A major challenge in neuroscience lies in elucidating how collective neural circuit activity drives information processing and adaptive behavior (Bassett and Sporns, 2017). Bridging molecular synaptic mechanisms and emergent network-level functions requires experimental paradigms capable of both observing and perturbing circuit dynamics with precise spatial and temporal resolution (Frank et al., 2019; Liu et al., 2023). In vitro neuronal preparations, from dissociated cultures to brain slices and organoids, have long provided controlled environments for dissecting cellular and network mechanisms (Lv et al., 2023; Osaki et al., 2024). Recent technological progress has moved these preparations from passive observation toward active, closed-loop interrogation of plasticity, dynamics, and computation (Li L. et al., 2024), including pattern recognition (Shao et al., 2025), adaptive response (Kagan et al., 2022), and reservoir computing (Cai et al., 2023) demonstrated in recent work.

The introduction of microelectrode arrays (MEAs) marked a pivotal advance, enabling long-term, parallel monitoring of neuronal ensembles (Thomas et al., 1972; Obien et al., 2015). As discussed in Section 2, MEAs established the foundation for investigating network-level activity patterns and plasticity (Eytan and Marom, 2006). Their utility, however, is limited by a lack of cell-type specificity, insensitivity to subthreshold dynamics, and diffuse current spread during stimulation, which obscure mechanistic insight (Buzsáki, 2004). The development of complementary metal-oxide-semiconductor (CMOS) MEAs substantially improved spatial resolution, reaching densities above 20,000 electrodes per array (Schröter et al., 2025), yet issues such as electrode crosstalk and bandwidth constraints persist, as shown in our earlier studies (Habibollahi et al., 2023). Consequently, while MEAs excel at capturing population-level spiking, they cannot precisely resolve the underlying cellular interactions (Buzsáki, 2004).

Optical methods, particularly calcium imaging, offer superior spatial resolution and improved signal-to-noise performance for mapping network activity, as demonstrated in our recent studies (Sun et al., 2021; Sun et al., 2023; Sun et al., 2024a; Sun et al., 2024b). A paradigm shift in in vitro research occurred with the advent of all-optical interrogation, described in Section 3. By coupling optogenetic actuators with genetically encoded indicators, researchers can simultaneously manipulate and monitor defined neuronal populations with subcellular precision (<10 μm) (Deisseroth, 2011; Emiliani et al., 2015; Fernandez-Ruiz et al., 2022). This capability has transformed the field from correlational to causal investigation, revealing direct links between microcircuit connectivity, plasticity rules, and emergent computation (Jazayeri and Afraz, 2017; Sumi et al., 2023).

Electrophysiology and optical approaches have complementary strengths: the former captures ground-truth voltage signals at millisecond timescales, whereas the latter offers unparalleled access to genetically defined and spatially resolved populations (Ramezani et al., 2021). This complementarity has motivated the development of integrated multimodal interfaces that merge MEAs with optical stimulation and imaging, enabling concurrent electrical and optical interrogation (Buzsáki et al., 2015; Miccoli et al., 2019; Middya et al., 2021; Xu et al., 2024). Such hybrid systems permit real-time, closed-loop modulation of network activity, thereby allowing direct tests of how circuit dynamics encode and transform information. Although many multimodal technologies were originally developed in vivo, related concepts are now emerging in vitro (Shew et al., 2010; Yakushenko et al., 2013; Kshirsagar et al., 2019; Shaik et al., 2020; Middya et al., 2021; Shin et al., 2021), highlighting the growing feasibility of electrical–optical integration in culture.

This mini-review charts the conceptual and technological evolution of these approaches—from MEAs to all-optical systems and finally to integrated multimodal platforms (Figure 1). These innovations enable precise interrogation of vitro biological neural networks (BNNs), providing scalable and mechanistic models for probing how synaptic plasticity gives rise to computation and complex function.

The development of MEAs in the 1970s represented a foundational advance in in vitro neurophysiology, enabling simultaneous extracellular recordings from multiple neurons and laying the groundwork for scalable circuit-level investigations (Thomas et al., 1972). By embedding microelectrodes in a planar substrate, MEAs allowed minimally invasive, parallel recordings of extracellular spikes and local field potentials in cultured neurons and acute slices (Gross et al., 1977; Pine, 1980). This design enabled stable, long-term tracking of population activity, opening access to coordinated network dynamics previously beyond reach (Potter and DeMarse, 2001).

Initial MEA studies provided some of the first experimental demonstrations of network-level plasticity. Patterned electrical stimulation through MEA electrodes induced long-term potentiation (LTP) and depression (LTD), demonstrating that synaptic learning rules operate not only at individual connections but across neuronal ensembles (Jimbo et al., 1999). Later, spike-timing-dependent plasticity (STDP) paradigms demonstrated that precise temporal relationships between pre- and postsynaptic activity modulate circuit connectivity during development (Wagenaar et al., 2006). These findings extended classical synaptic principles into the mesoscale, linking cellular plasticity with emergent network behavior (Eytan and Marom, 2006).

However, these pioneering applications also revealed the limitations of electrode-based approaches, as shown in Figure 1A. MEAs detect only extracellular spikes but cannot access subthreshold or dendritic potentials essential for synaptic computation (Buzsáki, 2004). Although high-density CMOS MEAs approach micrometer-level spatial resolution, their non-planar surfaces are less compatible with microfabricated structures, such as PDMS microchannels, that require precise topography for axon guidance or modular network design (Duru et al., 2022). Moreover, extracellular signals lack inherent cell-type specificity: spikes from excitatory, inhibitory, or genetically defined neurons cannot be readily distinguished, limiting the interpretability of population activity and its mechanistic origins (Buzsáki et al., 2015). Electrical stimulation also suffers from current spread from the electrode tip, which activates a broad neuronal population beyond the target region (Yizhar et al., 2011). As a result, MEAs excel at tracking the timing and structure of network activity, but offer limited resolution into the cellular identities and synaptic mechanisms shaping those dynamics. These constraints also restrict the translational applications. In pharmacological studies, MEAs measure overall changes in the network, but cannot precisely resolve the receptor systems or intracellular pathways through which compounds act from electrophysiological data alone (McConnell et al., 2012; Saavedra et al., 2021). Similarly, developmental studies describe stereotyped patterns of network maturation, yet the absence of spatial and molecular specificity hinders mechanistic interpretation, such as identifying the contributions of receptor expression, synaptogenesis, or axon guidance (Lv et al., 2023).

In summary, MEAs established a durable and scalable platform for extracellular electrophysiology, and they remain a mainstay for long-term population-level recordings. However, their inherent spatial and functional limitations have motivated the pursuit of complementary techniques. Optical approaches have emerged to fill this gap, extending circuit interrogation beyond spiking activity to include underlying cellular and synaptic processes, and enabling the transition toward causal and multimodal paradigms (Emiliani et al., 2015).

All-optical platforms have transformed in vitro circuit neuroscience by integrating genetic specificity, precise spatiotemporal control, and high-content imaging within a single experimental framework. Optogenetic actuators, such as channelrhodopsins and their red-shifted or fast-kinetic variants, enable temporally precise control of defined neuronal populations at the millisecond scale, minimal spectral crosstalk and reduced phototoxic effects (Kishi et al., 2022; Tanaka et al., 2024). Complementary optical reporters extend observation across scales: genetically encoded calcium indicators (GECIs) like the GCaMP family, report population-level calcium transients that correlate with spiking (Zhang Y. et al., 2023), whereas genetically encoded voltage indicators (GEVIs) provide access to fast subthreshold (millisecond to sub-millisecond) and dendritic voltage fluctuations (Bando et al., 2019) (Figure 1A).

Spatial precision is achieved through structured-illumination strategies (Figure 1). One-photon systems employing digital micromirror devices (DMDs), which use high-speed arrays of tiltable micromirrors to project programmable light patterns, support rapid patterned excitation across wide fields (Dudley et al., 2003; Zhu et al., 2012). In contrast, two-photon holographic photostimulation via spatial light modulators (SLMs) enables volumetric targeting of user-defined ensembles in three dimensions (Papagiakoumou et al., 2018). These stimulation approaches are increasingly coupled with advanced imaging techniques, like resonant-scanning multiphoton microscopy, a high-speed method using resonant galvanometer mirrors to achieve micrometer-level resolution recording across planes and populations (Prevedel et al., 2016; Hsu et al., 2023). Moreover, MAPSI (Miniscope with All-optical Patterned Stimulation and Imaging) (Zhang J. et al., 2023) recently demonstrates compact integration of calcium imaging and patterned photostimulation, building on advances from the open-source Miniscope project (Cai et al., 2016). Although developed for in vivo use, these platforms exemplify all-optical interrogation strategies that are inspiring modular and miniaturized approaches for in vitro applications.

Collectively, these developments establish a versatile experimental framework for probing synaptic plasticity, functional connectivity, and emergent circuit dynamics (Bassett and Sporns, 2017; Zhang et al., 2022). Although many of these applications have been extensively explored in vivo, in vitro platforms remain highly valuable, offering experimental control and scalability that continue to drive methodological innovation and mechanistic discovery.

As shown in Figure 1B, integrating optical and electrophysiological modalities creates a cohesive platform for probing in vitro circuits across complementary domains of measurement and control (Kim et al., 2017; Ramezani et al., 2025) (Figure 1). While optical methods afford genetic specificity and volumetric access at submicron to micron spatial resolution, electrophysiology provides direct readout of voltage dynamics with sub-millisecond temporal precision. By co-registering both modalities within a unified spatial reference frame, perturbations and recordings can be aligned across techniques, enabling systematic comparisons between optically and electrically derived signals.

The following subsections trace the progression from parallel acquisition and cross-modal validation, through real-time feedback architectures for adaptive control, to closed-loop frameworks that interrogate learning rules and computational motifs (Alon, 2007). This progression illustrates how multimodal systems evolve from passive observation toward active modulation, ultimately supporting hypothesis-driven dissection of high-order function.

The progression from planar MEAs to all-optical interrogation and, ultimately, to integrated multimodal systems represents a pivotal shift in neural interfacing. The field has evolved from the correlational, spike-based population analyses enabled by MEAs to precise, cell-type-specific interrogation achieved through optical techniques. By fusing the temporal fidelity of electrophysiology with the spatial and genetic targeting of optogenetics and imaging, current platforms define a new experimental paradigm. This convergence has redefined in vitro models: from passive observation platforms to interactive, closed-loop environments capable of simultaneously sensing and manipulating neural activity in real time (Ramezani et al., 2021).

Looking forward, as outlined in Figure 2, this multimodal convergence will depend on three core enabling mechanisms, spanning hardware integration, computational analysis, and standardized experimental frameworks for translational research. On the hardware side, a key challenge is seamlessly integrating high-density electrode arrays with optical access while maintaining stable, low-noise performance. Advances in transparent electrode materials have demonstrated strong potential for simultaneous imaging and electrophysiology, yet optimizing transparency, impedance, durability, and fabrication scalability remains an active area of research (Qiang et al., 2018; Thunemann et al., 2018; Yang et al., 2024; Shankar et al., 2025).

From a translational standpoint, multimodal platforms are well positioned to reshape disease modeling and therapeutic screening (Wang et al., 2022). Integrating multimodal tools with patient-derived iPSC neurons and brain organoids enables the creation of in vitro models that approximate circuit-level dysfunctions underlying neurological and psychiatric disorders (Lv et al., 2023; Birtele et al., 2025). Such preparations offer controlled yet scalable context for dissecting disease mechanisms and linking pharmacological perturbations to changes in network function, though issues of variability and reproducibility still limit large-scale applications (Saavedra et al., 2021).

Complementing the experimental advances, progress will also rely on improving analytical pipelines. High-dimensional multimodal datasets require both established methods, such as spike sorting, dimensionality reduction, and functional connectivity analysis (Cunningham and Yu, 2014; Bastos and Schoffelen, 2016; Pachitariu et al., 2024), and emerging machine-learning and AI-driven approaches for state estimation and cross-modal data fusion (Saxe et al., 2021; Hong et al., 2024). The development of these computation tools will enable robust interpretation of neural activity across scales and modalities, maximizing the scientific yield of multimodal platforms.

Still, few studies have achieved the full integration of (i) bidirectional electrical–optical interfacing, (ii) simultaneous patterned stimulation and multimodal recording, (iii) in vitro system implementation, and (iv) closed-loop feedback (Xu et al., 2024; Ramezani et al., 2025). Realizing all four elements would complete the loop between perturbation and readout, offering a coherent framework for testing hypotheses about neural computation. Such a framework opens new territory at the interface of neuroscience and engineering, enabling direct investigation of how neural circuits adapt to structured feedback, reorganize connectivity, and develop predictive, goal-directed dynamics. Extending these paradigms to more complex organoid architectures may reveal how BNNs realize reinforcement and generalization—a cornerstone of adaptive learning capability (Neftci and Averbeck, 2019; Wülfing et al., 2019; Li Q. et al., 2024).

Taken together, multimodal in vitro systems are no longer solely experimental tools, but rather serve as both investigative platforms for fundamental neuroscience and living substrates for neuron-based computation. As such, they represent a critical inflection point: one in which mechanistic understanding, computational modeling, and physical embodiment converge within a unified experimental framework.