Information & Decisions Across Scales: Constraints and Optimality in Neural and Biological Networks

The cell comprises millions of biomolecules that are interacting and reacting in a crowded millieu. There is growing recognition that cooperative self-organization by phase transitions contributes to spatial organization of molecules into organelles called condensates. In my talk, I will discuss how the collective processes underlying condensation can enable decision making by selectively assembling condensates at particular cellular surfaces - performing a formation of surface classification. I will elaborate on the capabilities and constraints that emerge in this form of computation.

Cells assemble complex molecules such as proteins using template polymers with information-carrying sequences. These sequences are used to direct the ordered assembly of a set of molecular building blocks to form exactly the right product, such as a specific protein, on demand. A crucial feature of this process is that the templates act catalytically, since they accelerate the assembly of a specific product without being consumed in the process. A single template can then be used to direct the assembly of many copies of a product. It has long been suspected that this transfer of information from catalytic templates to products must be associated with a minimal energy cost, in the same way that computation has lower bounds on energy input. To explore this hypothesis further, we consider arbitrarily complex templating networks involving many assembly and disassembly processes. We show that the accuracy with which a specific set of products can be maintained is, indeed, constrained by energetic properties of the underlying network. Surprisingly, the operation of the most efficient templating networks is completely different from the behavior observed in nature, where molecules such as proteins are overwhelmingly assembled by templates and degraded in a distinct, template-free fashion. Instead, the most efficient approach, which both achieves the highest accuracy and has a negligible energetic cost, is to disassemble products via a reversal of the pathway by which they were formed. This observation raises interesting questions about why nature does not operate in this way and whether synthetic templating systems can be built that exploit this approach.

A distinctive feature of many biological systems is their ability to adapt to persistent stimuli or disturbances that would otherwise drive them away from a desirable steady state. This resulting stasis enables reliable function across a wide range of external environments. We focus on a stringent form of this behavior—robust perfect adaptation (RPA)—which remains resilient to certain network and parameter perturbations. As in engineered control systems, RPA is not incidental: it requires the regulating network to satisfy specific, unavoidable structural constraints. Using examples from systems biology and synthetic biology, we show how these constraints arise in natural and engineered circuits. We argue that identifying the structural basis of RPA allows us to move beyond implementation details and provides a principled lens for understanding regulatory complexity and information processing in biological systems.

Many cellular components are present in such low numbers that individual stochastic production and degradation events can lead to significant fluctuations in biomolecular abundances. Although feedback control can, in principle, suppress such low-copy-number fluctuations, general rules have emerged that suggest fundamental performance constraints on feedback control in biochemical systems. In this talk, I will discuss how fluctuation control through regulated molecular synthesis is limited by stationary information flow. I will then show how this general fluctuation bound implies a kinetic uncertainty relation, whereby reducing one component’s fluctuations through feedback control unavoidably leads to increased fluctuations for another. This result suggests that fluctuating cellular components do not necessarily generate "noise" but may correspond to control species that are involved in removing "noise" from another component’s dynamics.

Living cells are capable of responding to environmental cues under noisy conditions, even in the absence of a centralised control unit or brain. Such systems therefore pose an ideal test-space in which to uncover the physical principles behind robust, decentrailised computations in soft materials. Motivated by the role interfaces play in relaying signals across cell boundaries, we investigate how spatial patterns on interfaces facilitate subcellular information transmission. Starting from a statistical description of particle dynamics, we show that these patterns act as signal filters that non-linearly amplify heterogeneties in their environment. This mechanism permits a form of pattern recognition, with interparticle interactions tuning the response function of the filter. We explicitly identify both thresholding and edge-detecting regimes and, accounting for thermal noise in the filters, quantify the flow of information across the interface. We find that, when suitably tuned, the noisy filters selectively compress input signals, resulting in the efficient encoding of information relevant to downstream tasks. Overall, our results indicate that the noisy patterning of membranes may be used to selectively sense environmental cues in biologically inspired systems, suggesting exciting implications for how physical interactions may encode computational logic in soft active materials.

In physical learning, information processing (such as the forward pass of a deep neural network) is implemented through the dynamics of a physical system. Here, we consider Competitive Dimerization Networks (CDNs) in which molecules belonging to multiple species dimerize following mass action laws. This system functions as a neural network in which the neuronal variables and synaptic weights are represented, respectively, by the monomer concentrations and the affinities controlling the propensity to dimerize. Trained CDNs can sort between different noisy patterns, encoded into the concentration of input species, by overexpressing the concentration of specific predefined output species. We compare different learning algorithms, including direct evolutionary learning, gradient descent, and contrastive learning methods. We highlight the important role played by the cost function to achieve a large contrast between the distributions of activated and suppressed output species. Furthermore, we extend our analysis to non-equilibrium open systems.

Collective behavior in animals provides a compelling example of self-organization in biological systems. Such behavior, together with the underlying social interactions, is thought to have emerged through evolutionary adaptation and is widely assumed to confer fitness benefits to individuals—for instance by enabling the exchange of social information, improving the accuracy of collective decisions, or providing protection from predators. In this context, it has been proposed that animal collectives, much like other distributed biological information-processing systems such as the brain, operate near critical points—special regions of parameter space where collective computation and responsiveness are optimized. Here, we investigate this criticality hypothesis in the context of biological multi-agent systems, focusing on collective escape responses in fish. To this end, we combine mathematical modeling with laboratory and field experimental data to assess whether and how critical dynamics shape collective behavior.

We address the problem of inferring the location of a target that releases odor in the presence of turbulence. Input for the inference is provided by many sensors scattered within the odor plume. Drawing inspiration from distributed chemosensation in biology, we ask whether the accuracy of the inference is affected by noise on the measure and the perceived location of the sensors. Surprisingly, in the presence of a net fluid flow, positional noise improves Bayesian inference, rather than degrading it. An optimal noise exists that efficiently leverages additional information hidden within the geometry of the odor plume. Empirical tuning of noise functions well across a range of distances and may be implemented in practice. Noise on the sensory process also improves accuracy, owing at least in part to its ability to break the spatiotemporal correlations of the turbulent plume. These counterintuitive benefits of noise may be leveraged to improve sensory processing in biology and robotics.

Time-series mutual information is an indispensable theoretical metric for quantifying information transmission and processing in biological systems and for characterizing their efficiency. However, its measurement requires time-resolved characterization and analysis of both input and output distributions, which poses a substantial constraint in experimental applications. In this study, we propose an alternative approach based on a dual-reporter system to alleviate the requirement of simultaneous input–output measurements. By extending the framework of extrinsic and intrinsic noise decomposition originally developed for gene expression, we derive an estimator of mutual information that eliminates the need to measure the input distribution. We apply this method to the bacterial chemotaxis signaling pathway and demonstrate its validity by regarding multiple flagellar motors as natural dual reporters. By comparing the measured information flow with theoretical upper bounds on sensory information, we reveal that the signaling network of Escherichia coli possesses sufficient transmission efficiency to convey the information acquired at the sensory level to the downstream output, namely the flagellar motors. This framework opens new directions for quantifying information flow in cellular signaling pathways.

Artemy Kolchinsky and Richard J.G. Löffler Living systems exhibit a remarkable capacity for goal-directed behavior, acting on their environments to achieve target outcomes. Similar behaviors are increasingly realized in simple nonequilibrium systems, such as active matter and nonequilibrium chemical systems. However, our theoretical understanding of goal-directedness in formal, quantitative terms remains limited. Here, we propose a framework for inferring and quantifying goal-directed behavior in many-body active systems. We fit a hierarchy of nested statistical models to trajectory data, where goal-directedness is defined as approximate gradient descent on an inferred objective. This yields an information-geometric decomposition of dynamical predictability into contributions from individual, collective, and non-reciprocal goals. We illustrate our approach on self-propelled droplets that exhibit rich emergent behaviors, including predator–prey dynamics.

Bacteria, and many other living organisms, navigate their environments collectively by exchanging information, both through direct interactions, and indirectly through their shared environment. While these flows of information are necessary to coordinate behavior in navigating collectives, we lack a principled framework to quantify and measure them. Inspired by stochastic thermodynamics, we have developed a framework to quantify the flow of information between collectively-navigating organisms. In this talk I will show how quantifying these information flows enable us to understand how different aspects of collective dynamics contribute to maintaining spatial structure in a traveling wave, and how information flows bound both the precision of spatial and phenotypic structure, and ultimately navigation performance.

Life hinges on information processing: to navigate the world, animals rely on their brains – and thus on the fine, collective interactions of millions or billions of neurons. Importantly, these networks can learn a model of the world in a self-organized manner, i.e., they update the connectivity and information flow between the neurons without any explicit teacher signal. We aim to understand the basic principles of this emergent information flow or “infogenesis” using the approaches from statistical physics and information theory. Understanding the emergence of infogenesis is very interesting per se; in addition, as evolution has optimized brains over >250M years, our studies can also inspire novel, robust, and energy-efficient compute principles for AI.

Physical learning machines, such as adaptive resistor and mechanical networks, learn desired functions through local physical processes instead of machine learning algorithms. Yet even in these simple systems, we lack a clear understanding of how learned tasks are encoded in the physical landscape and what ultimately determines their functional capacity and expressiveness. Here, we explore linear learning networks as a first analytically tractable setting to address these questions. These systems can adapt both their equilibrium state and the energy landscape around it, revealing distinct modes of physical learning that shape which tasks can be acquired and how they are encoded. Building on recent results linking learned functions to measurable structural changes, we show how these modes imprint signatures that allow us to deduce whether, and what, a physical system has learned. This framework lays essential groundwork for interpreting learning in more complex learning machines, both synthetic and living.

Cells are constantly tasked with making accurate measurements of their surroundings. A paradigmatic example is the sensing of signalling molecule concentrations: the work of Berg and Purcell derived limits for the precision and speed of this sensing through ligand-receptor binding. However, recent experimental work has identified the formation of condensates (liquid droplets coexisting with the cell cytoplasm through phase separation) as a potential mechanism for selectively initiating downstream processes by effectively amplifying small concentration differences between competing signalling molecules. Using a minimal model for droplet nucleation and growth in a fluid mixture, we observe that phase separation can distinguish concentration differences of 1% in minutes, a significant improvement upon well-established pathways for precise concentration sensing.

Understanding how living organisms process sensory information from their surroundings and translate it into decisions is a fundamental problem across biological scales – from biochemical signalling in single-cells to neural computations in animal brains. We address this challenge by introducing a method to reconstruct general decision processes directly from behavioral observations alone, applicable to any biological agent without prior knowledge of its internal mechanisms or its environment. We infer minimal, interpretable agent models from experimental data of rats performing evidence accumulation tasks and of mice making decisions under uncertainty and in changing environments.Our results show that this method can decode the computational structure underlying decision-making strategies from behavioral trajectories alone, providing a broadly applicable framework for understanding how general agents make decisions in complex environments.

During chemical navigation organisms need to continuously adapt their strategy to the changing signal statistics and context in which the navigation take place. What information to extract from signal and how to utilize it can change over time. I will report on experiments we conducted that examine how bacteria and fruit flies perform such adaptation.

Single-celled organisms live in a highly dynamic environment to which they continually have to respond and adapt. To this end, they must acquire information about their surroundings. These systems are indeed excellent information-processing machines, often operating close to fundamental limits as set by information theory. Yet, living systems must not only acquire information, they also must act on it, to enable function. Using cellular navigation as a paradigm for biological function, I will discuss how single-celled organisms not only acquire information, but also use it to generate function.

Living systems are characterized by their ability to react to, and affect, their complex environments. In this talk I will focus on the interplay between bacterial systems (in different forms) and their external world, discussing in particular recent results on spore germination and biofilm dispersion, using a variety of biophysical and biochemical theoretical descriptions, in combination with experiments at the cellular level.

Chemotactic navigation of biological cells represents a prime example of cellular information processing corrupted by noise, with known input and output, and a single objective. Intriguingly, cells of different size use different chemotaxis strategies, comparing concentrations of signalling molecules in either time or space: Small and fast bacteria sense concentration gradients in time while moving actively, whereas large and slow eukaryotic cells with crawling motility sense concentration gradients across their diameter in space. Only heuristic arguments exist to explain this evolutionary choice. To address this open question, we present an information theory of an ideal agent that combines both gradient-sensing strategies. By quantifying ‘chemotaxis in bits’, we can compare the information gain of temporal and spatial gradient-sensing. This theory of Bayesian chemotaxis generalizes information-greedy infotaxis to spatially extended agents with motility noise and an egocentric map. Using information decomposition, we can predict when each gradient-sensing strategy provides more information as function of an empirical powerlaw that combines agent size, motility noise and sensing noise into a single predictor. This predictor is consistent with data from 250 chemotactic cells, including unusual bacteria that use spatial gradient-sensing. Our idealized model of Bayesian chemotaxis assuming unlimited information processing capabilities thus serves as a benchmark for the chemotaxis of biological cells. Rode et al., PRX Life 2, 023012 (2024)

Navigating up a concentration gradient is one of the simplest behaviours, and organisms use a variety of different strategies: E. coli perform run-and-tumble motion, V. cholerae do run-reverse-flick, while some larger organisms implement continuous steering. What drives their choices? Here we explore the idea that limited sensory information might be the binding constraint, and show that qualitatively different strategies are optimal with different information rates. For an organism which responds to the instantaneous rate of change of local concentration, without directional information, we show that the optimal strategy is always some pattern of straight runs and sudden turns, never continuous steering. What kind of turn is preferred changes with the information rate (or gradient steepness), and the ideal response uses always a discrete set of turn angles, whose number changes increases when more information is available.

Biological and artificial systems make decisions from noisy signals under strict energetic constraints. I will discuss a trajectory-based framework that links information processing and dissipation in stochastic systems, showing how adaptive strategies can balance information gain and energetic cost under partial observability. This approach provides a concrete way to quantify how noise, limited observability, and energetic costs constrain information processing, from biochemical sensors to artificial learning systems.

Eukaryotic gene regulation is based on stochastic yet controlled promoter switching, during which genes transition between transcriptionally active and inactive states. Despite the molecular complexity of this process, recent studies reveal a surprising invariance of the "switching correlation time”, which characterizes promoter activity fluctuations. This quantity is constant across gene expression levels in diverse genes and organisms. A biophysically plausible explanation for this invariance remains missing. I will present recent work in collaboration with our experimental colleagues (Gregor lab), which provides a mechanistic and functional insights into the origin and meaning of the said invariance.

Cells in embryos need to decode extracellular signals to regulate their genes and eventually acquire the desired fate. In vertebrates, morphogen gradients typically control cell fate inductions, with cells expressing different genes depending on the morphogen concentration they detect. The role of the gradient can also be fulfilled by different contact areas between the cells and the signaling molecules. Cells exposed to the same concentration of signaling molecules can tune the level of signal they receive, and therefore acquire different fates, according to the surface area they expose to the morphogen. We developed minimal conceptual models to compare these two opposite frameworks for cell differentiation. We explored under what conditions, such as system size, cell number, and cell geometry, one mechanism is more efficient than the other in terms of information transmission.

Complex biological patterns emerge spontaneously in nature from branching of trees to synchronized motion of fish and birds. One remarkable example occurs during early embryonic development, when a network of blood vessels emerges from a group of mesodermal cells that cluster together and fuse to form an interconnected network. We explore the mechanical forces that regulate vascular network formation and the functional role of such networks during the early development of chick embryos. By integrating live and fixed imaging of vascularisation combined with topological data analysis and geometrical parameters, we track the transformation of individual mesodermal cell clusters into a fully connected vascular network. This network is thought to arise through the fusion of these clusters. We quantify their size, growth, and spatial dynamics, monitoring their progression over time to understand how surrounding tissues influence this behaviour. Through targeted drug perturbations we disrupt normal network formation, and the resulting changes will help us identify the physical and biological parameters critical for efficient vascular development. Through this approach our aim is to define the minimal set of rules or “ingredients” required to construct a functional transport network, both in embryonic systems, diseases such as cancer and in broader biological and engineered contexts.

The current paradigm in decision theory is based on the assumption that decisions, once taken, cannot be revoked. However, living systems often amend their decisions if accumulated evidence favors an alternative hypothesis. I will present a theory for amendable decision making in living systems. In contrast with irrevocable decisions, optimal amendable decisions can be taken in a finite average time at zero error probability. The theory for amendable decisions successfully predicts the outcome of a visual experiment involving human participants and the accuracy of cell-fate decisions in early development. These case studies demonstrate the broad applicability of the amendable decision paradigm in biology.

In a genetically heterogeneous population of single-celled organisms every cell performs an evolutionary experiment every generation. This information, about the mapping from genome to fitness is computationally infeasible to extract except through these natural experiments. In a multicellular organism, somatic cells perform such experiments, but in the usual picture of animal evolution the results of these experiments are discarded in future generations. In this talk I will argue that (1) information from these somatic experiments could be useful to the next generation, if an imperfect readout of organismal fitness and (2) that known biochemistry can be reinterpreted as providing a mechanism to transmit this information between somatic cells and from somatic cells to the germline.

Living systems utilize biochemical networks to transduce environmental inputs into internal modifications that can be read downstream to mount different responses. When the dimensionality of the intermediate driving signal is lower than that of the output, for example, if a single transcription factor affects the expression of multiple genes, it can be useful to exploit the information contained in the trajectory of such a signal. This is particularly relevant to development, where experiments have shown that different activity trajectories of transcription factors such as ERK can influence differentiation, effectively associating environmental signals with particular transcription factor dynamics. In this work, we introduce a minimal model of decision-making in multistable gene networks that exploits the history of a forcing signal and the structure of the phase space to reliably map different environmental conditions to end states of the gene network. The performance of such classification can be scored for an ensemble of environments using information-theoretical quantities. We will show that this metric can be efficiently optimized using an extended gradient estimation technique, finding sets of trajectories that partition the input space into different final states, from which the label of each input state can be reliably recovered.

The crossword-like patterns of tiles in Scrabble form connected graphs of occupied sites on a square lattice. We are interested in describing the ensemble of these Scrabble graphs and comparing them across different languages. To find the most structureless description of Scrabble graphs, we build a maximum-entropy probability distribution; using real tournament data, we adapt a pseudo-likelihood method to the case of connected graphs on a lattice. We find that a maximum-entropy distribution that includes means and pairwise correlations captures the data: it correctly predicts simultaneous square occupation, word-length statistics, and geometric features of the Scrabble graphs, as well as the hierarchy among square types. To explore how language affects the structure of the Scrabble graphs, we adapt a Scrabble bot to self-play and generate graphs using different lexica. We find that the graphs produced by the bot have lower entropy compared to human players, and that increasing the lexicon size reduces the entropy given the same tile set. Remarkably, the pairwise maximum-entropy distribution is almost sufficient to correctly assign Scrabble graphs to their corresponding language.

Foraging and feeding are essential behaviors for survival. The neural circuits underlying these behaviors are typically especially robust, and operate with a high level of redundancy to buffer against fatal effects at both short timescale changes such as during development as well as detrimental mutation and circuit changes occurring over the longer scales of evolution. I will discuss the enteric nervous system of nematodes as an example of such a system, which displays both hallmarks of degeneracy and redundancy, and simultaneously follows principles of evolvability. Specifically, we find that a weak synaptic linkage between somatic and enteric nervous systems allows for robust feeding behavior, yet this synaptic linkage is supplemented by extra-synaptic signaling, for example via neuromodulators such as serotonin and dopamine. We found that a model of neuromodulatory control in a Hopfield-like network allows for encoding of vastly more memories, suggesting that gating by neuromodulators may be advantageous for increasing brain state complexity. Interestingly, across different nematode species, variations in circuit organization and neuromodulator usage reflect divergent feeding strategies, such as grazing on bacterial lawns versus predatory behaviors. These findings demonstrate how weak synaptic connectivity, when paired with neuromodulatory plasticity, allows for both robust performance and evolutionary flexibility—key features of evolvable biological systems.

One of the hallmarks of adaptive matter is the ability to sense and respond to the environment. How can we engineer soft matter systems to exhibit such behavior? In this talk, I will present our recent findings on systems driven by centimeter-scale actuators (electric motors) and nanometer-scale actuators (myosin motors), using methods from control theory to either design function or probe system properties. Such control-theoretical approaches are often exploited in engineering applications. I will instead emphasize how control in active and robotic systems can help physicists better understand information processing, orchestration of function, and adaptability in living systems.

Economic choice involves comparing the subjective values of goods, a process linked to the orbitofrontal cortex (OFC). Neurophysiological studies have identified OFC neurons encoding variables that capture both the inputs and outputs of this process, such as the offer values and the chosen goods. However, the precise nature of these representations and the circuit mechanisms that give rise to choice remain open questions. To investigate these mechanisms, we combined large-scale neuronal recordings, network inference analyses, and theoretical modeling of circuit dynamics. We used simultaneous recordings of 50–150 neurons while monkeys performed a binary juice choice task and inferred Ising (Hopfield) models using the Adaptive Cluster Expansion algorithm. As an initial step, we characterized each neuron's tuning by identifying the variable it most strongly encoded. We then simulated the inferred network models under different offer inputs, and analyzed their stationary points. Simulations revealed attractor-like avalanche dynamics: stimulating offer value cells associated with one particular good (A or B) induced avalanches that led to the co-activation of other cell groups. Co-activated cells predominantly encoded the corresponding choice outcome (A or B). Compared to random networks, inferred OFC networks showed stronger avalanches. Ablation analyses indicated that direct connections between offer-value and chosen-good cells were crucial for generating correct decision-related avalanches, while other connections played a secondary, reinforcing role. These results suggest that OFC functions as an attractor network transforming offer values into binary choices.

Understanding how neural architecture constrains information processing is fundamental to explaining adaptive computation in the brain. While stimulus-evoked trials reveal how brain areas co-activate during decision-making, such approaches are labour-intensive and can be limited by the choice of stimuli. Prior work shows that scale-free spontaneous cortical activity reflects underlying structural connectivity, suggesting that information flow during rest may reveal computational constraints inherent to network architecture. Here, we capitalize on recent advances in high spatio-temporal resolution recording modalities to trace neural activity across the mouse cortical surface to determine effective connectivity from spontaneous activity. Using wide-field calcium imaging, we demonstrate that this effective connectome during quiet wakefulness correlates well, though imperfectly, with structural connectivity, revealing how information can propagate through multiple pathways despite anatomical constraints. Conversely, anesthesia dramatically decreases this correlation, indicating that altered brain states fundamentally change how structure constrains information flow. These findings suggest that effective information processing requires a balance between structural constraints and dynamic flexibility—a principle that may extend to understanding computational efficiency in both biological and artificial systems.

Sensory prediction is thought to be vital to organisms, but few studies have tested how well organisms and parts of organisms efficiently predict their sensory input in an information-theoretic sense. In this talk, we report results on how well cultured neurons ("brain in a dish") and humans efficiently predict artificial stimuli. We find that both are efficient predictors of their artificial input. That leads to the question of why, and to answer this, we study artificial neural networks, finding that LSTMs show similarly efficient prediction but do not model how humans learn well. Instead, it appears that an existing model of humans as order-R Markov modelers explains their performance on these prediction tasks.