Coming up short: generative network models fail to accurately capture long-range connectivity

1530 

Abstract Submission 

Stuart Oldham¹, Gareth Ball¹

¹Murdoch Children's Research Institute, Melbourne, VIC

Stuart Oldham  
Murdoch Children's Research Institute
Melbourne, VIC

Gareth Ball  
Murdoch Children's Research Institute
Melbourne, VIC

Generative network models (GNMs) are used to create synthetic networks which aim to capture key properties of human brain connectomes. GNMs are parameterised using simple wiring-rules, commonly a trade-off between minimising connection cost and promoting advantageous features, such as topological value or the connection of biologically similar regions[1-3]. Differences in GNM parameters and wiring-rules may give insight into the mechanisms that shape brain connectivity. Indeed, GNM wiring constraints vary among individuals and are linked to cognition and psychopathology[4-6]. Drawing biological insight from GNMs rests on their ability to accurately model human brain networks. While GNMs capture connectome topology–including long-tailed degree distributions, clustering, and small-world properties[1-6]–they are less able to reproduce connectome topography (spatial embedding of network properties)[3]. Such embedding is critical for enabling brain dynamics, thus is a key property GNMs should capture[7,8]. Therefore, we aimed to evaluate the ability of GNMs to reproduce the topography of human brain networks.

Diffusion tractography from 326 participants of the Human Connectome Project was used to construct a consensus 200 cortical region empirical network in the left hemisphere[9]. The GNM was formulated as a trade-off between connection cost and regional similarity of cortical metrics (Fig. 1A). Free parameters were fit using an optimisation procedure that maximised the topological similarity of the model networks to the empirical data. Networks produced using the best fit parameters were evaluated based on topological similarity, overlapping connections, and degree sequence similarity with the empirical data (Fig. 1B). Nine different GNMs, each using different cortical features[9], were evaluated (Fig. 1C).

Overall, GNMs showed high topological similarity to empirical networks (maximum dissimilarity:0.16–0.35) but reproduced only 36–44% of empirical connections and showed weak correlations with the empirical degree sequence (r=-0.48–0.18), highlighting the mismatch between synthetic and empirical topographies (Fig. 1C-1D). To isolate the underlying cause of this mismatch, we examined the overlap between connections of different lengths. GNMs were able to capture 51–84% of empirical short-range connections (<30mm), but this accuracy dropped to 2–20% for mid-range (30–90mm) and 0–10% for long-range (>90mm) connections (Fig. 2A). Empirical degree topography is mainly determined by connections >30mm (Fig. 2B). The discrepancies in spatial embedding therefore arise from GNMs' inability to accurately capture these connections. Importantly, no cortical feature was able to mitigate this error (Fig. 2C). Indeed, using a random similarity matrix achieves similar model performance to neurophysiological metrics, suggesting that the formation of long connections in GNMs is largely invariant to model input.

We show that GNMs, while matching brain network topological features, fail to capture topography due to inaccurate reconstruction of mid/long-range connections. GNMs can induce topologically-similar empirical networks by mapping most short-range connections and promoting a few long-range connections, even at random. Specific long-range connections enable topographic properties that are essential to brain dynamics/function[7,8]. Drawing conclusions based on GNMs' ability to reproduce topology–but not topography–is likely to lead to an incomplete or inaccurate understanding of the mechanisms and constraints which shape brain networks. To improve GNMs, mechanisms which are preferentially biased towards long-range connectivity should be explored. Factors like connection timing may be relevant, as evidence suggests that long-range connections between hubs form earlier than other connection types[10]. Investigating these ways to improve GNMs is required so they can be used to explore the underpinning developmental mechanisms of the connectome.

Connectivity (eg. functional, effective, structural) ¹

Cortical Anatomy and Brain Mapping ²

Computational Neuroscience

Cortex

Modeling

Other - Connectome; Network; Connectivity; Generative network model; Brain maps