BioNeuralNet: Bridging Biology and Computation

• BioNeuralNet refers to a family of biologically grounded neural systems that embed biological structure directly into computational substrates, using living neurons, GNNs, and biophysical constraints.
• It combines methodologies such as physical reservoir computing, multi-omics network analysis, and connectomics-driven models to leverage modularity, transient dynamics, and adaptive computation.
• Implementations showcase measurable performance gains in tasks like pattern recognition, disease prediction, and real-time control by harnessing biological organization.

BioNeuralNet denotes a family of biologically grounded neural systems in which biological structure, biophysics, biological data, or living neural tissue is treated as part of the computational substrate rather than as a loose source of analogy. In the available literature, the label is used for at least three major objects: living biological neuronal networks functioning as reservoirs in physical reservoir computing, graph-neural-network pipelines for multi-omics network analysis, and biohybrid or biophysically constrained neural architectures that incorporate cortical connectivity, spiking dynamics, memristive synapses, or closed-loop interaction with living tissue (Sumi et al., 2022). A common thread is the explicit coupling of representation learning or adaptive computation to biological organization, whether through cultured neurons, connectomics-derived masks, mechanistic state variables, or graph representations of biomolecular and neural systems (Ramos et al., 27 Jul 2025).

1. Terminological scope and principal usages

In the available literature represented here, “BioNeuralNet” is used in several distinct but related senses. The term can denote a living, modularly structured neuronal culture used as the nonlinear recurrent core of a reservoir computing system; a general-purpose Python framework for multi-omics network analysis built around Graph Neural Networks (GNNs); and broader classes of biologically inspired, biophysically grounded, or biohybrid neural systems (Sumi et al., 2022).

Usage	Realization	Representative source
Living reservoir	Micropatterned biological neuronal network used as a physical reservoir with linear readout	(Sumi et al., 2022)
Multi-omics software framework	End-to-end Python pipeline for network construction, GNN embeddings, and downstream analysis	(Ramos et al., 27 Jul 2025)
Biohybrid closed-loop system	Biological neural networks treated as computing substrates with deterministic stimulation and recording semantics	(Hogan et al., 12 Feb 2026)
Connectomics-constrained ANN	Feedforward network whose connectivity is masked by synaptic graphs from a cortical column	(Prasanth et al., 20 Jan 2026)
Biohybrid neural circuit	Rat neurons, VLSI neurons, and memristive synaptors linked over the Internet	(Serb et al., 2017)

This plurality suggests that BioNeuralNet is best understood as a technical umbrella for systems that operationalize biological organization directly in computation. A plausible implication is that the term marks a shift from metaphorical “brain inspiration” toward architectures in which biological constraints are primary design variables.

2. Living neuronal cultures as computational reservoirs

One established meaning of BioNeuralNet is a living neuronal culture used as a physical reservoir. In “Biological neurons act as generalization filters in reservoir computing,” the reservoir is a micropatterned biological neuronal network (mBNN) composed of primary rat cortical neurons cultured on glass coverslips patterned either by microcontact printing or PDMS microfluidic films, arranged into four square modules of $200\,\mu\text{m} \times 200\,\mu\text{m}$ connected by thin $5$–$10\,\mu\text{m}$ neurite pathways, stimulated optogenetically through ChrimsonR with patterned $623\,\text{nm}$ light and read out by Cal-520 AM calcium imaging at $20\,\text{Hz}$ (Sumi et al., 2022). In the reservoir-computing formalism used there, the biological network replaces the artificial recurrent core, the state is the calcium activity of $N=60$ neurons, and the only trainable part is a linear readout,

$$\mathbf{y}(t)=W\mathbf{x}(t), \qquad W = YX^\top(XX^\top+\lambda I)^{-1},$$

with $\lambda=1$ (Sumi et al., 2022).

The study reports that modular BNNs classify static spatial patterns with mean accuracy $74.8 \pm 20.2\%$ across $n=21$ networks, compared with $5$0 for a label-shuffled null, and that functional modularity $5$1 correlates strongly with a trajectory-separability metric with Pearson $5$2, $5$3 (Sumi et al., 2022). A timer task shows short-term memory on the order of $5$4, with mean $5$5 for $5$6 across $5$7 networks, and spoken digit classification reaches $5$8 over $5$9 networks (Sumi et al., 2022). Most notably, transfer learning is demonstrated: digit classification under gender switch yields $10\,\mu\text{m}$0 for $10\,\mu\text{m}$1, and gender classification under digit switch yields $10\,\mu\text{m}$2, whereas direct linear decoding of the input cochleagrams falls to chance under domain shift (Sumi et al., 2022). This is the empirical basis for the description of biological neurons as a “generalization filter.”

A complementary formulation appears in “From Neurons to Computation: Biological Reservoir Computing for Pattern Recognition,” where a cultured neuronal network on a $10\,\mu\text{m}$3 high-density MEA with 4096 electrodes is used as a biological reservoir, with inputs delivered on a subset of electrodes and the post-stimulus spike counts over all electrodes forming a $10\,\mu\text{m}$4-dimensional reservoir state vector (Iannello et al., 6 May 2025). In that system, a single-layer perceptron trained on 10 ms post-stimulus spike-count windows achieves $10\,\mu\text{m}$5 on pointwise stimuli, $10\,\mu\text{m}$6 on oriented bars, and $10\,\mu\text{m}$7 on a three-digit recognition task, with performance close to or exceeding a 4096-unit artificial echo-state-network baseline depending on task structure (Iannello et al., 6 May 2025).

Taken together, these studies establish a concrete BioNeuralNet paradigm in which biological neurons are not merely training targets or inspiration but the reservoir itself. Their significance lies in showing that modularity, heterogeneity, noise, and transient biophysical dynamics can improve separability and cross-domain generalization in ways that are difficult to reproduce by direct linear decoding.

3. BioNeuralNet as a GNN-based multi-omics analysis framework

A different and explicitly named usage appears in “BioNeuralNet: A Graph Neural Network based Multi-Omics Network Data Analysis Tool,” where BioNeuralNet is a flexible and modular Python framework for end-to-end network-based multi-omics data analysis (Ramos et al., 27 Jul 2025). The framework operates on a graph $10\,\mu\text{m}$8 whose nodes represent molecular entities or modalities and whose edges encode similarity, co-expression, or supervised multi-omics associations. It supports ingestion of precomputed networks such as WGCNA or SmCCNet outputs, or construction from tabular omics matrices using similarity networks, correlation networks, $10\,\mu\text{m}$9-nearest-neighbor graphs, shared nearest-neighbor variants, and phenotype-driven networks (Ramos et al., 27 Jul 2025).

The core modeling layer provides interchangeable GNN back ends—GCN, GAT, GraphSAGE, and GIN—implemented through PyTorch and PyTorch Geometric interfaces (Ramos et al., 27 Jul 2025). The paper gives standard message-passing forms, including the GCN layer

$623\,\text{nm}$0

and the GAT attention mechanism

$623\,\text{nm}$1

For supervised disease prediction, the DPMON workflow combines GNN-derived embeddings, dimensionality reduction, fusion with original omics and clinical covariates, and a feed-forward classifier trained with cross-entropy loss (Ramos et al., 27 Jul 2025).

The TCGA-BRCA case study illustrates the full pipeline. After QC and harmonization, the dataset contains 769 subjects with complete data: DNA methylation $623\,\text{nm}$2, mRNA $623\,\text{nm}$3, miRNA $623\,\text{nm}$4, and clinical $623\,\text{nm}$5; feature selection retains the top 6,000 features for mRNA and DNA methylation by variance, ANOVA-F, and random forest importance, all 503 miRNA features, and 10 clinical features via random forest importance; a $623\,\text{nm}$6-nearest-neighbor cosine similarity network is then built with $623\,\text{nm}$7 (Ramos et al., 27 Jul 2025). For PAM50 subtype classification, BioNeuralNet reports accuracy $623\,\text{nm}$8, weighted F1 $623\,\text{nm}$9, and macro F1 $20\,\text{Hz}$0, compared with RandomForest $20\,\text{Hz}$1, MOGONET $20\,\text{Hz}$2, MLP $20\,\text{Hz}$3, and SUPREME $20\,\text{Hz}$4 in accuracy (Ramos et al., 27 Jul 2025).

In this software-framework sense, BioNeuralNet does not denote living neurons. It denotes an extensible GNN-centered environment for converting molecular interaction structure into task-ready embeddings. The significance of this usage is methodological: it transfers the BioNeuralNet label from physical neurocomputation to computational biology, while retaining the idea that networks with biological semantics should be first-class computational objects.

4. Biophysically grounded and connectomics-constrained neural architectures

A broader BioNeuralNet lineage consists of artificial or hybrid neural architectures whose variables, topology, or learning rules are explicitly tied to biophysics or connectomics. In “Neurobiological reality simulation through an Artificial Neural Network at criticality,” an ANN with binary units $20\,\text{Hz}$5, asymmetric fully connected synapses, and a global energy

$20\,\text{Hz}$6

is evolved with a Fermi–Dirac update rule,

$20\,\text{Hz}$7

so that the local order parameter becomes $20\,\text{Hz}$8 and the membrane-potential analog is defined by

$20\,\text{Hz}$9

(Contoyiannis et al., 2019). At criticality, the model produces “grass” fluctuations between spikes with laminar-length statistics consistent with a power law, reporting $N=60$0, $N=60$1, and $N=60$2, in close agreement with experimentally reported CA1 pyramidal-neuron values $N=60$3 and $N=60$4 (Contoyiannis et al., 2019). The paper presents this as a route toward neural computation grounded in statistical mechanics rather than conventional rate-based abstractions.

A more recent connectomics-driven variant is “BioNIC: Biologically Inspired Neural Network for Image Classification Using Connectomics Principles,” which builds a multi-layer feedforward classifier whose hidden connectivity is masked by adjacency matrices derived from a single cortical column in the MICrONS minnie65_public dataset (Prasanth et al., 20 Jan 2026). The model preserves laminar organization, includes graded inhibition based on inhibitory in-degree, uses Hebbian synaptic plasticity with homeostatic regulation, LayerNorm, synaptic noise, and convolutional front ends “mimicking retinotopic mapping,” and attains $N=60$5 accuracy on FER-2013, comparable to a conventional baseline at $N=60$6 (Prasanth et al., 20 Jan 2026). Its ablations show that data augmentation, convolutions, and LayerNorm dominate performance, whereas connectivity masks, intra-layer connections, and graded inhibition contribute smaller but measurable gains (Prasanth et al., 20 Jan 2026).

Other papers extend the same logic to mechanistic modeling or biochemical realization. “Biologically-informed neural networks guide mechanistic modeling from sparse experimental data” constrains neural approximators by the reaction–diffusion form

$N=60$7

with $N=60$8, $N=60$9, and $$\mathbf{y}(t)=W\mathbf{x}(t), \qquad W = YX^\top(XX^\top+\lambda I)^{-1},$$0 learned jointly with the solution field under PDE residual and biological monotonicity constraints, then distilled into an explicit delay–reaction–diffusion model (Lagergren et al., 2020). “Automatic Implementation of Neural Networks through Reaction Networks — Part I” goes further by constructing a programmable biochemical reaction network that realizes a fully connected neural network under mass-action kinetics, including feedforward propagation, backpropagation, automatic sample assignment, and a judgment termination module (Siddiq et al., 2023). These systems suggest that BioNeuralNet can also denote neural computation embedded directly in mechanistic biological or biochemical dynamics.

5. Closed-loop, embodied, and biohybrid BioNeuralNets

A major strand of the literature treats BioNeuralNet as a closed-loop hybrid system in which living neural tissue is interfaced with digital hardware under strict timing and stimulation semantics. “CL API: Real-Time Closed-Loop Interactions with Biological Neural Networks” frames biological neural networks as computing substrates that require temporally and structurally consistent input delivery, microsecond-scale timing, multi-channel synchronization, and real-time response to activity (Hogan et al., 12 Feb 2026). The reference implementation uses an FPGA timing engine with base frequency $$\mathbf{y}(t)=W\mathbf{x}(t), \qquad W = YX^\top(XX^\top+\lambda I)^{-1},$$1, minimum $$\mathbf{y}(t)=W\mathbf{x}(t), \qquad W = YX^\top(XX^\top+\lambda I)^{-1},$$2 period, and a $$\mathbf{y}(t)=W\mathbf{x}(t), \qquad W = YX^\top(XX^\top+\lambda I)^{-1},$$3 frame clock for observation loops; a neurons.stim(...) call executes in approximately $$\mathbf{y}(t)=W\mathbf{x}(t), \qquad W = YX^\top(XX^\top+\lambda I)^{-1},$$4, with stimulation beginning $$\mathbf{y}(t)=W\mathbf{x}(t), \qquad W = YX^\top(XX^\top+\lambda I)^{-1},$$5–$$\mathbf{y}(t)=W\mathbf{x}(t), \qquad W = YX^\top(XX^\top+\lambda I)^{-1},$$6 after admission when channels are idle, while the loop at $$\mathbf{y}(t)=W\mathbf{x}(t), \qquad W = YX^\top(XX^\top+\lambda I)^{-1},$$7 provides roughly $$\mathbf{y}(t)=W\mathbf{x}(t), \qquad W = YX^\top(XX^\top+\lambda I)^{-1},$$8–$$\mathbf{y}(t)=W\mathbf{x}(t), \qquad W = YX^\top(XX^\top+\lambda I)^{-1},$$9 of user logic budget per iteration (Hogan et al., 12 Feb 2026). The API’s contract specifies deterministic per-channel ordering, explicit synchronization, transactional admission, and semantic consistency across hardware and simulator, thereby turning the BNN plus interface into a reproducible hybrid control system (Hogan et al., 12 Feb 2026).

“Embodied Neurocomputation” uses this type of interface to define an encoding–biology–decoding–feedback loop,

$\lambda=1$0

where $\lambda=1$1 is an odor-gradient observation, $\lambda=1$2 is a binary spatiotemporal stimulation pattern, $\lambda=1$3 is the biological response, and $\lambda=1$4 is the decoded action (Zhou et al., 13 May 2026). In a 6×6 grid-world, the study evaluates approximately 1,300 parameter combinations across 26 human stem-cell-derived neural cultures over about 4,000 hours of real-time agent–environment interaction, identifying 12 configurations that consistently demonstrate learning (Zhou et al., 13 May 2026). Those top configurations outperform optimized DQN baselines under the same interaction budget, with BNN agents achieving approximately $\lambda=1$5 DQN performance in a single 150-step episode regime and approximately $\lambda=1$6 DQN performance in a five-episode regime, and much larger gains over non-adaptive baselines (Zhou et al., 13 May 2026).

The biohybrid interpretation can be even more literal. “A geographically distributed bio-hybrid neural network with memristive plasticity” links rat hippocampal neurons in Padova, memristive synaptors in Southampton, and VLSI neurons in Zurich into a single feed-forward three-node circuit, ANPRE $\lambda=1$7 BN $\lambda=1$8 ANPOST, using a modified BCM rule with LTP above $\lambda=1$9, no plasticity between $74.8 \pm 20.2\%$0 and $74.8 \pm 20.2\%$1, and LTD below $74.8 \pm 20.2\%$2 (Serb et al., 2017). Communication occurs over UDP with measured static delays of $74.8 \pm 20.2\%$3–$74.8 \pm 20.2\%$4 and jitter below $74.8 \pm 20.2\%$5, and the experiment demonstrates that memristive weights can modulate EPSP-like responses and spiking in the biological neuron and the firing of the downstream neuromorphic neuron despite geographical separation (Serb et al., 2017). In a different embodied direction, “Gridbot” shows an autonomous robot controlled by a spiking neural network inspired by head-direction, grid, place, and border cells, using local STDP and sparse, structured connectivity rather than all-to-all coupling (Tang et al., 2018).

These systems move BioNeuralNet from analysis or simulation to deployable hybrid control. Their significance lies in establishing deterministic software contracts, physically realized learning loops, and task-driven validation for biological neural computation.

6. Design principles, limitations, and outlook

Several recurrent design principles emerge across these usages. In living reservoirs, functional modularity improves class separability, task stability, and transfer performance, while excessive global synchrony degrades effective dimensionality (Sumi et al., 2022). In multi-omics software, modularity appears as an engineering principle: BioNeuralNet exposes network construction, subgraph detection, GNN back ends, and downstream tasks as separable but composable pipeline stages (Ramos et al., 27 Jul 2025). In connectomics-constrained and criticality-based models, inductive biases are supplied by laminar organization, inhibitory structure, homeostatic regulation, statistical-mechanical order parameters, or explicit energy functions (Prasanth et al., 20 Jan 2026). In the broader comparative literature, representational complexity, complex network structure and energetics, robust function, and embodiment are proposed as guiding principles for future biologically grounded neural systems (Katyal et al., 2021).

The limitations are equally consistent. Living-neuron reservoirs are measured through calcium imaging at $74.8 \pm 20.2\%$6, so the accessible state is effectively low-pass filtered; readouts are modest in scale ($74.8 \pm 20.2\%$7 neurons in one study), training is offline, and culture-to-culture variability is substantial (Sumi et al., 2022). The multi-omics framework does not explicitly address batch effects or complex experimental designs and notes that scaling to very large graphs may require specialized sampling or distributed training (Ramos et al., 27 Jul 2025). Closed-loop BNN interfaces cannot guarantee deterministic biological responses even when timing semantics are deterministic, and finite queue capacity or large stimulation plans may trigger transaction rejection (Hogan et al., 12 Feb 2026). Embodied neurocomputation currently explores a constrained six-parameter encoding space and a simple low-dimensional task, so its conclusions do not yet establish generality for high-dimensional sensory or continuous-control settings (Zhou et al., 13 May 2026).

A plausible synthesis is that BioNeuralNet now names a research program rather than a single architecture. In that program, biology can appear as substrate, topology, dynamics, semantics, or software object. Living neurons may act as reservoirs or adaptive agents; biomolecular networks may be embedded as graphs and learned by GNNs; cortical columns may become adjacency masks in feedforward models; and standardized APIs may render neural cultures programmable under real-time contracts. The unifying technical question is how much computational advantage follows from preserving biological structure rather than abstracting it away.