Biology subjects

Sotero, R.

Publications and source records attributed to Sotero, R..

6 recordsLinked to original sources

bioRxiv2026

Metabolic Cost Asymmetry as a Structural Regulariser for Recurrent Spiking Neural Networks

Cortical energy expenditure is heavily dominated by synaptic transmission, with excitatory (glutamatergic) signalling demanding substantially more ATP than inhibitory (GABAergic) signalling. Yet existing trainable spiking neural networks treat energy as a generic, synapse-type-agnostic penalty. Here, we embed this fundamental glutamate/GABA cost asymmetry as a Lagrangian penalty within a surrogate-gradient learning framework, and use it to train a Dale's-law-compliant recurrent microcircuit on the Spiking Heidelberg Digits (SHD) benchmark. Because glutamatergic synapses are metabolically more expensive, a single scalar penalty applied uniformly to all weights automatically exerts stronger gradient pressure on excitatory pathways. We demonstrate that this differential pressure alone is sufficient to produce three major effects. First, it moves the network out of the suboptimal, inhibition-dominated regime of the unpenalised network, simultaneously improving accuracy and reducing total energy several-fold. Second, it induces an emergent pruning hierarchy in which excitatory connections are eliminated first, while inhibitory connections exhibit graded resistance. Third, it actively expands the dimensionality of the population code at moderate penalty strengths. These structural motifs remain robust across a two-dimensional sweep of five cost ratios and fifteen structural densities (75 distinct conditions). Ultimately, our results demonstrate that a single biophysical cost asymmetry, operating without auxiliary regularisers or architectural priors, is sufficient to organise sparse, pathway-specific circuit structure in recurrent spiking networks.

neuroscience↗

bioRxiv2024

Estimating the Excitatory-Inhibitory Balance from Electrocorticography Data using Physics-Informed Neural Networks

Understanding the excitatory/inhibitory (E/I) balance in the brain is crucial for elucidating the neural mechanisms underlying various cognitive functions and states of consciousness. Mathematical models have provided significant insights into these mechanisms, but they often face challenges due to high dimensionality, noisy observation signals, and nonlinearities. In this paper, we introduce a novel methodology using Physics-Informed Neural Networks (PINNs) to estimate the E/I balance from electrocorticography (ECoG) data, effectively addressing these limitations. By integrating physical laws via a neural mass model with neural network training, our approach enhances parameter estimation accuracy and robustness. Our analysis reveals a significant reduction in long-range connections (LRCs) and excitatory short-range connections (SRCs) under anesthesia, alongside an increase in inhibitory SRCs, highlighting anesthesias role in modulating neural dynamics to induce unconsciousness. These findings not only corroborate existing theories on the neural mechanisms of anesthesia but also provide new insights into brain connectivity and its relationship with consciousness. CCS CONCEPTSComputing methodologies[->]Machine learning ACM Reference Format

neuroscience↗

bioRxiv2024

Directed fMRI-based Functional Connectivity Estimation using Physics-Informed Neural Networks

Estimating directed functional connectivity (dFC) within the brain is crucial for comprehending neural interactions. However, conventional methodologies encounter constraints in accuracy, scalability, and interpretation. The method presented here harnesses Physics-Informed Neural Networks (PINNs) to amalgamate the governing physical principles of brain dynamics, thereby improving dFC estimation from resting-state functional magnetic resonance imaging (rsfMRI) data. In particular, during the training phase, we derive the input weights from a long-short term memory (LSTM) network, which, within our framework, represent the influence of all other brain areas on the specific region under consideration. These input weights are then integrated into the nonlinear differential equation that models the rsfMRI time series within the specific brain area. Through the training of the PINN model, we simultaneously estimate, for each brain area, the biophysical parameters of the model, including the dFC parameters from all the remaining areas. We applied this methodology to both autism spectrum disorder (ASD) and neurotypical data, revealing significant sex-specific differences in connectivity patterns. These findings underscore the potential of PINNs in advancing our understanding of neural dynamics and emphasize the significance of directionality in brain connectivity research.

neuroscience↗

bioRxiv2023

Exploring Correlation-Based Brain Networks with Adaptive Signed Random Walks

The human brain is a highly connected network with complex patterns of correlated and anticorrelated activity. Analyzing functional connectivity matrices derived from neuroimaging data can provide insights into the organization of brain networks and their association with cognitive processes or disorders. Common approaches, such as thresholding or binarization, often disregard negative connections, which may result in the loss of critical information. This study introduces an adaptive signed random walk (ASRW) model for analyzing correlation- based brain networks that incorporates both positive and negative connections. The model calculates transition probabilities between brain regions as a function of their activities and connection strengths, dynamically updating probabilities based on the differences in node activity and connection strengths at each time step. Results show that the classical random walk approach, which only considers the absolute value of connections, underestimates the mean first passage time (MFPT) compared to the proposed ASRW model. Our model captures a wide range of interactions and dynamics within the network, providing a more comprehensive understanding of its structure and function. This study suggests that considering both positive and negative connections, has the potential to offer valuable insights into the interregional coordination underlying various cognitive processes and behaviors.

neuroscience↗

bioRxiv2019

The Importance of Anti-Correlations in Graph Theory Based Classification of Autism Spectrum Disorder

In recent years, there has been a significant growth in the number of applications of machine learning (ML) techniques to the study and identification of neurological disorders. These methods rely heavily on what features are made available to the ML algorithm. Features such as graph theoretical metrics of resting-state fMRI-based brain networks have proven useful. However, the computation of functional brain networks relies on making an arbitrary choice about whether the obtained anti-correlations, representing the strengths of functional connections in the brain, should be discarded or not. In this study, we examine how this choice affects the performance of a support vector machine (SVM) model for classifying autism spectrum disorder. We extracted graph theoretical features using three different pipelines for constructing the functional network graph. These pipelines primarily used positive weights, negative weights (anti-correlations) and only the absolute value of weights of the correlation matrix derived from fMRI time-series. Our results suggest that in the presence of Global Signal Regression (GSR) the features extracted from anti-correlations play a major role in improving model performance. However, this does not undermine the importance of features from other pipelines.

neuroscience↗

bioRxiv2019

Spatiotemporal empirical mode decomposition of resting-state fMRI signals: application to global signal regression

Resting-state functional connectivity MRI (rs-fcMRI) is a common method for mapping functional brain networks. However, estimation of these networks is affected by the presence of a common global systemic noise, or global signal (GS). Previous studies have shown that the common preprocessing steps of removing the GS may create spurious correlations between brain regions. In this paper, we decompose fMRI signals into 5 spatial and 3 temporal intrinsic mode functions (SIMF and TIMF, respectively) by means of the empirical mode decomposition (EMD), which is an adaptive data-driven method widely used to analyze nonlinear and nonstationary phenomena. For each SIMF, brain connectivity matrices were computed by means of the Pearson correlation between TIMFs of different brain areas. Thus, instead of a single connectivity matrix, we obtained 5 x 3 = 15 functional connectivity matrices. Given the high value obtained for large-scale topological measures such as transitivity, in the low spatial maps (SIMF3, SIMF4, and SIMF5), our results suggest that these maps can be considered as spatial global signal masks. Thus, the spatiotemporal EMD of fMRI signals automatically regressed out the GS, although, interestingly, the removed noisy component was voxel-specific. We compared the performance of our method with the conventional GS regression and to the results when the GS was not removed. While the correlation pattern identified by the other methods suffers from a low level of precision, our approach demonstrated a high level of accuracy in extracting the correct correlation between different brain regions.

neuroscience↗