Biologically-Informed Neural Networks

• Biologically-Informed Neural Networks (BINNs) are neural models that embed explicit biological constraints, such as synaptic plasticity and connectivity, into their architecture.
• They employ local learning rules, Hebbian updates, and mechanistic modeling to enhance robustness and interpretability in tasks like continual learning and biomedical imaging.
• Applications span genomics, cancer stratification, and neuromorphic AI, demonstrating improved interpretability, performance, and resistance to catastrophic forgetting.

Biologically-Informed Neural Networks (BINNs) are a class of neural architectures, learning rules, and modeling methodologies that integrate explicit biological principles and empirical knowledge from neuroscience, physiology, or systems biology into the design and training of artificial neural networks. BINNs span a broad methodological range, from synaptic-level learning rules inspired by metaplasticity and spike-timing-dependent plasticity, to structural constraints derived from connectomics and pathway knowledge, to mechanistic modeling of biological dynamics using neural surrogates embedded in physical or biological law. The key unifying feature of BINNs is the direct incorporation of biological constraints or structures: connectivity, plasticity, noise, cell type ratios, or operator equations, into the functional or architectural specification of a neural network.

1. Biological Motivations and Foundational Principles

Leading work on BINNs draws inspiration from diverse biological mechanisms, including but not limited to:

• Synaptic metaplasticity and synaptic consolidation, where plasticity is modulated by prior history, preventing catastrophic forgetting via a mechanism analogous to the activity-dependent adjustment of plasticity thresholds seen in the brain (Laborieux et al., 2021).
• Structural and functional principles from connectomics and neuroanatomy, for example, imposing adjacency masks, laminar stratification, and graded inhibition as observed in mammalian cortex (Prasanth et al., 20 Jan 2026).
• Local learning rules rooted in Hebbian plasticity, spike-timing-dependent plasticity (STDP), and homeostatic regulation, paralleling mechanisms in biological neurons (Liu et al., 2019, Inoue et al., 3 Mar 2026).
• Definition of neural ordinary/partial differential equation surrogates constrained by governing laws of biological systems (e.g. reaction-diffusion equations), coupling interpretability with powerful function approximation (Lagergren et al., 2020, Lavery et al., 20 Apr 2026, Nardini, 2023, Crossley et al., 11 Jun 2026).
• Institution of biologically-realistic cell type ratios (e.g. an 80:20 excitatory:inhibitory neuron balance) and adherence to Dale’s law, supporting both functional stability and energy efficiency (Kilgore et al., 2024, Inoue et al., 3 Mar 2026).

Biological inspiration in BINNs typically manifests through the imposition of locality, architectural sparsity, plasticity constraints, or mechanistic priors, each aimed at emulating properties observed in neurobiological circuits or physiological systems.

2. Architectural and Learning Mechanisms

BINNs leverage a spectrum of mechanisms to encode biological knowledge into neural network training and structure:

Synaptic and Plasticity Mechanisms

• Metaplastic Binarized Neural Networks: In BNNs for continual learning, hidden real-valued weights are updated by a metaplasticity-informed rule: large-magnitude weights are made resistant to flipping, mirroring synaptic consolidation. The update incorporates a modulating factor based on the derivative of $\tanh$ to achieve exponential attenuation of plasticity as a function of synaptic strength (Laborieux et al., 2021).
• Hebbian/Oja Updates and STDP: Networks can eschew global error signals, instead relying on local correlation-based updates (Hebbian rules), competition (Oja’s rule), and, in the spiking domain, spike-timing dependent plasticity consistent with biological STDP curves (Liu et al., 2019, Inoue et al., 3 Mar 2026).
• Homeostatic Regulation: Some frameworks explicitly adjust biases or scale weights to enforce target firing rates or activity statistics, thereby stabilizing network dynamics (Prasanth et al., 20 Jan 2026, Inoue et al., 3 Mar 2026).

Connectivity and Structural Constraints

• Pathway-encoded Sparsity: Models may employ binary masks encoding gene-pathway relationships, connectome-derived adjacency, or pathway-level priors, imposing biologically plausible sparsity on the connection patterns between units (Pedersen et al., 2023, Kontolati et al., 16 Oct 2025, Prasanth et al., 20 Jan 2026).
• Connectome-inspired Lamination and Inhibition: Networks have been constructed with laminar hierarchical structure, graded inhibition, and intra-/inter-layer connections that strictly follow observed anatomical adjacency matrices (Prasanth et al., 20 Jan 2026).

Cell Type Ratios and Inhibitory Control

• Excitatory/Inhibitory Balance: BinNs model the true 80:20 E:I neuron ratio by enforcing nonnegativity and nonpositivity constraints on respective weight populations, with functional implications for robustness, low activity stability, and hardware compatibility (Kilgore et al., 2024, Inoue et al., 3 Mar 2026).

Mechanistic Modeling

• Neural-augmented Mechanistic ODE/PDE Surrogates: BINNs generalize PINNs by using neural networks to represent unknown constitutive laws (e.g., density-dependent diffusivities, growth or reaction rates) within a mechanistic operator framework, enforcing data fidelity together with PDE or ODE residual constraints via soft penalties (Lagergren et al., 2020, Lavery et al., 20 Apr 2026, Nardini, 2023, Crossley et al., 11 Jun 2026).
• Explicit Likelihood-based Loss Functions: Recent extensions simultaneously learn both biological dynamics and noise models (e.g., heteroscedastic variance), formulating the loss as a penalized negative log-likelihood, leveraging statistical structure in biological variability (Crossley et al., 11 Jun 2026).

3. Applications Across Domains

The BINN framework has been successfully applied to an array of domains:

• Continual Learning: Synaptic metaplasticity rules demonstrably improve retention across sequential learning scenarios on benchmarks such as permuted MNIST/Fashion-MNIST and binarized VGG-16 CIFAR-10 (Laborieux et al., 2021).
• Biomedical Imaging: Hierarchically constrained, multi-stage 3D U-Net trees, each encoding tissue compartment relationships (e.g., brain tumor regions and mutual-exclusion constraints), achieve significantly improved Dice scores and specificity in clinical segmentation tasks (Beers et al., 2017).
• Genomic Prediction and Multi-omics Modeling: Pathway-masked BINNs integrating genotype, transcriptomic, and metabolomic priors deliver up to 56% improvements in rank-correlation accuracy and increased interpretability in crop genomics and synthetic benchmarks, outperforming standard deep and linear models especially in data-scarce scenarios (Kontolati et al., 16 Oct 2025).
• Cancer Stratification: Graph-based and pathway-masked BINNs outperform and differ systematically from uninformed architectures for clinical risk prediction, with pathway-based sparsification critical to performance and model interpretability (Pedersen et al., 2023).
• Spiking Neural Networks and Neuromorphic AI: BINNs built with E/I ratio constraints and local learning rules are robust to synaptic noise and support efficient event-driven neuromorphic computation (Kilgore et al., 2024, Liu et al., 2019).
• Mechanistic Model Discovery: BINNs constrained by reaction–diffusion or growth dynamics accurately reconstruct governing equations and discover interpretable analytic forms from sparse, noisy experimental data or agent-based model outputs (Lagergren et al., 2020, Lavery et al., 20 Apr 2026, Nardini, 2023, Crossley et al., 11 Jun 2026).

4. Quantitative Performance and Empirical Insights

BINNs routinely demonstrate significant empirical benefits over both unconstrained neural baselines and classical regression approaches:

• Continual learning: Metaplastic BNNs preserve accuracy within 1–2% of EWC-BNN baselines, while naïve BNNs drop to <50% after multiple tasks (Laborieux et al., 2021).
• Sparse and generalizable representations: BINNs employing Hebbian/local updates and Dale’s law demonstrate near-lognormal synaptic weight distributions, achieve ≈90% sparsity, and exhibit strong adversarial robustness and few-shot generalization surpassing backpropagation and Hopfield-style models (Inoue et al., 3 Mar 2026).
• Biomedical imaging: Sequential U-Net models that encode biological hierarchy improve segmentation accuracy by up to 15% in tumor sub-regions and produce fewer spurious islands (Beers et al., 2017).
• Genomics and multi-omics: BINNs incorporating pathway and omics masks reduce prediction error by up to 75% in synthetic metabolomics and identify functionally relevant latent features with high correlation to experimental measurements (Kontolati et al., 16 Oct 2025).
• Mechanistic inference: BINN-guided PDEs and likelihood-based BINNs yield robust, accurate models of collective migration and population growth, with improved mechanistic accuracy and uncertainty calibration compared to standard mean-field and neural surrogates (Nardini, 2023, Crossley et al., 11 Jun 2026).

5. Interpretability, Robustness, and Biological Plausibility

• Latent Variable Interpretability: BINN designs allow mapping of intermediate latent variables to omics entities, pathways, or anatomical regions, supporting direct interpretation and enabling post-hoc sensitivity analyses and biomarker discovery (Kontolati et al., 16 Oct 2025, Pedersen et al., 2023).
• Architectural Sensitivity and Error Modes: Separate architectural choices (e.g., pathway-masked feedforward nets vs. graph neural networks) yield complementary and sometimes divergent clinical stratifications, underscoring the importance of architectural audit and ensemble approaches in translational domains (Pedersen et al., 2023).
• Robustness and Regularization: The explicit inclusion of adaptation, homeostasis, or metaplasticity mitigates catastrophic forgetting and overfitting, aligns learning dynamics with biological timescales, and improves generalization under noise and adversarial perturbation (Laborieux et al., 2021, Inoue et al., 3 Mar 2026, Kubo et al., 2022).

6. Methodological Advances and Outlook

Recent developments in the BINN field emphasize:

• Hybrid and Self-organizing Architectures: Brain-like Neural Networks (e.g., LuminaNet) instantiate dynamic structural changes—neurogenesis, synaptic competition, pruning—driven by local variance statistics, achieving state-of-the-art results on image and text benchmarks without human-designed macrostructure (Liu, 27 Jan 2026).
• Integration with Neuromorphic Hardware: Structural and functional constraints, such as E/I balance, local learning, and sparseness, are directly compatible with the requirements and constraints of neuromorphic and memristive hardware design (Kilgore et al., 2024).
• Scalability via Connectomics: Direct instantiation of large-scale connectome-derived mask patterns is computationally feasible and positions BINNs as a principled route to integrating next-generation connectomics data into large-scale AI systems (Prasanth et al., 20 Jan 2026).
• Mechanistic Discovery from Data: BINNs provide a practical framework for merging black-box function fitting with interpretable mechanistic discovery, enabling rapid exploration and hypothesis formation in biological modeling (Lagergren et al., 2020, Lavery et al., 20 Apr 2026, Nardini, 2023, Crossley et al., 11 Jun 2026).

7. Challenges, Open Questions, and Future Directions

Key challenges for BINN research include:

• Generalization of Bio-principles: Systematically identifying which biological principles yield consistent gains across domains, and under which conditions explicit biological constraints may conflict with data-driven inductive biases.
• Architectural Sensitivity and Clinical Translation: Deconvolving the effects of different BINN architectures on safety-critical predictions remains a frontier, with ensemble and cross-model consistency checks proposed as potential remedies (Pedersen et al., 2023).
• Uncertainty Quantification and Bayesian Extensions: Extensions to Bayesian BINNs, enabling robust quantification of epistemic uncertainty in latent laws and parameter estimates, are an active research area (Lagergren et al., 2020, Crossley et al., 11 Jun 2026).
• Scalable Hebbian/Local Learning: Methods to scale strictly local rules to high-dimensional, deep convolutional networks or transformers, ideally with minimal weight transport and maximum parallelizability, remain only partially addressed (Inoue et al., 3 Mar 2026, Liu, 27 Jan 2026).
• Dynamic Adaptation and Structural Plasticity: Further integration of dynamic rewiring, homeostatic plasticity, and adaptation principles may unlock higher degrees of biological realism and functional robustness (Liu, 27 Jan 2026, Kubo et al., 2022).

The field is positioned at the intersection of neuroscience, machine learning, and complex systems modeling, with ongoing cross-disciplinary research likely to yield further advancements in both foundational understanding and practical AI deployment.