Shiling Liang: MPI-PKS
We uncover fundamental physical constraints on information processing in nonequilibrium networks through rigorous analysis of response theory in Markov systems. Our work establishes two key principles: a universal precision bound of 1/2 for state observable responses that cannot be overcome by increasing network complexity, and exact mathematical relations describing system responses to strong perturbations. These results provide physical insights into why simple network architectures - as illustrated with a push-pull motif - can be optimal despite evolutionary pressure for precision. By bridging statistical physics and information theory, our framework reveals how thermodynamic constraints shape the design of nonequilibrium networks, with implications for biological sensing and information processing systems.
Tetsuhiro S. Hatakeyama: Earth-Life Science Institute, Institute of Science Tokyo
Monod's law is a widely accepted phenomenology for bacterial growth. Since it has the same functional form as the Michaelis--Menten equation for enzyme kinetics, cell growth is often considered to be locally constrained by a single reaction. In contrast, we shows that a global constraint principle of resource allocation to metabolic processes can well describe the nature of cell growth. This concept is a generalization of Liebig's law, a growth law for higher organisms, and explains the dependence of microbial growth on the availability of multiple nutrients, in contrast to Monod's law.
Henrique Oyama: Okinawa Institute of Science and Technology (OIST)
The phenomenon of mind wandering, characterized by a spontaneous drift away from task-focused thought, reflects a dynamic interplay between focused attention and off-task mental states. Various studies have investigated the psychological and systematic mechanisms underlying these shifts. However, previous models have not yet provided an account for the underlying neural mechanisms for autonomous shifts between the two states.
Recent works investigated mind-wandering mechanisms using the Predictive Variational Recurrent Neural Network (PV-RNN), a hierarchically organized model rooted in the Free Energy Principle (FEP). The PV-RNN’s dynamic behavior is governed by a meta-level parameter, the meta-prior w, which balances the complexity term against the accuracy term in free energy minimization. While these studies provide critical insights into macroscopic neural mechanisms, the transition from FS to MW was generated by manual resetting of the meta-prior from a low to a high setting, leaving the mechanism for autonomous FS-MW shifts unexplored.
In light of the above, we propose an adaptive mechanism for w, wherein w is modulated depending on the reconstruction error accumulated over a past window period. A simulation experiment is presented to showcase the proposed model. In particular, using PV-RNN, we trained the model to predict sensory patterns generated by probabilistic transitions among multiple cyclic patterns. Simulation results demonstrate that autonomous shifts between FS and MW emerged as w switched dynamically: high w enhanced top-down predictions, promoting MW, while low w emphasized bottom-up sensory perception, favoring FS. Finally, we speculate that agents could become consciously aware of MW when the accumulated error exceeds a certain threshold.
Pranshu Malik: Okinawa Institute of Science and Technology (OIST)
Contexts and intentions continually shape our movement goals, often requiring us to replan and adjust our actions in real time. While it is well-established that stretch reflexes are tuned to these goals, the timescale at which they update during ongoing movement remains unclear. Our investigation shows that long-latency feedback responses can update almost as rapidly as visuomotor response times and that the motor periphery can also be tuned independently and simultaneously for the new movement goal. These findings not only offer valuable insights into the close coupling between reflexive and voluntary control during dynamic motor tasks but also, through nuanced observations, suggest a resource- and time-efficient strategy employed in the computation of corrective feedback responses — a promising area for further exploration — which will be discussed as well.
Olga Bagrova: Okinawa Institute of Science and Technology (OIST)
Proteins play a central role in biological systems due to their diverse functions. Elucidating these functions is an important task for drug development and disease research, but it often requires extensive experimental work. Despite the vast availability of protein sequences and 3D structural data, these resources do not always enable the identification of functions for novel proteins or the understanding of the mechanisms of known proteins. Given that a protein’s function is determined by its structure, identifying any fundamental structural patterns related to a specific protein function could bridge the gap between available data and our comprehension of these biomolecules.
In this study, we analyzed the distribution of secondary structures along the polypeptide chains of various protein groups distinguished by function and homology to reveal the relationship between protein structure and function and to explore the evolution of this relationship. We developed a method for encoding structural information into tables to provide the analysis of spatial structure distributions for potential property prediction.
Our results revealed distinct structural patterns characteristic of specific protein groups. For example, we identified regions of frequent occurrence of alpha-helices and irregular structures in DNA-binding proteins and microtubule-associated proteins. Additionally, an overall predominance of irregular structures was observed at the termini of polypeptide chains across all protein groups. By comparing secondary structure distributions among homologous protein groups with similar folds but differing functions and origins, we discovered similarities in minor secondary structural elements, specifically three-ten-helices and irregular structures. This observation may indicate potential divergent evolutionary pathways for protein groups such as globins and phycocyanins.
It was expected that these results could contribute to solving the problem of annotating and predicting the functions of unknown polypeptides.
Alexandru Mihai: Okinawa Institute of Science and Technology (OIST)
"Automating fish image classification presents new opportunities for developmental biology and species identification, particularly in data-heavy fields like ichthyology and fisheries science. This study applies unsupervised machine learning to streamline the categorization of tropical fish species, reducing dependence on manual annotation while improving classification accuracy.
Using pre-trained neural networks (VGG16, ResNet, EfficientNet) for feature extraction and foreground detection, we implemented clustering techniques to identify developmental stages and differentiate species. UMAP proved particularly effective in preserving high-dimensional feature relationships, while HDBSCAN revealed previously uncategorized developmental groupings. The model classified A. ocellaris developmental stages with 87% accuracy and successfully distinguished A. ocellaris from A. percula based on image embeddings.
These results highlight the potential of unsupervised learning in fishery management, offering a scalable, data-driven approach for species identification and developmental classification. Ongoing work will extend this methodology to hybrid taxa and broader datasets, refining its applicability in evolutionary and ecological studies."
Aarón Cid Ramírez: ECSU, OIST / RC51, ISA / DCNI, UAM
In recent months, the integration of artificial intelligence (AI) tools into various scientific disciplines has become inevitable. This is particularly evident in their application within recent Nobel Prize-winning projects (Chemistry and Medicine) and their increasing role in everyday scientific workflows.
One of the greatest advantages of these AI tools is their continuous development by multidisciplinary teams, often with open-access frameworks, ensuring broad accessibility. However, keeping track of the most recently developed tools with significant potential impact remains challenging.
This poster aims to analyze the state of the art of AI tools in molecular biology and their role in transforming biological data analysis. By understanding their fundamental principles, these tools could be integrated into current and future experiments or used to replicate past studies more efficiently.
Key AI-based tools such as AlphaFold, RoseTTAFold, DeepCRISPR, CellProfiler, DeepCell, Scikit-learn, TensorFlow/PyTorch, ESMFold, DeepMind's AlphaMissense, Scanpy, IntegrOmics, and OpenMM will be presented, highlighting their applications in structural biology, gene editing, image analysis, and omics data integration.
Anzhelika Koldaeva: Okinawa Institute of Science and Technology (OIST)
Genes interact in complex, correlated ways. These relationships are crucial for understanding both an organism's function and its evolutionary relatedness. Traditional phylogenetic methods, which simplify gene evolution by assuming independence, often overlook key coevolutionary signals. As a result, phylogenetic trees can be noisy due to factors like horizontal gene transfer, incomplete lineage sorting, and rate variation, making accurate species tree inference challenging. We present a deep denoising framework based on a Set Transformer, which learns the complex interactions among clusters of orthologous groups (COGs). In our approach, each genome is represented as a fixed-length vector over a curated vocabulary of COGs, where each value corresponds to the count of a specific COG in the genome. We use a masked language modeling objective to train the model to capture coevolutionary patterns and interdependencies that classical methods ignore. This enables more accurate reconstruction of complete gene repertoires from partial data, with potential applications in improving ancestral phenotype inference from phylogenetic reconciliations and denoising metagenomic data.
Atsushi Kamimura: The University of Tokyo
Regulating cell growth and division is fundamental in cell physiology. Microfluidic devices enable precise single-cell measurements, revealing key physiological parameters like cell size. Constructing mathematical models is essential to elucidate key aspects of these mechanisms. However, this task is challenging, especially when ensuring consistency with noisy multidimensional data.
A neural network (NN) method, designed for temporal point processes, effectively separates deterministic factors from noise. Applied to bacteria and fission yeast, this method infers birth and division size distributions, successfully identifying known mechanisms such as the adder model. Our study demonstrates that, compared to conventional statistics, the NN method is more effective in uncovering hidden dynamic laws in complex data.
Kento Nakamura: RIKEN Center for Brain Science
Mutual information provides a theoretically grounded measure for quantifying information flow in biological signaling systems. However, experimental measurement is demanding because it requires complete distributions of both input and output variables. In this poster, we propose a method that eliminates the need for measuring input statistics by using dual reporter systems. We apply our method to bacterial chemotaxis where multiple flagellar motors serve as dual reporters, and establish the biological relevance of the estimated value by comparing it with our previously derived optimality theory of chemotactic information processing.
Takehiro Tottori: RIKEN
While the information processing and decision-making of organisms have been optimized through evolution, they exhibit significant variability among species. One potential source of this diversity is the limitation of resources available to organisms. However, we lack a theoretical framework for investigating biological information processing and decision-making under resource limitations. To address this issue, we propose a novel optimal estimation and control theory that explicitly accounts for the resource limitations inherent in biological systems. We apply this theory to simple models of biological information processing and decision-making and demonstrate that resource limitations induce discontinuous phase transitions in optimal strategies. This result suggests that resource limitations may contribute to the rich diversity observed in biological information processing and decision-making.
Shuhei Horiguchi: Kanazawa University
Optimal control problems for the population of interacting particles arise in various fields, including pandemic management, species conservation, cancer therapy, and chemical engineering. When the population size is small, the time evolution of the particle numbers is inherently noisy and modeled by stochastic reaction networks, a class of jump processes on the space of particle number distributions. However, compared to deterministic and other stochastic models, optimal control problems for stochastic reaction networks have not been studied well. This is partly due to the absence of a mathematically and computationally nice framework, such as the Linear-Quadratic-Gaussian (LQG) setting for diffusion processes. Without such a framework, one has to solve a large system of nonlinear (differential) equations, known as the Hamilton–Jacobi–Bellman (HJB) equation, to calculate the optimal solution.
In this study, we propose a mathematically and computationally nice framework for stochastic optimal control of reaction networks. By utilizing the Kullback–Leibler (KL) divergence as a control cost, the HJB equations can be linearized. This approach allows us to efficiently obtain the optimal solution, which shares a similar mathematical structure with previously discovered linearly solvable optimal control problems. We apply this framework to the control of interacting random walkers, birth-death processes, and stochastic SIR models, presenting numerical solutions, and even analytical solutions for some problems.
Ignacio Madrid: IIS, The University of Tokyo
Under antibiotics causing DNA damage, bacteria can trigger a stress response known as the SOS response. While the expression of this stress response can make individual cells transiently able to overcome antibiotic treatment, it can also delay cell division, thus impacting in different ways the population's ability to grow and survive.
Our goal is to contrast the different trade-offs that emerge from this phenomenon when we consider the survival probability of the population issued from a single cell, and when we consider the population growth rate (Lyapunov exponent) in constant and periodic environments. We show that the sensitivity of these two different notions of fitness with respect to the parameters describing the phenotypic plasticity differs between the stochastic approach (survival probability) and the deterministic approach (population growth rate). As a consequence, the optimal parameters that characterize the single-cell lineages that survive by the end of the experiment, and the parameters that maximise the exponential growth of the population might not be the same.
Finally, under a more realistic configuration of periodic stress, our results indicate that optimal population growth can only be achieved through fine-tuning simultaneously both the induction of the stress response and the repair efficiency of the damage caused by the antibiotic.
Praful Gagrani: IIS, The University of Tokyo
Autocatalytic reaction networks play a key role in self-replication and metabolism, appearing in diverse systems from chemistry to ecology. These networks consume and net-produce their own species, capturing the essence of self-replication. We introduce a formalism to quantify autocatalytic strength and develop efficient algorithms to identify the strongest, maximal, and minimal autocatalytic subnetworks. Applying our methods to the Formose and E. coli datasets, we find that while the strongest autocatalytic subnetworks in early chemical systems are minimal, evolution favors autocatalytic metabolisms composed of interlinked minimal subnetworks, whose combined strength surpasses that of any individual minimal autocatalytic subnetwork.
Masaki Kato: Department of Mathematical Informatics, Graduate School of Information Science and Technology, The University of Tokyo
Cue-mediated collective exploration by chemotactic cells represents a distinctive form of collective information processing. Although the optimality of single-cell responses has been established, the optimality of collective behaviors remains underexplored. In this study, we formulate the collective exploration problem within a reinforcement learning framework and reveal an optimal relationship that has been previously overlooked. Our derived optimal exploration dynamics not only correspond structurally to established models of collective dynamics but also suggest a class of more complex nonlinear models of cooperative behaviors.
Kyohei Kinoshita: Graduate school of Engineering, The University of Tokyo
In this study, we developed a method combining protein language model (SCEPTR) and optimal transport to efficiently identify low-frequency antigen-specific T-cell receptor (TCR) clones. While TCRs play a crucial role in pathogen recognition, conventional frequency-based and similarity-based methods struggle to detect low-frequency clones. Our approach generates embedded representations of TCR sequences and applies optimal transport algorithms at the V-gene level, achieving both computational efficiency and accuracy. Validation using COVID-19 infection data demonstrated higher accuracy for low clone count TCRs compared to existing methods, successfully identifying antigen-specific TCRs undetectable by conventional approaches.
Keisuke Sugie: Department of Mathematical Informatics, Graduate School of Information Science and Technology, The University of Tokyo
Slow relaxation in the CRN is important to better understand living cells under limited resources, and its mechanisms and characteristics have been studied. However, it remains unclear how the relaxation rates of CRNs can be quantitatively characterized by stoichiometric and thermodynamic features, as previous studies have been limited to qualitative or simulation-based approaches. In this work, we explicitly calculate the general bounds of the KL divergence in mass-action CRN dynamics for the two dissipative formulations. We exploit the convex structure of equilibrium CRNs and use convergence analysis methods in convex optimization algorithms to quantify the relaxation process. The bounds are determined by the stoichiometric singular values, the convexity indicators, and the activities of the reactions. We numerically validate the divergence bounds on specific CRNs exhibiting slow relaxation, and identify the minimal and sufficient pairs of quantities for plateau formulation.
Masamitsu Yasutake: The University of Tokyo
Metagenome analysis is a culture-independent method for analyzing microorganisms. A crucial step in metagenomic analysis is binning, which classifies DNA fragments by species. Existing binning methods rely on a sequence feature referred to as tetra-nucleotide frequency (TNF). TNF represents the occurrence frequency of 4-mer in a given DNA sequence and is empirically known to exhibit species-specific patterns. However, TNF does not retain the entire information of the DNA sequence, potentially leading to less accurate binning results. To address this issue, deep learning can be used to estimate alternative representations of DNA sequences instead of TNF. Recent advancements in deep state-space models (DSSMs) have enabled the development of models that directly input sequence information from metagenomic analysis with lengths exceeding 10,000 bps. Therefore, this study aims to learn DNA sequence representations using DSSMs and apply the model to metagenomic binning.
First, a training dataset was created based on a public microbial genome database. Using this dataset, representation learning for DNA sequences was performed. The network structure used was Hyena, a type of DSSMs. The model's embeddings evaluation demonstrated the model’s sufficient species discrimination ability. Additionally, binning experiments confirmed that the proposed method can detect microbial genomes from metagenomic datasets. Furthermore, evaluations with shorter metagenomic DNA fragments revealed that the proposed method is less susceptible to accuracy degradation compared to existing methods. In this presentation, we would like to discuss the similarity between representations derived from 4-mer and those learned by DSSMs.
Dimitri Loutchko (Proxy presentor: Tetsuya J. Kobayashi): The University of Tokyo
Absolute sensitivity is a novel concept in chemical reaction network (CRN) theory which quantifies how steady state concentrations react to changes in the linear conserved quantities [1]. It is defined independently of the choice of a basis for the space of conserved quantities and therefore its numerical values carry biochemical significance. By employing information geometrical tools, we derive results on the absolute sensitivity, even for CRNs with non-ideal thermodynamics. A linear algebraic characterization and explicit results on first order corrections to the ideal solution case are provided. The IDHKP-IDH glyoxylate bypass regulation system serves to illustrate the concepts.
The novelty of the theory is that it is applicable to CRNs with non-ideal thermodynamical behavior, which are prevalent in highly crowded cellular environments due to various interactions between the chemicals. Indeed, the analyzed example shows remarkable behavior ranging from hypersensitivity to negative-self regulations. These are effects which usually require strongly nonlinear reaction kinetics. However, here, they are obtained by tuning thermodynamical interactions providing a complementary viewpoint on such phenomena.
The technical basis is the information geometry of chemical reaction networks (CRNs) which has recently emerged as a framework to naturally connect the geometries appearing in CRN theory to their inherent thermodynamical [2, 3] and kinetic properties [4]. The classical Legendre duality in chemical thermodynamics between the spaces of chemical concentrations and potentials, which is induced by the convex potential functions, is thereby geometrically formalized by dually flat Riemannian manifolds [3]. For the present sensitivity analysis of CRNs, we introduce new Riemannian geometrical tools into the framework, the main result being a multivariate Cramer-Rao bound [5].
References
[1] D. Loutchko, Y. Sughiyama, and T. J. Kobayashi, arXiv preprint arXiv:2401.06987, (2024).
[2] T. J. Kobayashi, D. Loutchko, A. Kamimura, and Y. Sughiyama, Phys. Rev. Res. 4(3), 033066, (2022).
[3] Y. Sughiyama, D. Loutchko, A. Kamimura, and T. J. Kobayashi, Phys. Rev. Res. 4(3), 033065, (2022).
[4] T. J. Kobayashi, D. Loutchko, A. Kamimura, S. Horiguchi and Y. Sughiyama, Inform. Geom. 7(1), 97-166, (2024).
[5] D. Loutchko, Y. Sughiyama, and T. J. Kobayashi, in preparation, (2025).
Zi Hui (Proxy presenter: Tetsuya J. Kobayashi): The University of Tokyo
Predicting odor perception from chemical structures is a complex chemoinformatics task informing food and fragrance design. Nonetheless, the performance of conventional methods is still limited. To overcome this, we introduced a novel architecture (Mol-ABC) that fuses message-passing and attention layers, integrating atom, bond, and Coulomb matrices. The predictive power of our model surpasses traditional methods. Latent-space analysis reveals the nonlinear representation of odor molecules attained by our model, which identified odor-dependent global and local structures in the space and the impact of association with other senses to prediction.
Yusuke Himeoka & Tetsuya J. Kobayashi: Universal Biology Institute, The University of Tokyo
TBA
Keita Nakao: Nagoya University
In recent years, large-scale clinical data, such as electronic health records (EHR), has been accumulated and made available. The use of machine learning on this data is expected to lead to the prediction of disease onset and patient stratification, and technological development is underway for this purpose. Generally, EHR data is given as linguistic data such as words. Since machine learning models cannot process linguistic data as it is, it is necessary to convert it into numerical data appropriately. For linguistic data, natural language processing techniques such as Word2Vec and BERT can be used to convert them into numerical data[1]. However, these features are high-dimensional and may not be suitable for interpretable non-black box machine learning models. Furthermore, when the words consist of the same noun with adjectives of opposite meanings, yielded features will be very similar, which leads to misinterpretation by the machine learning models. Therefore, we propose a method to represent EHR data as numerical data suitable for use in machine learning models. We first cluster the diagnostic words in the EHR, then evaluate the goodness or badness of the words within the cluster, to represent the EHR data as a numerical vector for the number of clusters. When we applied the proposed method to the EHR data of Parkinson’s disease patients and healthy subjects, we were able to visualize the worsening or improvement of symptoms by looking at changes in the low-dimensional numerical data. With this representation, we could quantify the differences between the diagnostic time series data of Parkinson's disease and healthy subjects. We also analyzed the data digitized by the proposed method using a continuous-time hidden Markov model. The aim is to identify the prodromal stage of Parkinson’s disease. Parkinson’s disease is known to include constipation, depression, and REM sleep behavior disorder as prodromal symptoms, but whether and when these symptoms develop can vary among patients[2]. Based on these findings, we aim to identify and better understand the prodromal stage of Parkinson’s disease.
References
[1] Alsentzer, E., Murphy, J. R., Boag, W., Weng, W.-H., Jin, D., Naumann, T., McDermott, M. B. A, Publicly
Available Clinical BERT Embeddings, arXiv:1904.03323, 2019.
[2] Kalia, L. V., Lang, A. E. (2015). Parkinson’s disease. The Lancet, 386(9996), 896–912.
Kengo Inutsuka: Hiroshima University(Nagoya University)
The real world is filled with complex phenomena arising from uncertainty and diverse interactions. In such an environment, higher animals, including humans, exhibit a remarkable adaptive ability—termed “insight”—which enables them to spontaneously generate novel solutions to complex problems through trial and error. This insight emerges from the sudden realization of perspectives that differ from conventional approaches and represents a discontinuous process, in contrast to the continuous refinement of solutions. This discontinuous mechanism is believed to promote creative thinking that leads to new solutions. To efficiently address the complex problems of the real world by harnessing insight in an applied context, it is imperative to elucidate the discontinuous information processing mechanisms that occur within the brain.
To date, efforts have primarily focused on modeling and elucidating the neural basis of insight in humans (Jung-Beeman, Mark, et al., 2004). However, even in humans, the neural mechanisms underlying insight are not yet fully understood. Moreover, exploring analogous cognitive phenomena in non-human species is not only crucial for assessing the evolutionary underpinnings and universality of cognitive functions, but also provides important clues for supplementing and validating mechanisms that remain unresolved in humans. In humans, the moment of insight can be pinpointed via self-report, yet such determination is challenging in other species. Therefore, a deeper understanding of the neural basis of insight necessitates the development of time-series analyses of insight in non-human species, along with techniques for its quantitative characterization.
In this study, we formulate the decision-making process in a mouse behavioral task that requires insight within the framework of generative models. Using this formulation, we introduce a machine learning method to decode the time-series data of latent strategies and their switching points from experimentally acquired mouse behavioral data. In the poster presentation, we will present the results obtained using these methods.
Jigen Koike: Hiroshima University
"Neural circuits are formed through the extension of axons to appropriate target locations, where they establish synaptic connections with other neurons. The projection sites of these axons are believed to be determined by the concentration gradients of guidance molecules and their receptors, a hypothesis widely recognized as the “chemoaffinity theory.” For instance, in the retinotectal system, Eph receptors and their ligands, ephrins, form concentration gradients in the retina and its target, the tectum, respectively. These molecular gradients function as positional information for axonal projections, enabling the establishment of topographic connectivity patterns. Various other molecules involved in axonal guidance have also been identified, further reinforcing the chemoaffinity theory as a fundamental mechanism of neural wiring. However, previous studies have mainly focused on relatively simple neural structures, such as the retinotectal circuit, while the wiring mechanisms of more complex circuits, such as those in the cerebral cortex, remain poorly understood.
To address this gap, we developed a novel data-driven analytical approach based on the chemoaffinity theory to investigate the wiring mechanisms of complex neural circuits. Our method employs canonical correlation analysis (CCA), a machine learning technique, to analyze vector pairs of gene expression levels at the source and target regions of neural projections. This approach allows us to compute gradient pairs that best explain actual neural wiring patterns within the framework of the chemoaffinity theory. Furthermore, we applied this method to the publicly available mouse brain dataset from the Allen Brain Atlas to examine the relationship between whole-brain connectivity patterns and the chemoaffinity theory. Our analysis revealed that the gradient information governing the wiring structure of the mouse brain connectome exhibits a two-layered organization: (1) global connectivity patterns that define projection frequencies between regions, and (2) local connectivity patterns that reflect the spatial relationships between specific brain areas. These findings provide novel insights into the principles underlying the formation of complex neural circuits and offer a data-driven framework for understanding brain connectivity."
Misako Kimura: CiNet
Misako Kimura(A)(B), Tsutomu Murata(A), Shigeto Seno(C), Izumi Ozawa(A)(C), Kunihiko Kaneko(A)(D), Yanagida Toshio(A), Kazufumi Hosoda(A)(C)(E)
(A)NICT CiNet, (B)University of Hyogo, (C)Osaka University, (D)Niels Bohr Insititute, (E)Kobe University
One aspect of thought, the eureka effect, is a phenomenon in which existing information is unconsciously reorganized based on prior experiences and knowledge, allowing for the sudden discovery of novel solutions or perspectives for unfamiliar problems and situations. For example, when observing an extremely degraded binary image for some time, one may suddenly recognize the hidden object within it. The mechanism underlying such eureka moments remains largely unexplored.
Previous studies in psychophysics have shown that the difficulty of degraded images can be characterized by natural numbers. Based on this finding, the difficulty of a degraded image corresponds to the number of missing parts of the hidden object. During the search process, these missing parts probabilistically activate, and when multiple missing parts are simultaneously activated and perceptually completed, a eureka moment occurs. Another study built an artificial neural network model that reproduces this discreteness. In this model, deep learning is used to extract image features, and during the search process, features randomly fluctuate. A eureka moment arises when the number of activated features exceeds a threshold. However, the randomness and threshold-based logic in this model deviate from the framework of neural networks, making it questionable whether it accurately reflects the mechanisms of the human brain.
In this study, we applied a recurrent neural network (RNN) described by time-differential equations as a dynamical systems model for the search process in eureka moments. This model explores solutions through chaotic dynamics and autonomously converges when the necessary features are acquired. Although integration with deep learning has not yet been implemented, our simulations using the dynamical systems model alone demonstrate that the discrete characteristics observed in previous studies can be reproduced.
Thoma Itoh: SOKENDAI / NIBB / Hiroshima Univ.
Organisms are systems capable of making decisions by integrating multiple signals. Understanding the mechanisms and reasons behind signal integration is a primary question in biology. This study approaches this from two perspectives.
1. The mechanisms of signal integration in gonad development in medaka.
Medaka gonad development exhibits a distinct annual cycle. While the previous study modeled the relationship between environmental signals and gonad size as a steady system, this may overlook seasonal shifts in signal prioritization. By applying a time-variable model, we quantitatively revealed seasonal changes in signal prioritization. Moreover, a subset of genes show strong correlation with these changes, suggesting their involvement in adjustment of signal integration.
2. The evolutionary process of signal integration in molecular networks.
The ERK pathway, which receives multiple inputs, exhibits input-specific responses despite the signals are mixed within the pathway. This is enabled by input-specific ERK dynamics, known as temporal encoding. While the mechanism is gradually understood, its evolutionary process remains unclear. We developed an abstract model to investigate the conditions that evolves temporal encoding, and found that temporal encoding emerges under the pressure for rapid response.
Kota Mitsumoto: University of Tokyo
In morphogenetic processes, a characteristic gene expression pattern is formed using the regulations among genes, called a gene regulatory network. Essential regulations for the pattern in each species can be determined by experiments. However, there is no theoretical study to unbiasedly search gene regulatory networks to form the pattern. We construct an ensemble of gene regulatory networks forming a segmentation pattern using multicanonical Monte Carlo simulation with the Wang-Landau method. We aim to investigate the statistical properties of the ensemble and classify the networks, based on network theory and information theory.
Sota Shimamura: Graduate school of arts and science, The University of Tokyo
Collective cell motion is observed in various biological contexts, including morphogenesis, wound healing, and cancer metastasis. This collective movement arises from both individual cell motility and intercellular adhesion, yet the conditions that enable coordinated motion remain unclear. Here, we investigate an active Brownian particle model with attractive interactions and find that a multi-cluster state emerges when motility competes with inter-particle attraction. Surprisingly, each cluster in this regime moves faster than expected, which can be attributed to the resultant alignment of motility directions when particles collectively escape from large clusters. Our results suggest that cells can migrate more effectively when motility and adhesion are balanced. In this poster, we present a phase diagram of attractive active Brownian particles derived from numerical simulations and theoretical analysis.
TBAi: TBA
TBA