Explicit-duration hidden Markov model inference of UP-DOWN states from continuous signals

Neocortical neurons show UP-DOWN state (UDS) oscillations under a variety of conditions. These UDS have been extensively studied because of the insight they can yield into the functioning of cortical networks, and their proposed role in putative memory formation. A key element in these studies is determining the precise duration and timing of the UDS. These states are typically determined from the membrane potential of one or a small number of cells, which is often not sufficient to reliably estimate the state of an ensemble of neocortical neurons. The local field potential (LFP) provides an attractive method for determining the state of a patch of cortex with high spatio-temporal resolution; however current methods for inferring UDS from LFP signals lack the robustness and flexibility to be applicable when UDS properties may vary substantially within and across experiments. Here we present an explicit-duration hidden Markov model (EDHMM) framework that is sufficiently general to allow statistically principled inference of UDS from different types of signals (membrane potential, LFP, EEG), combinations of signals (e.g., multichannel LFP recordings) and signal features over long recordings where substantial non-stationarities are present. Using cortical LFPs recorded from urethane-anesthetized mice, we demonstrate that the proposed method allows robust inference of UDS. To illustrate the flexibility of the algorithm we show that it performs well on EEG recordings as well. We then validate these results using simultaneous recordings of the LFP and membrane potential (MP) of nearby cortical neurons, showing that our method offers significant improvements over standard methods. These results could be useful for determining functional connectivity of different brain regions, as well as understanding network dynamics.     

Automatic decomposition of multichannel intramuscular EMG signals

We describe an automatic algorithm for decomposing multichannel EMG signals into their component motor unit action potential (MUAP) trains, including signals from widely separated recording sites in which MUAPs exhibit appreciable interchannel offset and jitter. The algorithm has two phases. In the clustering phase, the distinct, recurring MUAPs in each channel are identified, the ones that correspond to the same motor units are determined by their temporal relationships, and multichannel templates are computed. In the identification stage, the MUAP discharges in the signal are identified using matched filtering and superimposition resolution techniques. The algorithm looks for the MUAPs with the largest single channel components first, using matches in one channel to guide the search in other channels, and using information from the other channels to confirm or refute each identification. For validation, the algorithm was used to decompose 10 real 6-to-8-channel EMG signals containing activity from up to 25 motor units. Comparison with expert manual decomposition showed that the algorithm identified more than 75% of the total 176 MUAP trains with an accuracy greater than 95%. The algorithm is fast, robust, and shows promise to be accurate enough to be a useful tool for decomposing multichannel signals. It is freely available at http://emglab.stanford.edu.     

Hierarchical Feedback Modules and Reaction Hubs in Cell Signaling Networks

Despite much effort, identification of modular structures and study of their organizing and functional roles remain a formidable challenge in molecular systems biology, which, however, is essential in reaching a systematic understanding of large-scale cell regulation networks and hence gaining capacity of exerting effective interference to cell activity. Combining graph theoretic methods with available dynamics information, we successfully retrieved multiple feedback modules of three important signaling networks. These feedbacks are structurally arranged in a hierarchical way and dynamically produce layered temporal profiles of output signals. We found that global and local feedbacks act in very different ways and on distinct features of the information flow conveyed by signal transduction but work highly coordinately to implement specific biological functions. The redundancy embodied with multiple signal-relaying channels and feedback controls bestow great robustness and the reaction hubs seated at junctions of different paths announce their paramount importance through exquisite parameter management. The current investigation reveals intriguing general features of the organization of cell signaling networks and their relevance to biological function, which may find interesting applications in analysis, design and control of bio-networks.     

Natural selection governs local, but not global, evolutionary gene coexpression networks in Caenorhabditis elegans

Background

Signal duration (e.g. the time over which an active signaling intermediate persists) is a key regulator of biological decisions in myriad contexts such as cell growth, proliferation, and developmental lineage commitments. Accompanying differences in signal duration are numerous downstream biological processes that require multiple steps of biochemical regulation.

Results

Here we present an analysis that investigates how simple biochemical motifs that involve multiple stages of regulation can be constructed to differentially process signals that persist at different time scales. We compute the dynamic, frequency dependent gain within these networks and resulting power spectra to better understand how biochemical networks can integrate signals at different time scales. We identify topological features of these networks that allow for different frequency dependent signal processing properties.

Conclusion

We show that multi-staged cascades are effective in integrating signals of long duration whereas multi-staged cascades that operate in the presence of negative feedback are effective in integrating signals of short duration. Our studies suggest principles for why signal duration in connection with multiple steps of downstream regulation is a ubiquitous motif in biochemical systems.     

Cyclic GMP-gated CNG channels function in Sema3A-induced growth cone repulsion

Cyclic nucleotide-gated channels (CNGCs) transduce external signals required for sensory processes, e.g., photoreception, olfaction, and taste. Nerve growth cone guidance by diffusible attractive and repulsive molecules is regulated by differential growth cone Ca2+ signaling. However, the Ca2+-conducting ion channels that transduce guidance molecule signals are largely unknown. We show that rod-type CNGC-like channels function in the repulsion of cultured Xenopus spinal neuron growth cones by Sema3A, which triggers the production of the cGMP that activates the Xenopus CNGA1 (xCNGA1) subunit-containing channels in interneurons. Downregulation of xCNGA1 or overexpression of a mutant xCNGA1 incapable of binding cGMP abolished CNG currents and converted growth cone repulsion to attraction in response to Sema3A. We also show that Ca2+ entry through xCNGCs is required to mediate the repulsive Sema3A signal. These studies extend our knowledge of the function of CNGCs by demonstrating their requirement for signal transduction in growth cone guidance.     

Learning in human neural networks on microelectrode arrays

This paper describes experiments involving the growth of human neural networks of stem cells on a MEA (microelectrode array) support. The microelectrode arrays (MEAs) are constituted by a glass support in which a set of tungsten electrodes are inserted. The artificial neural network (ANN) paradigm was used by stimulating the neurons in parallel with digital patterns distributed on eight channels, then by analyzing a parallel multichannel output. In particular, the microelectrodes were connected following two different architectures, one inspired by the Kohonen's SOM, the other by the Hopfield network. The output signals have been analyzed in order to evaluate the possibility of organized reactions by the natural neurons.f The results show that the network of human neurons reacts selectively to the subministered digital signals, i.e., it produces similar output signals referred to identical or similar patterns, and clearly differentiates the outputs coming from different stimulations. Analyses performed with a special artificial neural network called ITSOM show the possibility to codify the neural responses to different patterns, thus to interpret the signals coming from the network of biological neurons, assigning a code to each output. It is straightforward to verify that identical codes are generated by the neural reactions to similar patterns. Further experiments are to be designed that improve the hybrid neural networks' capabilities and to test the possibility of utilizing the organized answers of the neurons in several ways.     

Differential Imaging of Biological Structures with Doubly-resonant Coherent Anti-stokes Raman Scattering (CARS)

Coherent Raman imaging techniques have seen a dramatic increase in activity over the past decade due to their promise to enable label-free optical imaging with high molecular specificity ¹. The sensitivity of these techniques, however, is many orders of magnitude weaker than fluorescence, requiring milli-molar molecular concentrations ^1,2. Here, we describe a technique that can enable the detection of weak or low concentrations of Raman-active molecules by amplifying their signal with that obtained from strong or abundant Raman scatterers. The interaction of short pulsed lasers in a biological sample generates a variety of coherent Raman scattering signals, each of which carry unique chemical information about the sample. Typically, only one of these signals, e.g. Coherent Anti-stokes Raman scattering (CARS), is used to generate an image while the others are discarded. However, when these other signals, including 3-color CARS and four-wave mixing (FWM), are collected and compared to the CARS signal, otherwise difficult to detect information can be extracted ³. For example, doubly-resonant CARS (DR-CARS) is the result of the constructive interference between two resonant signals ⁴. We demonstrate how tuning of the three lasers required to produce DR-CARS signals to the 2845 cm^-1 CH stretch vibration in lipids and the 2120 cm^-1 CD stretching vibration of a deuterated molecule (e.g. deuterated sugars, fatty acids, etc.) can be utilized to probe both Raman resonances simultaneously. Under these conditions, in addition to CARS signals from each resonance, a combined DR-CARS signal probing both is also generated. We demonstrate how detecting the difference between the DR-CARS signal and the amplifying signal from an abundant molecule''s vibration can be used to enhance the sensitivity for the weaker signal. We further demonstrate that this approach even extends to applications where both signals are generated from different molecules, such that e.g. using the strong Raman signal of a solvent can enhance the weak Raman signal of a dilute solute.     

Linear Population Allocation by Bistable Switches in Response to Transient Stimulation

Many cellular decision processes, including proliferation, differentiation, and phenotypic switching, are controlled by bistable signaling networks. In response to transient or intermediate input signals, these networks allocate a population fraction to each of two distinct states (e.g. OFF and ON). While extensive studies have been carried out to analyze various bistable networks, they are primarily focused on responses of bistable networks to sustained input signals. In this work, we investigate the response characteristics of bistable networks to transient signals, using both theoretical analysis and numerical simulation. We find that bistable systems exhibit a common property: for input signals with short durations, the fraction of switching cells increases linearly with the signal duration, allowing the population to integrate transient signals to tune its response. We propose that this allocation algorithm can be an optimal response strategy for certain cellular decisions in which excessive switching results in lower population fitness.     

Recording Large Extracellular Spikes in Microchannels along Many Axonal Sites from Individual Neurons

The numerous connections between neuronal cell bodies, made by their dendrites and axons, are vital for information processing in the brain. While dendrites and synapses have been extensively studied, axons have remained elusive to a large extent. We present a novel platform to study axonal physiology and information processing based on combining an 11,011-electrode high-density complementary metal-oxide semiconductor microelectrode array with a poly(dimethylsiloxane) channel device, which isolates axons from somas and, importantly, significantly amplifies recorded axonal signals. The combination of the microelectrode array with recording and stimulation capability with the microfluidic isolation channels permitted us to study axonal signal behavior at great detail. The device, featuring two culture chambers with over 30 channels spanning in between, enabled long-term recording of single spikes from isolated axons with signal amplitudes of 100 μV up to 2 mV. Propagating signals along axons could be recorded with 10 to 50 electrodes per channel. We (i) describe the performance and capabilities of our device for axonal electrophysiology, and (ii) present novel data on axonal signals facilitated by the device. Spontaneous action potentials with characteristic shapes propagated from somas along axons between the two compartments, and these unique shapes could be used to identify individual axons within channels that contained many axonal branches. Stimulation through the electrode array facilitated the identification of somas and their respective axons, enabling interfacing with different compartments of a single cell. Complex spike shapes observed in channels were traced back to single cells, and we show that more complicated spike shapes originate from a linear superposition of multiple axonal signals rather than signal distortion by the channels.     

Self-selected modular recurrent neural networks with postural and inertial subnetworks applied to complex movements

 It has been shown that dynamic recurrent neural networks are successful in identifying the complex mapping relationship between full-wave-rectified electromyographic (EMG) signals and limb trajectories during complex movements. These connectionist models include two types of adaptive parameters: the interconnection weights between the units and the time constants associated to each neuron-like unit; they are governed by continuous-time equations. Due to their internal structure, these models are particularly appropriate to solve dynamical tasks (with time-varying input and output signals). We show in this paper that the introduction of a modular organization dedicated to different aspects of the dynamical mapping including privileged communication channels can refine the architecture of these recurrent networks. We first divide the initial individual network into two communicating subnetworks. These two modules receive the same EMG signals as input but are involved in different identification tasks related to position and acceleration. We then show that the introduction of an artificial distance in the model (using a Gaussian modulation factor of weights) induces a reduced modular architecture based on a self-elimination of null synaptic weights. Moreover, this self-selected reduced model based on two subnetworks performs the identification task better than the original single network while using fewer free parameters (better learning curve and better identification quality). We also show that this modular network exhibits several features that can be considered as biologically plausible after the learning process: self-selection of a specific inhibitory communicating path between both subnetworks after the learning process, appearance of tonic and phasic neurons, and coherent distribution of the values of the time constants within each subnetwork. Received: 17 September 2001 / Accepted in revised form: 15 January 2002     

The potential for signal integration and processing in interacting MAP kinase cascades

The cellular response to environmental stimuli requires biochemical information processing through which sensory inputs and cellular status are integrated and translated into appropriate responses by way of interacting networks of enzymes. One such network, the mitogen-activated protein (MAP) kinase cascade is a highly conserved signal transduction module that propagates signals from cell surface receptors to various cytosolic and nuclear targets by way of a phosphorylation cascade. We have investigated the potential for signal processing within a network of interacting feed-forward kinase cascades typified by the MAP kinase cascade. A genetic algorithm was used to search for sets of kinetic parameters demonstrating representative key input-output patterns of interest. We discuss two of the networks identified in our study, one implementing the exclusive-or function (XOR) and another implementing what we refer to as an in-band detector (IBD) or two-sided threshold. These examples confirm the potential for logic and amplitude-dependent signal processing in interacting MAP kinase cascades demonstrating limited cross-talk. Specifically, the XOR function allows the network to respond to either one, but not both signals simultaneously, while the IBD permits the network to respond exclusively to signals within a given range of strength, and to suppress signals below as well as above this range. The solution to the XOR problem is interesting in that it requires only two interacting pathways, crosstalk at only one layer, and no feedback or explicit inhibition. These types of responses are not only biologically relevant but constitute signal processing modules that can be combined to create other logical functions and that, in contrast to amplification, cannot be achieved with a single cascade or with two non-interacting cascades. Our computational results revealed surprising similarities between experimental data describing the JNK/MKK4/MKK7 pathway and the solution for the IBD that evolved from the genetic algorithm. The evolved IBD not only exhibited the required non-monotonic signal strength-response, but also demonstrated transient and sustained responses that properly reflected the input signal strength, dependence on both of the MAPKKs for signaling, phosphorylation site preferences by each of the MAPKKs, and both activation and inhibition resulting from the overexpression of one of the MAPKKs.     

Recovering motifs from biased genomes: application of signal correction

A significant problem in biological motif analysis arises when the background symbol distribution is biased (e.g. high/low GC content in the case of DNA sequences). This can lead to overestimation of the amount of information encoded in a motif. A motif can be depicted as a signal using information theory (IT). We apply two concepts from IT, distortion and patterned interference (a type of noise), to model genomic and codon bias respectively. This modeling approach allows us to correct a raw signal to recover signals that are weakened by compositional bias. The corrected signal is more likely to be discriminated from a biased background by a macromolecule. We apply this correction technique to recover ribosome-binding site (RBS) signals from available sequenced and annotated prokaryotic genomes having diverse compositional biases. We observed that linear correction was sufficient for recovering signals even at the extremes of these biases. Further comparative genomics studies were made possible upon correction of these signals. We find that the average Euclidian distance between RBS signal frequency matrices of different genomes can be significantly reduced by using the correction technique. Within this reduced average distance, we can find examples of class-specific RBS signals. Our results have implications for motif-based prediction, particularly with regards to the estimation of reliable inter-genomic model parameters.     

Brain connectivity at different time-scales measured with EEG

We present an overview of different methods for decomposing a multichannel spontaneous electroencephalogram (EEG) into sets of temporal patterns and topographic distributions. All of the methods presented here consider the scalp electric field as the basic analysis entity in space. In time, the resolution of the methods is between milliseconds (time-domain analysis), subseconds (time- and frequency-domain analysis) and seconds (frequency-domain analysis). For any of these methods, we show that large parts of the data can be explained by a small number of topographic distributions. Physically, this implies that the brain regions that generated one of those topographies must have been active with a common phase. If several brain regions are producing EEG signals at the same time and frequency, they have a strong tendency to do this in a synchronized mode. This view is illustrated by several examples (including combined EEG and functional magnetic resonance imaging (fMRI)) and a selective review of the literature. The findings are discussed in terms of short-lasting binding between different brain regions through synchronized oscillations, which could constitute a mechanism to form transient, functional neurocognitive networks.     

Simulated Evolution of Signal Transduction Networks

Signal transduction is the process of routing information inside cells when receiving stimuli from their environment that modulate the behavior and function. In such biological processes, the receptors, after receiving the corresponding signals, activate a number of biomolecules which eventually transduce the signal to the nucleus. The main objective of our work is to develop a theoretical approach which will help to better understand the behavior of signal transduction networks due to changes in kinetic parameters and network topology. By using an evolutionary algorithm, we designed a mathematical model which performs basic signaling tasks similar to the signaling process of living cells. We use a simple dynamical model of signaling networks of interacting proteins and their complexes. We study the evolution of signaling networks described by mass-action kinetics. The fitness of the networks is determined by the number of signals detected out of a series of signals with varying strength. The mutations include changes in the reaction rate and network topology. We found that stronger interactions and addition of new nodes lead to improved evolved responses. The strength of the signal does not play any role in determining the response type. This model will help to understand the dynamic behavior of the proteins involved in signaling pathways. It will also help to understand the robustness of the kinetics of the output response upon changes in the rate of reactions and the topology of the network.     

Intrinsic fluctuations, robustness, and tunability in signaling cycles

Covalent modification cycles (e.g., phosphorylation-dephosphorylation) underlie most cellular signaling and control processes. Low molecular copy number, arising from compartmental segregation and slow diffusion between compartments, potentially renders these cycles vulnerable to intrinsic chemical fluctuations. How can a cell operate reliably in the presence of this inherent stochasticity? How do changes in extrinsic parameters lead to variability of response? Can cells exploit these parameters to tune cycles to different ranges of stimuli? We study the dynamics of an isolated phosphorylation cycle. Our model shows that the cycle transmits information reliably if it is tuned to an optimal parameter range, despite intrinsic fluctuations and even for small input signal amplitudes. At the same time, the cycle is sensitive to changes in the concentration and activity of kinases and phosphatases. This sensitivity can lead to significant cell-to-cell response variability. It also provides a mechanism to tune the cycle to transmit signals in various amplitude ranges. Our results show that signaling cycles possess a surprising combination of robustness and tunability. This combination makes them ubiquitous in eukaryotic signaling, optimizing signaling in the presence of fluctuations using their inherent flexibility. On the other hand, cycles tuned to suppress intrinsic fluctuations can be vulnerable to changes in the number and activity of kinases and phosphatases. Such trade-offs in robustness to intrinsic and extrinsic fluctuations can influence the evolution of signaling cascades, making them the weakest links in cellular circuits.     

Shape-dependent control of cell growth, differentiation, and apoptosis: switching between attractors in cell regulatory networks

Development of characteristic tissue patterns requires that individual cells be switched locally between different phenotypes or "fates;" while one cell may proliferate, its neighbors may differentiate or die. Recent studies have revealed that local switching between these different gene programs is controlled through interplay between soluble growth factors, insoluble extracellular matrix molecules, and mechanical forces which produce cell shape distortion. Although the precise molecular basis remains unknown, shape-dependent control of cell growth and function appears to be mediated by tension-dependent changes in the actin cytoskeleton. However, the question remains: how can a generalized physical stimulus, such as cell distortion, activate the same set of genes and signaling proteins that are triggered by molecules which bind to specific cell surface receptors. In this article, we use computer simulations based on dynamic Boolean networks to show that the different cell fates that a particular cell can exhibit may represent a preprogrammed set of common end programs or "attractors" which self-organize within the cell's regulatory networks. In this type of dynamic network model of information processing, generalized stimuli (e.g., mechanical forces) and specific molecular cues elicit signals which follow different trajectories, but eventually converge onto one of a small set of common end programs (growth, quiescence, differentiation, apoptosis, etc.). In other words, if cells use this type of information processing system, then control of cell function would involve selection of preexisting (latent) behavioral modes of the cell, rather than instruction by specific binding molecules. Importantly, the results of the computer simulation closely mimic experimental data obtained with living endothelial cells. The major implication of this finding is that current methods used for analysis of cell function that rely on characterization of linear signaling pathways or clusters of genes with common activity profiles may overlook the most critical features of cellular information processing which normally determine how signal specificity is established and maintained in living cells.     

Simulation study on effects of signaling network structure on the developmental increase in complexity

The developmental increase in structural complexity in multicellular lifeforms depends on local, often non-periodic differences in gene expression. These, in turn, depend on a network of gene-gene interactions coded within the organismal genome. To see what architectural features of a network (size, connectivity, etc.) affect the likelihood of patterns with multiple cell types (i.e. patterns where cells express > or = 3 different combinations of genes), developmental pattern formation was simulated in virtual blastoderm embryos with small artificial genomes. Several basic properties of these genomic signaling networks, such as the number of genes, the distributions of positive (inductive) and negative (repressive) interactions, and the strengths of gene-gene interactions were tested. The results show that the frequencies of complex and/or stable patterns depended not only on the existence of negative interactions, but also on the distribution of regulatory interactions: for example, coregulation of signals and their intracellular effectors increased the likelihood of pattern formation compared to differential regulation of signaling pathway components. Interestingly, neither quantitative differences in strengths of signaling interactions nor multiple response thresholds to different levels of signal concentration (as in morphogen gradients) were essential for formation of multiple, spatially unique "cell types". However, those combinations of architectural features that greatly increased the likelihood for pattern complexity tended to decrease the likelihoods for pattern stability and developmental robustness. Nevertheless, elements of complex patterns (e.g. genes, cell type order within the pattern) could differ in their developmental robustness, which may be important for the evolution of complexity. The results show that depending on the network structure, the same set of genes can produce patterns of different complexity, robustness and stability. Because of this, the evolution of metazoan complexity with a combinatorial code of gene regulation may have depended at least as much on selection for favorable distribution of connections between existing developmental regulatory genes as on the simple increase in numbers of regulatory genes.     

Cross-talk and information transfer in mammalian and bacterial signaling

In mammalian and bacterial cells simple phosphorylation circuits play an important role in signaling. Bacteria have hundreds of two-component signaling systems that involve phosphotransfer between a receptor and a response regulator. In mammalian cells a similar pathway is the TGF-beta pathway, where extracellular TGF-beta ligands activate cell surface receptors that phosphorylate Smad proteins, which in turn activate many genes. In TGF-beta signaling the multiplicity of ligands begs the question as to whether cells can distinguish signals coming from different ligands, but transduced through a small set of Smads. Here we use information theory with stochastic simulations of networks to address this question. We find that when signals are transduced through only one Smad, the cell cannot distinguish between different levels of the external ligands. Increasing the number of Smads from one to two significantly improves information transmission as well as the ability to discriminate between ligands. Surprisingly, both total information transmitted and the capacity to discriminate between ligands are quite insensitive to high levels of cross-talk between the two Smads. Robustness against cross-talk requires that the average amplitude of the signals are large. We find that smaller systems, as exemplified by some two-component systems in bacteria, are significantly much less robust against cross-talk. For such system sizes phosphotransfer is also less robust against cross-talk than phosphorylation. This suggests that mammalian signal transduction can tolerate a high amount of cross-talk without degrading information content. This may have played a role in the evolution of new functionalities from small mutations in signaling pathways, allowed for the development of cross-regulation and led to increased overall robustness due to redundancy in signaling pathways. On the other hand the lack of cross-regulation observed in many bacterial two-component systems may partly be due to the loss of information content due to cross-talk.     

Evolvability and robustness in a complex signalling circuit

Biological systems at various levels of organisation exhibit robustness, as well as phenotypic variability or evolvability, the ability to evolve novel phenotypes. We still know very little about the relationship between robustness and phenotypic variability at levels of organisation beyond individual macromolecules, and especially for signalling circuits. Here, we examine multiple alternate topologies of the Saccharomyces cerevisiae target-of-rapamycin (TOR) signalling circuit, in order to understand the circuit's robustness and phenotypic variability. We consider each of the topological variants a genotype, a set of alternative interactions between TOR circuit components. Two genotypes are neighbours in genotype space if they can be reached from each other by a single small genetic change. Each genotype (topology) has a signalling phenotype, which we define via the concentration trajectories of key signalling molecules. We find that the circuits we study can produce almost 300 different phenotypes. The number of genotypes with a given phenotype varies very widely among these phenotypes. Some phenotypes have few associated genotypes. Others have many genotypes that form genotype networks extending far through genotype space. A minority of phenotypes accounts for the vast majority of genotypes. Importantly, we find that these phenotypes tend to have large genotype networks, greater robustness and a greater ability to produce novel phenotypes. Thus, over a broad range of phenotypic robustness, robustness facilitates phenotypic variability in our study system. Our observations show parallels to studies on macromolecules, suggesting that similar principles might govern robustness and phenotypic variability in biological systems. Our approach points a way towards mapping genotype spaces in complex circuitry, and it exposes some challenges such mapping faces.     

Multimodal signalling in estrildid finches: song,dance and colour are associated with different ecological and life‐history traits

Sexual traits (e.g. visual ornaments, acoustic signals, courtship behaviour) are often displayed together as multimodal signals. Some hypotheses predict joint evolution of different sexual signals (e.g. to increase the efficiency of communication) or that different signals trade off with each other (e.g. due to limited resources). Alternatively, multiple signals may evolve independently for different functions, or to communicate different information (multiple message hypothesis). We evaluated these hypotheses with a comparative study in the family Estrildidae, one of the largest songbird radiations, and one that includes many model species for research in sexual selection and communication. We found little evidence for either joint evolution or trade‐offs between song and colour ornamentation. Some negative correlations between dance repertoire and song traits may suggest a functional compromise, but generally courtship dance also evolved independently from other signals. Instead of correlated evolution, we found that song, dance and colour are each related to different socio‐ecological traits. Song complexity evolved together with ecological generalism, song performance with investment in reproduction, dance with commonness and habitat type, whereas colour ornamentation was shown previously to correlate mostly with gregariousness. We conclude that multimodal signals evolve in response to various socio‐ecological traits, suggesting the accumulation of distinct signalling functions.