A ceLLM framework for molecular priors bioelectric context and adaptive form
John Coates | RF Safe | September 2026
Perspective and research proposal
Abstract
Physical systems can embody relationships whose mathematical description requires far more dimensions than the space they occupy. A neural network implements a high-dimensional transformation through ordinary computing hardware. A cell likewise couples many molecular, electrical, and mechanical variables within a small physical volume. We propose ceLLM as a framework for investigating how that organization constrains cellular responses. DNA and chromatin contribute an evolved and adjustable physical prior; cellular machinery implements context-sensitive processing; and the body provides a changing runtime that cells themselves continually reconstruct. This perspective connects the bioelectric research of Robert O. Becker and Michael Levin, models of collective computation, and the search for geometric organizing principles in quantum field theory associated with Nima Arkani-Hamed, Jaroslav Trnka, Carolina Figueiredo, and their collaborators. Stephen Wolfram’s computational proposals supply a separate candidate account of how relational structure could underlie physical description. The connection proposed here concerns physically realized constraints on possible transformations. We distinguish this synthesis from a demonstrated unification of the underlying theories. Its practical contribution is a research program for measuring how molecular organization changes response probabilities, how temporal context selects among responses, and when perturbations reduce a system’s ability to distinguish and act on biologically relevant inputs. A simple dynamical example and a controlled exposure design make these claims concrete.
Higher dimensional organization within physical systems
A trained artificial neural network can be recorded as arrays of numbers. Those numbers remain information when printed on paper, but paper does not execute the transformations that give the trained system its capabilities. Computation requires a physical implementation of relationships: stored parameters, changing activity, and operations that connect one state to the next. The geometry used to describe those relationships can have thousands or millions of coordinates while the hardware remains entirely within ordinary spacetime.
The same distinction clarifies ceLLM. A cell contains molecular arrangements in three-dimensional space, yet describing its condition may require coordinates for membrane voltage, ion concentrations, metabolic state, chromatin accessibility, protein activity, mechanical strain, and recent stimulation. A point in this state space represents one possible combination of those variables. A trajectory represents their coordinated change. The additional dimensions count relevant variables; they do not require additional spatial directions or an incoming signal from an external realm.
This relational organization can be physically consequential. Hopfield’s work showed how specified network interactions can support associative recall through collective dynamics [1]. Sadtler and colleagues later found that existing neural population structure constrained how readily monkeys learned different brain-computer interface mappings [2]. These results concern particular networks and tasks. They motivate a precise question for cells: which aspects of their organization constrain the responses they can learn, maintain, or recover?
Three uses of geometry need separate definitions. Real-space geometry concerns molecular positions, shapes, and contacts. State-space geometry concerns relationships among combinations of measured variables. Information geometry concerns relationships among probability distributions, often through their statistical distinguishability. A model can connect these descriptions, but a visual resemblance between two plots does not establish that connection. Distances in an embedding can change when its coordinates or scaling change. Biologically meaningful conclusions therefore need observable anchors, such as transition probabilities, recovery times, or performance on a defined task.
Hartl, Pio-Lopez, Fields, and Levin recently proposed a framework connecting natural and artificial cognition through the construction and remapping of embedding spaces and navigation within those spaces [3]. ceLLM adopts a complementary focus: identifying the molecular and physiological implementation that makes a cell’s response repertoire possible. Here, intelligence means an operational capacity to use information to sustain or improve performance under changing conditions. A claim about cellular intelligence does not by itself establish subjective experience or solve the problem of consciousness.
From pattern formation to bioelectric control
The historical foundation extends beyond the past fifty years. Turing’s 1952 account of morphogenesis showed how interacting and diffusing chemical species could produce spatial patterns under specified conditions [4]. It provided a constructive example of organized form emerging from local physical processes. Waddington’s developmental landscape later offered a way to think about alternative developmental paths and the stabilization of cell identities [5]. A landscape in this sense is an organizing model; its valleys need not be literal minima of thermodynamic energy.
Robert O. Becker helped bring electrical organization into the experimental study of regeneration. His 1960 work mapped direct-current potentials at the salamander surface, and his 1961 paper examined bioelectric factors in amphibian limb regeneration [6,7]. Small, preliminary groups received currents of different polarities at different stages: an early positive current enlarged the regeneration blastema while delaying differentiation, whereas a later negative current increased axial growth. The design did not isolate timing and polarity independently. These findings made electrical conditions part of the problem of how living tissue repairs itself, without establishing a universal electrical regeneration mechanism.
Modern developmental bioelectricity supplies more specific interventions. Adams, Masi, and Levin showed that proton-pump activity was required for regeneration in a Xenopus tail model and that an introduced proton pump could restore regenerative responses under tested conditions [8]. Pai and colleagues manipulated membrane voltage during Xenopus development and altered eye formation, including induction of eye tissue in ectopic locations; their experiments connected voltage to calcium-dependent transcriptional control [9]. Such interventions demonstrate that electrical state can influence how developmental machinery is deployed.
Persistent physiological history also matters. Durant and colleagues reported that transient disruption of bioelectric communication could produce planarians that repeatedly regenerated a two-headed morphology after later amputations, without further treatment [10]. The experiments support durable changes in pattern regulation. They do not establish that the genome becomes irrelevant, nor do they identify a single exclusive storage site for all anatomical information.
Together, these studies support the distinction at the center of ceLLM. Inherited machinery supplies capabilities. Physiological context helps select their expression. Stable tissue and cellular states can preserve earlier selections. Morphology emerges through their interaction and then changes the signals, forces, and boundaries encountered by the cells. The body is consequently an active runtime with persistent state, rather than a static container around otherwise independent processors.
What amplitude geometry contributes
Carolina Figueiredo received the 2026 Vera Rubin New Frontiers Prize for work on the geometric structure of scattering amplitudes and relations among quantum field theories [11]. Her work with collaborators belongs to a broader effort to uncover structures that conventional calculations can obscure.
Arkani-Hamed and Trnka’s amplituhedron gives a geometric formulation of scattering amplitudes and loop integrands in planar, maximally supersymmetric Yang-Mills theory [12]. The relevant mathematical object encodes information conventionally assembled through Feynman diagrams. Its canonical differential form should not be confused with the ordinary volume of a solid. Within that framework, physical consistency properties can follow from geometric conditions.
Subsequent work involving Arkani-Hamed, Qu Cao, Jin Dong, Figueiredo, and Song He identified hidden zeros and relations among selected scalar, pion, and gluon amplitudes [13]. Their surface-kinematics work also addressed the definition of planar nonsupersymmetric Yang-Mills loop integrands using curves on surfaces and demonstrated specified consistency properties [14]. These developments broaden the reach of geometric and combinatorial methods. They do not establish that a single amplituhedron describes every quantum system or a living cell.
The methodological lesson is valuable for biology. A long catalogue of interactions can sometimes be reorganized by identifying constraints that many interactions share. For ceLLM, the corresponding question is whether measured molecular relationships organize many cellular responses more economically and predictively than treating each response as an unrelated pathway. This is a proposed transfer of research strategy. A genuine mathematical bridge would require an explicit mapping that preserves relevant quantities and yields new predictions; shared geometric language alone does not supply one.
Computation and geometry in Wolfram models
Stephen Wolfram approaches the relation between computation and physics from a different starting point. His ruliad is a proposed limiting structure involving possible computations and their relationships. Particular rewriting systems and descriptions available to observers provide candidate routes toward physical laws [15]. These are foundational proposals, distinct from the experimentally established use of quantum field theory.
Arsiwalla and Gorard have investigated a direct connection from Wolfram-model rewriting systems to pregeometric spaces using homotopy types [16]. In this work, relations and transformations help construct topological structure without taking a background geometry as the initial ingredient. It is a relevant example of how a computational description can acquire geometric organization.
For ceLLM, this suggests a research hierarchy. Relational rules can define possible transformations; physical realization determines which transformations occur and at what rates; biological organization can use those transformations in regulation and adaptation. It remains necessary to derive the connections between these levels. No result discussed here identifies a biological regulatory network with the ruliad or with an amplitude geometry.
Topology, metric geometry, and dynamics also have different jobs. Topology describes qualitative connectedness. A metric defines quantitative distances. Dynamics specifies how states change. A perturbation can alter timing and transition rates while preserving network connectivity. Describing every functional change as a topological change would therefore lose information that ceLLM needs to explain.
The ceLLM architecture
We propose ceLLM as a model of distributed cellular inference whose implementation is molecular and physiological. The central unit is a living cell with inherited capabilities, retained history, access to local signals, and the capacity to alter its surroundings. Cells collectively construct the conditions under which subsequent responses occur.
DNA and chromatin contribute the physical prior. In this usage, a prior is a predisposition built into a system before the next input arrives. DNA sequence reflects evolutionary history and provides templates and regulatory elements. Chromatin contributes developmentally established and adjustable accessibility and organization. Calling this system trained emphasizes accumulated adaptation; it does not imply that evolution performs the same optimization procedure as training a modern language model.
There is direct evidence that physical regulatory organization matters. Lupiáñez and colleagues linked structural changes around chromatin domain boundaries to altered enhancer contacts, gene expression, and developmental abnormalities [17]. Yet geometry does not act as an unrestricted master variable: Rao and colleagues found that loss of cohesin removed chromatin loop domains while immediate transcriptional effects were comparatively limited [18]. The consequence of changing organization depends on the affected relationships, cellular state, timescale, and compensatory mechanisms.
The phrase atomic neural network can therefore serve as a research hypothesis about organized molecular interactions. To become a mechanism, it must specify the relevant units, couplings, nonlinear responses, persistence, and rules of adjustment. DNA, RNA, proteins, water, ions, and chromatin operate as a dynamic molecular system. Neural-network language does not establish a literal crystalline lattice or a particular quantum computing architecture.
| Biological component | Proposed computational role | What an experiment must identify |
| DNA and chromatin | Inherited and adjustable response constraints | Sequence, accessibility, contacts, and causal regulatory effects |
| Cellular machinery and retained state | Local processing and memory | Relevant kinetics, feedback, persistence, and response rules |
| Bioelectric and metabolic signals | Context and coordination | Voltage, ion activity, energy state, and temporal relationships |
| Cytoskeleton and extracellular matrix | Mechanical implementation and context | Forces, organization, coupling, and changes in response |
| Neighboring cells and body morphology | Shared runtime and recurrent input | Communication routes, tissue boundaries, and collective outcomes |
The roles overlap. Mitochondria supply both energy and signals. The cytoskeleton both implements actions and changes the inputs a cell receives. Chromatin can retain history while also responding rapidly. A cell can be a local inference engine without being an autonomous or sufficient explanation of tissue behavior.
Neural cellular automata offer an especially clear computational comparison. Mordvintsev, Randazzo, Niklasson, and Levin trained a common local update rule that could grow, maintain, and in some cases regenerate a target pattern [19]. The learned rule supplies shared capabilities; neighboring states supply context; repeated local updates produce the visible form. This is a computational demonstration of the architecture, not evidence that living cells implement the particular artificial network used in the model.
Figure 1. The proposed ceLLM architecture. Molecular organization constrains cellular responses, while physiological and tissue conditions supply context. Cellular actions alter morphology and the local environment, producing subsequent inputs. The diagram represents functional roles rather than a complete inventory of biological mechanisms. Persistent state can occur within several roles.
A model that can be measured
The smallest useful formal commitment is a conditional response model: how likely is a specified action after a particular input history, given measured organization and cellular condition?
P(a | u0:t, x0, G, θ)
Here a is a measured action or outcome, u from time zero to t is the input history, x is the initial cellular state, G records specified physical or regulatory relationships, and theta contains kinetic parameters. Outcomes might include a transcriptional response, directional migration, or a repair decision. The notation organizes a research problem; it does not claim that cells explicitly calculate probability tables.
Calling the response inference is stronger than calling it regulation. A rigorous inference model should identify the environmental quantity being estimated, the observations available to the cell, and the consequences of a correct or incorrect estimate. Kobayashi constructed a theoretical example in which intracellular kinetics implement Bayesian decision making [20]. That provides a precedent for mechanistic inference models, while leaving the implementation in any particular experimental cell to be demonstrated.
An operational test of cellular cognition would ask whether retained information about past inputs improves a defined response when conditions change. History dependence by itself is insufficient: a passive material can have hysteresis. Evidence becomes stronger when a model predicts useful adaptation and a targeted intervention identifies the contribution of the proposed memory mechanism.
We propose measuring response geometry through reproducible relationships among experimentally determined response distributions. Cells with similar responses across a controlled panel of input histories occupy nearby positions under a declared statistical comparison; cells with different responses are separated. The choice of metric, measured variables, and observation window must be reported. A low-dimensional picture is a display of the model, not proof that its axes are biological causes.
The important prediction runs from independently measured organization to responses under new conditions. Estimate contacts or coupling strengths and relevant timescales, then predict the consequences of an intervention or input sequence held out from fitting. Compare this prediction with conventional gene-regulatory and signaling models. Added geometric structure earns explanatory value if it improves prediction, transfer between conditions, or causal interpretation without merely increasing model flexibility.
Memory and regeneration across timescales
The ceLLM architecture distinguishes evolutionary inheritance, developmental state, physiological memory, and learned content. A genome can support the ability to construct or repair a nervous system without encoding every connection or every experience acquired by that organism. Persistent regulatory states can influence reconstruction, while remaining tissue and sensory interactions supply additional information.
Regeneration experiments therefore require careful interpretation. Shomrat and Levin’s planarian experiments reported evidence of savings after head regeneration: previously familiarized animals performed differently after a reminder exposure. Direct recall after regeneration was not statistically significant in their reported test [21]. The surviving body contained substantial biological structure, so the result did not localize the retained information exclusively to DNA. Recovering function, reconstructing an earlier connectome, and retaining an acquired memory are distinct outcomes.
The analogy with AI becomes productive when these distinctions are maintained. A trained model’s parameters, temporary context, and persistent working records can all affect its next output. Likewise, the genome, chromatin, signaling state, and tissue environment can jointly influence a cell. Assigning a primary role to DNA and chromatin does not require all other biological states to be memoryless.
This also clarifies the question of discovery versus creation. An engineer can discover an effective parameter configuration while creating a physical system that implements it. Evolution can select organizations permitted by physical law while development constructs their living realization. Whether mathematical possibilities possess an independent ontological existence remains a separate question. ceLLM’s experimental program can investigate their physical implementation without requiring that question to be settled first.
Timing and biological fidelity
Geometry becomes informative when connected to dynamics. Two systems can have the same network connections and different behavior because their delays, rates, or noise differ. Biological signals therefore need characterization over time as well as by average magnitude.
Dolmetsch, Xu, and Lewis demonstrated that controlled calcium oscillations could change the efficiency and specificity of transcriptional responses [22]. Kukushkin and colleagues showed that two immortalized human cell preparations could distinguish spaced chemical stimulation from a massed exposure with matched total stimulus-on time, using a transcriptional reporter as a memory proxy [23]. These findings establish temporal sensitivity in particular preparations. They do not establish an electromagnetic cause or human-like reasoning by the cells.
A simple model shows why timing can matter without changing the amount of input. Let a signaling intermediate x accumulate in response to an input u and decay with a measured timescale tau.
dxdt = −xτ + u(t)
Deliver two rectangular pulses of equal height and width, with their start times separated by Delta. Let A be the level reached at the end of one isolated pulse. Starting from zero, the level immediately after the second pulse is given by the following expression.
x2A = 1 + exp(−Δτ)
For separations of 0.2 tau and 2 tau, the peak after the second pulse is approximately 1.82 A and 1.14 A respectively, provided pulse width is less than 0.2 tau. A downstream threshold of 1.5 A distinguishes the two patterns. The pulses have the same height, width, count, and total input over the same observation window. Only their separation differs. This is temporal filtering; demonstrating inference or intelligence requires the additional task and performance criteria described above. Other networks can prefer spaced rather than clustered inputs; the response depends on the measured mechanism.
In this framework, biological fidelity means the reliability with which a system distinguishes and responds to relevant input histories for a specified function. A change in an average reporter value or a projected trajectory is not automatically a loss of fidelity. A useful measure would combine held-out discrimination of input patterns, response reproducibility, and functional performance. A fixed statistical decoder can fail after harmless rescaling of an output. Discriminability should therefore also be tested with an appropriately recalibrated decoder and compared with the actual downstream biological response. Reduced responsiveness might represent impairment, adaptation, or protection; the task and subsequent outcomes determine which interpretation is warranted.
Figure 2. An illustrative timing model, not experimental data. Identical pulse pairs produce different peaks in a decaying intermediate. Time is expressed in units of tau, pulse width is 0.1 tau, and intermediate abundance is normalized to the isolated-pulse peak A. The example concerns biological input after it has reached a signaling pathway. It assumes no coupling from radiofrequency exposure to that pathway.
Cellular organization is also maintained away from equilibrium. Work on sensory adaptation illustrates how accurate regulation can require continuing energy dissipation [24]. A stable biological attractor should consequently not be identified automatically with a thermodynamic ground state. ceLLM must include energy supply, material turnover, and feedback as part of the runtime that maintains its apparent geometry.
Experiments that distinguish the proposed mechanisms
The most useful experimental distinction is between a change in the cell’s response machinery and a change in the context acting on that machinery. Both can produce similar outward behavior. A crossed design can vary a specified regulatory state and an input pattern separately, then measure their interaction. Restoring the original context, removing the intervention, and testing again can help distinguish transient disruption from retained change. Because context can itself remodel chromatin or metabolism, the relevant internal states must be measured rather than inferred solely from recovery.
A first study could reproduce a known cellular timing response, estimate its kinetics, and predict its response to new input sequences. An independently measured decay or feedback timescale should predict where sensitivity changes. Selective perturbation of the proposed molecular relationship should then shift that response in the predicted direction. This tests whether the geometry-to-dynamics mapping explains more than a descriptive fit.
The resonance branch of ceLLM requires additional commitments. Specify a molecular or collective mode, its frequency, damping, coupling to an applied field, and a downstream response. Measure these under physiological conditions. An observed vibrational mode alone would establish molecular motion; the distinctive evidence would be a correctly predicted change in biological behavior mediated by that mode. If adequately sensitive measurements exclude the predicted mode or coupling, that particular mechanism fails in the tested regime.
For the RF Safe research program, a tractable question is whether a calibrated electromagnetic exposure alters a cell’s ability to distinguish biological temporal inputs. A reporter preparation with a reproducible spaced-versus-massed chemical response supplies a positive control for timing sensitivity [23]. The original study included a reporter measurement 24 hours from stimulation onset; comparisons should specify both time from onset and time from the final pulse. Randomized, blinded exposure can then be introduced while early signaling trajectories and later transcriptional outcomes are measured. The biological timing effect must be established independently of the proposed RF effect.
Exposure comparisons should include sham, continuous-wave, periodic-pulse, and temporally jittered conditions where appropriate. Average absorbed power alone does not isolate timing: duty cycle can change peak fields, and waveform changes can alter temperature histories. A periodic-versus-jittered comparison should match pulse height, width, count, and total energy over a common interval, with timing as the intended difference. Report dosimetry, spectra, spatial distributions, uncertainty, and temperature at the available resolution. Thermal replay controls and measurements insensitive to the exposure apparatus are needed to test alternative explanations.
Existing null findings constrain mechanism selection. Platano and colleagues found no significant acute effect of their tested 900 MHz continuous-wave and GSM-modulated exposures on currents through voltage-gated calcium channels in cultured rat cortical neurons [25]. That result does not resolve all exposure conditions, but it prevents treating direct channel activation as an established universal consequence of RF exposure.
The proposed causal chain is exposure, an early physiological change, an altered response to biological input, and a functional consequence. Interference with the candidate mediator should remove the predicted effect while an appropriate rescue restores it. Such interventions must preserve or account for baseline signaling; pathway dependence alone does not identify the primary field sensor. Replication should use independent preparations, concealed analysis labels, cell-free instrument controls, and an orthogonal endpoint measured after exposure. Randomization and uncertainty estimates must respect the independently exposed preparation or run: many cells in one exposed dish are subsamples, not independent exposure replicates.
Preregister effect sizes and failure criteria. A result explained by measured heating, peak exposure, or instrumentation does not support an additional timing mechanism. A replicated impairment in input discrimination would support a specified loss of cellular information fidelity, with further work required to establish injury or generalize to environmental exposure. A well-powered null result narrows the theory rather than becoming evidence of an undetectable effect.
The connection proposed by ceLLM
The central proposal is that physical organization constrains a repertoire of possible responses, and that living systems continually select and reshape that repertoire through interactions with their surroundings. DNA and chromatin contribute durable constraints. Bioelectric, biochemical, and mechanical processes implement the changing context. Cellular activity generates morphology, and morphology changes the conditions of subsequent activity.
Amplitude geometry shows that compact mathematical structure can expose relationships hidden in complicated physical calculations. Computational approaches investigate how relations and transformations can generate organized structure. Developmental bioelectricity demonstrates that physiological state can causally redirect biological form. ceLLM brings these research questions into contact by asking which physically realized relationships organize cellular inference and how perturbing them changes performance.
ceLLM is presented here as a synthesis and research proposal. The distinctive test would be a predeclared mapping from independently measured organization to response distributions that predicts new timing conditions and intervention outcomes better than comparably flexible conventional models. Its value will be established by predicting what a cell does next, identifying why that prediction changes, and controlling the relevant relationships well enough to restore function when it is lost.
References
1. Hopfield JJ. Neural networks and physical systems with emergent collective computational abilities. Proceedings of the National Academy of Sciences. 1982;79:2554–2558. Read source
2. Sadtler PT, et al. Neural constraints on learning. Nature. 2014;512:423–426. Read source
3. Hartl B, Pio-Lopez L, Fields C, Levin M. Remapping and navigation of an embedding space via error minimization: A fundamental organizational principle of cognition in natural and artificial systems. Physics of Life Reviews. 2026;58:336–359. Read source
4. Turing AM. The chemical basis of morphogenesis. Philosophical Transactions of the Royal Society of London B. 1952;237:37–72. Read source
5. Waddington CH. The Strategy of the Genes. London: George Allen and Unwin; 1957. Read source
6. Becker RO. The bioelectric field pattern in the salamander and its simulation by an electronic analog. IRE Transactions on Medical Electronics. 1960;ME-7:202–207. Read source
7. Becker RO. The bioelectric factors in amphibian-limb regeneration. Journal of Bone and Joint Surgery American Volume. 1961;43-A:643–656. Read source
8. Adams DS, Masi A, Levin M. H+ pump-dependent changes in membrane voltage are an early mechanism necessary and sufficient to induce Xenopus tail regeneration. Development. 2007;134:1323–1335. Read source
9. Pai VP, Aw S, Shomrat T, Lemire JM, Levin M. Transmembrane voltage potential controls embryonic eye patterning in Xenopus laevis. Development. 2012;139:313–323. Read source
10. Durant F, et al. Long-term, stochastic editing of regenerative anatomy via targeting endogenous bioelectric gradients. Biophysical Journal. 2017;112:2231–2243. Read source
11. Breakthrough Prize Foundation. Carolina Figueiredo. 2026 Vera Rubin New Frontiers Prize. Official award record. Read source
12. Arkani-Hamed N, Trnka J. The amplituhedron. Journal of High Energy Physics. 2014;10:030. Preprint first posted in 2013. Read source
13. Arkani-Hamed N, Cao Q, Dong J, Figueiredo C, He S. Hidden zeros for particle/string amplitudes and the unity of colored scalars, pions and gluons. arXiv:2312.16282. 2023. Read source
14. Arkani-Hamed N, Cao Q, Dong J, Figueiredo C, He S. Surface kinematics and “the” Yang-Mills integrand. Physical Review Letters. 2025;134:171601. Preprint first posted in 2024. Read source
15. Wolfram S. The concept of the ruliad. Stephen Wolfram Writings. November 10, 2021. Author’s foundational proposal. Read source
16. Arsiwalla XD, Gorard J. Pregeometric spaces from Wolfram model rewriting systems as homotopy types. arXiv:2111.03460. 2021. Read source
17. Lupiáñez DG, et al. Disruptions of topological chromatin domains cause pathogenic rewiring of gene-enhancer interactions. Cell. 2015;161:1012–1025. Read source
18. Rao SSP, et al. Cohesin loss eliminates all loop domains. Cell. 2017;171:305–320.e24. Read source
19. Mordvintsev A, Randazzo E, Niklasson E, Levin M. Growing neural cellular automata. Distill. 2020. Read source
20. Kobayashi TJ. Implementation of dynamic Bayesian decision making by intracellular kinetics. Physical Review Letters. 2010;104:228104. Read source
21. Shomrat T, Levin M. An automated training paradigm reveals long-term memory in planarians and its persistence through head regeneration. Journal of Experimental Biology. 2013;216:3799–3810. Read source
22. Dolmetsch RE, Xu K, Lewis RS. Calcium oscillations increase the efficiency and specificity of gene expression. Nature. 1998;392:933–936. Read source
23. Kukushkin NV, Carney RE, Tabassum T, Carew TJ. The massed-spaced learning effect in non-neural human cells. Nature Communications. 2024;15:9635. Read source
24. Lan G, Sartori P, Neumann S, Sourjik V, Tu Y. The energy-speed-accuracy trade-off in sensory adaptation. Nature Physics. 2012;8:422–428. Read source
25. Platano D, et al. Acute exposure to low-level CW and GSM-modulated 900 MHz radiofrequency does not affect Ba2+ currents through voltage-gated calcium channels in rat cortical neurons. Bioelectromagnetics. 2007;28:599–607. Read source
