All Organisms?

Principle 1: Connectome Constrains Dynamics (Partially)

Quantified:

Implication:

Principle 2: Cell-Type-Level Parameterization is Powerful

Evidence:

Why It Works:

Practical Benefit:

Principle 3: Emergent Computation from Structure + Local Nonlinearity

Examples:

Fly Motion Detection (Turaga):

C. elegans Locomotion (Zhao):

Rat Cortex Oscillations (Blue Brain):

General Rule:

Principle 4: Recurrent Amplification is Ubiquitous

Observed In:

Mechanism:

Output=Input1−Recurrent Gain\text{Output} = \frac{\text{Input}}{1 - \text{Recurrent Gain}}

Implications:

Principle 5: Inhibitory Diversity is Functionally Critical

Evidence:

Mouse V1 (Billeh):

Fly optic lobe (Borst):

C. elegans (Zhao):

General Pattern:

3. The "Connectome Ladder": Levels of Abstraction

Different modeling goals require different levels of detail:

Level Connectome Info Neuron Model Example Use Case
L1: Binary Who connects to whom Point neuron (LIF) Potjans 2014 Network structure analysis
L2: Weighted + Synapse counts Point neuron + types Creamer C. elegans Dynamics prediction (coarse)
L3: Biophysical + Spatial locations Compartmental HH Blue Brain Emergent properties
L4: Functional + In vivo measurements Mechanistic + data fit Turaga Fly Single-neuron prediction

Choosing the Right Level:

Trend: Moving up the ladder as data and compute improve

🔑 Cross-Species Insights

Common Principles Across Organisms

Aspect Drosophila C. elegans Mouse
Connectome Completeness ✅ Full (FlyWire) ✅ Full ⚠️ Partial (1 mm³)
Neuron Count ~140,000 ~302 ~75 million (whole brain)
Neuron Model Complexity Mechanistic point → Compartmental Compartmental (simple) Point (practical limit)
Synapse Type Chemical (mostly) Chemical + Gap junctions Chemical (gap junctions less characterized)
Functional Data Abundant (calcium imaging) Abundant (whole-brain imaging) Large-scale (MICrONS)
Connectome Predictive Power High (~70-90%) Moderate (~40-50%) Moderate (~50-60%)
Key Challenge Dendritic computations Graded transmission, embodiment Incomplete connectome, scale

Why Connectome Alone is Insufficient

All three organisms show the same pattern:

  1. Connectome provides strong scaffold (30-50% variance explained)

  2. Need physiological parameters:

  3. Dynamics matter:

The Optimal Modeling Strategy (Synthesis)

Based on all reviewed papers:

Connectome Structure + Cell-Type Parameters + Functional Data → Accurate Model

Step-by-Step Recipe:

  1. Start with connectome:

  2. Add cell-type biophysics:

  3. Parameterize at cell-type level (not individual synapses):

  4. Optimize using functional data:

  5. Validate with held-out data:

Future Directions

Technical:

Biological:

Applications:

Summary Table: Key Papers at a Glance

Drosophila Models

Paper Scale Neuron Model Connectome Use Free Parameters Key Innovation
Turaga 2024 ⭐️ 60K neurons Mechanistic point Full FlyWire EM ~10K-100K Single-neuron prediction (r=0.7-0.9)
Fiete 2025 439 neurons (HD circuit) Point Full FlyWire EM 57 (!) Massive parameter reduction via cell types
Borst 2024 325 neurons (5 columns) Conductance-based Optic lobe connectivity ~1K Temporal filtering cascade
Whole-brain LIF 2024 140K neurons LIF Full FlyWire EM ~1K First whole-brain sensorimotor model

C. elegans Models

Paper Scale Neuron Model Connectome Use Free Parameters Key Innovation
Zhao 2024 ⭐️ 302 neurons + 95 muscles Compartmental HH Full + gap junctions ~10K-100K Brain-body-environment closed loop
Morrison 2025 ~30 neurons Conductance-based Premotor circuit ~100 Data-driven fit to calcium imaging
Creamer 2024 302 neurons Linear dynamical system Full connectome 302² Quantifies connectome insufficiency

Mammalian Cortex Models

Paper Organism Scale Neuron Model Connectome Use Free Parameters Key Innovation
Blue Brain 2015 ⭐️ Rat 31K neurons, 37M synapses Multi-compartmental HH Statistical (touch-detection) Millions First mammalian cortical simulation
Billeh 2020 ⭐️ Mouse 230K neurons, 280M synapses Hybrid (GLIF + detailed) Mixed (real + inferred) ~100K Large-scale V1 with cell type diversity
Potjans 2014 Generic 80K neurons LIF Statistical (8 populations) ~100 Canonical microcircuit benchmark
MICrONS Mouse 100K neurons Point (in models) Partial EM (1 mm³) ~10K Structure-function co-registration
Rajan 2020 Mouse Multi-region RNN Inferred from function 1000s CURBD: Effective connectivity

Theory & Principles

Paper Focus Key Contribution
Beiran & Litwin-Kumar 2024 Theoretical limits Connectome alone has prediction floor; need weights + function
Cross-species synthesis General principles 5 universal principles (see section)
Connectome Ladder Abstraction levels 4-level framework for choosing model complexity

🎯 Synthesis: The Current State of Connectome-Based Modeling

Where We Are (2025)

Complete Connectomes Available:

Modeling Maturity:

Organism Connectome Neuron Models Functional Data Predictive Models Behavioral Validation
Fly ⭐️⭐️⭐️ Complete ⭐️⭐️⭐️ Excellent ⭐️⭐️⭐️ Abundant ⭐️⭐️⭐️ High accuracy ⭐️⭐️ Good
C. elegans ⭐️⭐️⭐️ Complete ⭐️⭐️ Good ⭐️⭐️⭐️ Whole-brain imaging ⭐️⭐️ Moderate ⭐️⭐️⭐️ Excellent
Mouse ⭐️ Partial ⭐️⭐️⭐️ Excellent (Allen) ⭐️⭐️⭐️ Large-scale ⭐️⭐️ Moderate ⭐️ Limited
Rat ❌ None ⭐️⭐️⭐️ Excellent (BBP) ⭐️⭐️ Good ⭐️ Qualitative ⭐️ Limited

What We've Learned: The "Connectome Equation"

The field has converged on a consensus formula for predicting neural activity:

Neural Activity=f(Connectome⏟Structure+Cell Types⏟Parameters+Biophysics⏟Dynamics+Input⏟Context)\boxed{\text{Neural Activity} = f(\underbrace{\text{Connectome}}_{\text{Structure}} + \underbrace{\text{Cell Types}}_{\text{Parameters}} + \underbrace{\text{Biophysics}}_{\text{Dynamics}} + \underbrace{\text{Input}}_{\text{Context}})}

Component Contributions (approximate variance explained):

  1. Connectome alone: 30-50%

  2. + Cell-type parameters: +20-30%

  3. + Functional data: +10-20%

  4. Remaining (~10-20%):

Key Insight: Each component is necessary; none is sufficient alone.

The Connectome Taxonomy: What Type of Model Do You Need?

Research Goal
    ├── Understand network structure
    │   └── → Binary connectome + graph theory (L1)
    │
    ├── Predict coarse dynamics
    │   └── → Weighted connectome + LIF neurons (L2)
    │
    ├── Study emergent properties
    │   └── → Statistical connectome + HH neurons (L3, Blue Brain)
    │
    ├── Predict single-neuron responses
    │   └── → Full EM connectome + mechanistic models (L4, Turaga)
    │
    └── Design perturbation experiments
        └── → Hybrid models + functional data (L3.5, Billeh)

No single "best" approach — depends on question, data, and resources.

Outstanding Questions & Challenges

1. The Weight Problem:

2. The Completeness Problem (Mammals):

3. The Dynamics Problem:

4. The Validation Problem:

5. The Interpretation Problem:

1. Hybrid Models:

2. Multi-Modal Integration:

3. Whole-Organism Modeling:

4. GPU-Accelerated Simulation:

5. Foundation Models Meet Connectomes:

Practical Recommendations for Researchers

If you want to build a connectome-based model:

Step 1: Define Your Question

Step 2: Inventory Your Data

Step 3: Choose Model Complexity

Step 4: Optimize Intelligently

Step 5: Validate Rigorously

Step 6: Iterate

The Big Picture: Why This Matters

Scientific Impact:

Technological Impact:

Medical Impact (Future):

Future Vision (2025-2035)

Near-term (2-5 years):

Medium-term (5-10 years):

Long-term (10-20 years):

The Goal: Not to replace experimental neuroscience, but to complement it with computational models that:

📚 Additional Resources

Software & Tools:

Databases:

Key Labs & Projects:

TODO

Hebbian Learning based Orthogonal Projection for Continual Learning of Spiking Neural Networks https://openreview.net/forum?id=MeB86edZ1P

https://github.com/pkuxmq/HLOP-SNN https://arxiv.org/abs/2402.11984

Learning using switching synaptic plasticity rules