BioRNNs: biologically plausible recurrent neural network models of brain dynamics
Poster No:
1569
Submission Type:
Abstract Submission
Authors:
Ahmad Beyh1, Jason Kim2, David Zald1, Linden Parkes1
Institutions:
1Department of Psychiatry, Rutgers University, Piscataway, NJ, USA, 2Department of Physics, Cornell University, Ithaca, NY, USA
First Author:
Co-Author(s):
Introduction:
Modeling the brain's neural dynamics offers a unique opportunity to bridge multiple scales, from microscopic features like gene expression and cytoarchitecture to macroscopic ones like regional organization and network topology. By capturing the interactions across these levels, we can gain deeper insights into how structural and functional hierarchies integrate to shape behavior, cognition, and disease. Recurrent neural networks (RNNs) have emerged as powerful tools for multiscale modeling of neural dynamics that can quantitatively incorporate information about the brain's spatial structure (Mante et al., 2013; Tanner et al., 2023; Wang et al., 2021; Yang et al., 2019). RNNs consist of nodes that recurrently interact through time (the hidden layer), enabling the system to capture temporal dependencies in sequential inputs. Typical RNN architectures overlook the spatio-temporal constraints inherent to biological neural systems (Kim et al., 2024). For example, vanilla RNNs follow an all-to-all mapping where all nodes receive inputs and contribute directly and equally to decision-making outputs. This setup departs from known biology where, e.g., visual information enters the brain through the visual cortex and complex decision-making primarily involves association regions. A second key issue with vanilla RNNs is that the formation of the hidden layer connections is typically unconstrained. Together, these issues almost always lead RNNs to reach biologically unrealistic final architectures, limiting their value as multiscale models of the brain. Here, we overcome these limitations by introducing biophysically embedded RNNs (bioRNNs) whose hidden layer connectivity is shaped by two novel constraints: input/output (I/O) masking and biophysical embedding (Fig. 1). By integrating these two simple features into their design, bioRNNs emerge as more biologically plausible deep learning models of brain dynamics.
Methods:
We trained 100 RNNs on two behavioral tasks implemented in neurogym (github.com/neurogym/neurogym): Go-Nogo and Perceptual Decision Making. Each RNN consisted of 100 nodes representing cortical regions taken from the left hemisphere of the Schaefer 200-node brain atlas (Schaefer et al., 2018). For each RNN, we used (1) I/O masking to restrict task inputs to the 14 nodes designated as 'visual cortex' and outputs to the 27 nodes designated as 'association cortex'; and (2) spatial embedding, which leverages regions' multi-modal neurobiology to physically constrain the hidden layer's recurrent connections (Fig. 1). We then assessed the intrinsic neural timescales (INTs) of the trained networks by examining the decay in the autocorrelation of node-level time series, and we compared them to INTs obtained empirically from resting-state fMRI data from the Human Connectome Project (humanconnectome.org).
Results:
In the presence of I/O masking, which is a fundamental constraint of brain physiology, biophysically embedded RNNs outperformed those with simple physical embedding and those without any embedding (Fig. 2A). This was evident in the initial stages of learning where bioRNNs reached a high and stable performance level faster than other RNNs. We also observed a positive correlation between bioRNN INTs (trained on the Go-Nogo task) and empirical INTs obtained from fMRI (ρ = 0.46, p = 2e-06; Fig. 2B). This result indicates that our in silico models of the brain give rise to biologically-plausible brain dynamics.
Conclusions:
We demonstrated that introducing two fundamental brain-based constraints to RNNs endows them with a performance advantage and, critically, makes their neural dynamics more biologically plausible. Finally, our approach can flexibly incorporate diverse brain features for bioRNN embedding, including gene expression and cytoarchitecture maps. This flexibility makes our bioRNNs powerful test beds for addressing basic and clinical neuroscience questions of brain connectivity and dynamics.
Modeling and Analysis Methods:
fMRI Connectivity and Network Modeling 2
Methods Development 1
Task-Independent and Resting-State Analysis
Other Methods
Keywords:
Computational Neuroscience
Machine Learning
Modeling
Other - recurrent neural networks; neuroAI
1|2Indicates the priority used for review
Abstract Information
By submitting your proposal, you grant permission for the Organization for Human Brain Mapping (OHBM) to distribute your work in any format, including video, audio print and electronic text through OHBM OnDemand, social media channels, the OHBM website, or other electronic publications and media.
The Open Science Special Interest Group (OSSIG) is introducing a reproducibility challenge for OHBM 2025. This new initiative aims to enhance the reproducibility of scientific results and foster collaborations between labs. Teams will consist of a “source” party and a “reproducing” party, and will be evaluated on the success of their replication, the openness of the source work, and additional deliverables. Click here for more information. Propose your OHBM abstract(s) as source work for future OHBM meetings by selecting one of the following options:
Please indicate below if your study was a "resting state" or "task-activation” study.
Healthy subjects only or patients (note that patient studies may also involve healthy subjects):
Was this research conducted in the United States?
Are you Internal Review Board (IRB) certified? Please note: Failure to have IRB, if applicable will lead to automatic rejection of abstract.
Were any human subjects research approved by the relevant Institutional Review Board or ethics panel? NOTE: Any human subjects studies without IRB approval will be automatically rejected.
Were any animal research approved by the relevant IACUC or other animal research panel? NOTE: Any animal studies without IACUC approval will be automatically rejected.
Please indicate which methods were used in your research:
For human MRI, what field strength scanner do you use?
Provide references using APA citation style.
2. Mante, V. et al. (2013). Context-dependent computation by recurrent dynamics in prefrontal cortex. Nature, 503, 78–84. https://doi.org/10.1038/nature12742
3. Schaefer, A. et al. (2018). Local-global parcellation of the human cerebral cortex from intrinsic functional connectivity MRI. Cerebral Cortex, 28(9), 3095–3114. https://doi.org/10.1093/cercor/bhx179
4. Sydnor, V. et al. (2021). Neurodevelopment of the association cortices: patterns, mechanisms, and implications for psychopathology. Neuron 109(18), 2820–46. https://doi.org/10.1016/j.neuron.2021.06.016
5. Tanner, J. et al. (2023). Functional connectivity modules in recurrent neural networks: function, origin and dynamics. arXiv.
6. Wang, P.Y. et al. (2021). Evolving the olfactory system with machine learning. Neuron, 109(23), 3879-3892.e5. https://doi.org/10.1016/J.NEURON.2021.09.010
7. Yang, G.R. et al. (2019). Task representations in neural networks trained to perform many cognitive tasks. Nature Neuroscience, 22, 297–306. https://doi.org/10.1038/s41593-018-0310-2