Daily Summary for 2026-11-08
daily
Transcript
Introduction to the show: ident: Genomics Radio. Generated commentary on the latest computational biology and genomics papers.
Ines: It's the eighth of November, twenty twenty-six, and this is the day's research.
Marcus: 11 new papers came out today.
Ines: I'm Ines.
Marcus: And I'm Marcus, here with Yuki.
Ines: We'll take the day in one pass, then pull out the papers we're staying with.
Marcus: Start us off with GINsim 3.1: Qualitative modelling and dynamical analysis of logical regulatory networks.
The day's research: Ines: The focus is regulatory networks dictating cell behavior and modeling these dynamics with GINsim 3.1, introducing symbolic computation for trap-spaces.
Marcus: GINsim 3.1 helps analyze protein interactions; the mammalian cell cycle case study showed negative circuits drive oscillations for rhythm generation.
Yuki: Modeling gene expression noise examines how cell growth and division leave distinct fluctuation signatures on molecule numbers using that approach.
Ines: That work suggests static models fail because they ignore the cell cycle; incorporating growth dynamics is necessary for cellular heterogeneity understanding.
Marcus: Research also touched upon evolving neural networks for seizure detection by changing calcium parameters per neuron.
Yuki: That study showed evolving calcium parameters improved seizure detection performance on held-out patient data, suggesting network evolution helps find better detectors.
Ines: The most significant finding relates to agents autonomously building complex virtual tissue models using Morpheus.AI agents.
Marcus: Morpheus.AI uses a frontier large language model to generate and refine these models from existing benchmark studies, achieving high accuracy in under two hours for published models.
Yuki: This agentic workflow involves agents calling tools, running code, and using feedback to repair failed attempts within the simulator.
Ines: The construction of a multiscale virtual tissue model is framed as a slow task requiring abstraction into mathematical frameworks.
Marcus: Performance was strong, meeting eighty-three to eighty-six percent of study criteria in less than two hours for models with published descriptions.
Yuki: Furthermore, the agentic loop proved robust; an LLM scorer agent checked the generated model against criteria and agreed with expert scores in nearly ninety-four percent of runs.
Ines: Testing five tool setups revealed that GPT-6.1 Sol outperformed other models across all configurations, suggesting the LLM choice is critical.
Marcus: GINsim is software for defining and dynamically analyzing qualitative models of biological regulatory networks.
Yuki: BIOTRAM is a neural network designed to estimate per-taxon specimen counts by learning read yield specific to each specimen.
Ines: From Connection Loss to Collision: Closed-Loop Fault Assessment of a Fly-Connectome SNN quantifies finite-horizon collisions and record check gains in a fly spiking neural network.
Marcus: A Bayesian Optimized Biquad Filter-Burg Model for Real-Time EEG Phase Estimation presents a filter model achieving real-time phase estimation on EEG signals better than existing methods.
Yuki: The Kinetic Basis of the Genetic Code investigates codon-anticodon association using a model based on asymmetric cooperativity.
Ines: Evolved Spiking Neural Networks for EEG Seizure Detection: Calcium Dynamics and Delay Mechanics evolving calcium parameters per neuron gave detectors higher MCC and fewer false alarms on held-out patients.
Marcus: Modeling Gene Expression Noise: From Transcriptional Dynamics to Cell Growth and Division reviews models of stochastic gene expression in growing cells coupled to cell cycle dynamics.
Yuki: Mathematical modeling to study combination strategies for neural stem cell therapy in glioblastoma develops a hybrid model of NSC migration coupled with chemoattractant profiles.
Ines: Fly Whiskers Meet LiDAR extends a navigation network by integrating a 32-ray LiDAR interface to improve robustness when vision fails.
Marcus: Kinetic models of a K channel with 2 stage voltage sensor activation and an opening step reduces a complex kinetic model to an effective three-state system under specific conditions.
Yuki: Morpheus.AI uses large language model agents to autonomously build and refine virtual tissue models from biological studies.
Lucky paper: 2611.04109: Ines: Okay, so we’re looking at GINsim three point one now, and the paper focuses on symbolic computation for trap-spaces and Booleanisation of multilevel models for large logical networks.
Marcus: So this software tool lets you define qualitative models of biological regulatory networks using directed signed graphs where vertices are proteins or genes, right?
Yuki: It uses Boolean variables where zero means inactive and one means active, and it handles multilevel variables with rules like AND or OR to specify activity levels.
Ines: The paper details several tools, including computing state transition graphs, finding stable states using simulation-free procedures based on decision diagrams, and identifying trap-spaces symbolically without running simulations.
Marcus: And they also have model reduction functionalities that compress the regulatory graph by hiding components while updating the logical rules for targets.
Yuki: They introduce Booleanisation of models to convert multi-valued logical models into equivalent Boolean ones using a mapping proposed by van Ham, which prevents exploring non-admissible states.
Ines: The application example uses a reference model of the mammalian cell cycle to show how the trap-spaces identification tool leads to two stable states when CycD is inactive and one cyclic attractor when it is active.
Marcus: That's interesting because they also showed that a double perturbation, knocking out Rb and ectopically activating E2F, leads to two cyclic attractors, which implies cyclin oscillations even without growth factors.
Yuki: And the paper highlights that functional negative circuits are what account for the observed oscillatory behavior in that model.
Ines: They also talk about model reversal, which generates a model whose asynchronous successors correspond to predecessors of the original model, effectively reversing the direction of dynamics.
Marcus: The software is distributed with a repository containing around seventy published logical models spanning three decades of work and supports scripting through the CoLoMoTo environment for reproducible workflows.
Yuki: It mentions future plans to extend the GINsim Python interface to give access to more functionalities from that CoLoMoTo environment.
Lucky paper: 2611.05374: Ines: Alright, we're looking at the Bayesian Optimized Biquad Filter-Burg Model for Real-Time EEG Phase Estimation now. This paper tackles the need for real-time phase estimation on stimulus-aligned transient EEG, which is crucial for closed-loop neuromodulation systems targeting the theta band.
Marcus: Most existing real time phase estimation methods have been validated on resting state, high fidelity signals in the alpha band, so this study characterizes those spectral and temporal properties of non anticipative bandpass filtering.
Yuki: The authors present a framework that jointly tunes five parameters of a biquad filter-Burg model using a Gaussian process surrogate over a short calibration split.
Ines: They did record EEG during peripheral vibrotactile stimulation of the left median nerve for evoked theta evaluation, sampling at three hundred Hertz.
Marcus: The stimulus event markers were extracted from the raw BCI2000 files to define epochs from negative one second to positive three seconds relative to stimulus onset for various network inputs.
Yuki: This Bayesian Optimized Biquad Filter-Burg model uses a biquad filter parameterized by a nominal center frequency fc and an exponential decay time tau, which determine the pole angle and radius of the filter.
Ines: They optimize the objective function MACEcal by minimizing it over the search space for parameters p, C, fc, tau, and P.
Marcus: The ablation study showed that optimization reduced mean theta MACE from fifty-four point five three degrees to forty-eight point three three degrees after one hundred twenty evaluations of that model.
Yuki: The filter order was identified as the dominant parameter in both bands and reconstruction modes, with an order one filter selected throughout both cohorts, which is a specific detail we should keep in mind for deployment.
Ines: The study found that backward reconstruction reduced alpha MACE by three point zero seven degrees but increased theta MACE by one point eight one degrees.
Marcus: BOB achieved mean accuracy of seventy-three point twenty-five percent in theta and eighty-four point forty-five percent in alpha, ranking first and second, respectively, among the five evaluated estimators.
Yuki: These results suggest that temporal alignment should be considered alongside spectral selectivity when choosing a filter for real time phase estimation.
Lucky paper: 2611.03841: Ines: So, we're looking at Evolved Spiking Neural Networks for EEG Seizure Detection: Calcium Dynamics and Delay Mechanics. The paper talks about evolving calcium parameters per neuron in integrate-and-fire networks giving seizure detectors better performance on held-out CHB-MIT patients.
Marcus: That sounds promising, but what exactly does that mean for how we build these detection systems? Are we talking about a small change in the network structure, or is it something deeper?
Yuki: The core methodology involves the EONS framework which evolves the topology, weights, thresholds and delays of neuromorphic spiking networks without gradients. It compares fix parameters to evolved ones.
Ines: The results show that evolving the calcium parameters per neuron gave a gain of plus zero point one three four over the integrate-and-fire baseline on CHB-MIT data, and it lowered the false alarm rate from zero point one seven down to zero point one zero.
Marcus: That reduction in false alarms is pretty significant, but I wonder if that gain holds up when we look at different datasets or patient populations outside of CHB-MIT.
Yuki: The false alarm reduction was consistent across patients for the Ca-IFneu setting, with it lowering the FPR in seven out of ten held-out patients. Also, short evolved synaptic delays, specifically those between one and three ticks or twenty five to seventy five milliseconds, gave a similar gain on CHB-MIT data.
Ines: So the paper suggests that both calcium evolution and evolved delays might help the evolution find better networks, rather than just changing how the finished networks work.
Marcus: That's a key distinction; it implies we aren't just tuning existing models but letting them adapt to the signal characteristics of different patients. Where does this leave us regarding generalizability?
Yuki: On Siena data, Ca-IFneu gained plus zero point zero five nine over the integrate-and-fire baseline in one patient-disjoint split, and plus zero point zero nine six in the other split.
Ines: But there's a caveat mentioned about generalizability: scoring CHB-MIT trained networks on Siena without retraining resulted in lower MCC scores than on CHB-MIT, and the Ca-IFneu advantage wasn't significant then.
Marcus: That points to a limitation, right? The study notes that the modest cohort size and limited number of seizure records are primary limitations they are flagging.
Yuki: They also mention that longer or fixed delays and axonal delays gave no significant gain over the baseline in terms of improvement on CHB-MIT.
Ines: It’s interesting how they're pointing toward neuromorphic hardware next, given these software results on small cohorts. That suggests they see this as a path forward for larger evaluations.
Marcus: If we look at what this means for deployment, does implementing this evolution process add significant computational overhead compared to just using the fixed parameters?
Yuki: The study indicates that synaptic delays were fixed at the minimum transmission time of one tick in their main experiments, so those specific delay parameters weren't evolved.
Ines: So the gain came primarily from tuning the calcium dynamics, and that’s what they focused on for this initial study.
Marcus: I think it shows that even in complex systems like spiking neural networks, there's a lot of room for finding better configurations by allowing parameters to evolve.
Lucky paper: 2611.06537: Ines: We’re looking at Modeling Gene Expression Noise from Augustinas Sukys et al., which focuses on how cell life history leaves a signature on molecule numbers.
Marcus: That paper reviews models of stochastic gene expression in growing and dividing cells, aiming to understand that distinct fluctuation signature.
Yuki: The standard model is the telegraph or two-state model, but both the telegraph and bursty models predict steady-state distributions that often fit snapshot data.
Ines: But the authors point out that their simplicity comes at the cost of neglecting key biological detail about transcription states.
Marcus: They also assume a static cell description, which means they exclude a cell cycle description involving birth, DNA replication, growth and division.
Yuki: Incorporating cell growth and division adds several sources of biological noise, like random partitioning of molecules at division or the doubling of gene copy number upon DNA replication.
Ines: The paper discusses three phenomenological strategies for cell size control: the sizer which divides once a threshold is reached, the timer which occurs after a fixed time, and the adder which occurs after adding a fixed volume.
Marcus: Cell size-dependent transcription is thought to be important because it helps maintain stable mRNA concentrations during growth.
Yuki: The authors suggest these three strategies can be distinguished by looking at the slope of added size over the cycle against birth size, assuming exponential growth.
Ines: The authors say that a good statistical fit of the steady-state solution to single-cell data doesn't guarantee biophysically realistic predictions or meaningful kinetic parameters.
Marcus: They stress that fitting mechanistic models explicitly incorporating cell growth and division to cell-cycle resolved single-cell data is the gold standard for overcoming these limitations.
Yuki: Beyond inference, they say the scope needs to widen to include multi-state promoters and regulatory interactions, but deriving exact likelihoods remains difficult.
Ines: The conclusion is that the telegraph model and its bursty limit are still widely used descriptions of stochastic gene expression, but fitting them to cell-cycle resolved data is promising.
Marcus: We also see that spatial transcriptomics allows measurement of expression variability at the subcellular and tissue levels, but combining stochastic gene expression with intracellular transport remains undeveloped.
Yuki: Morpheus.AI agents autonomously build and refine virtual tissue models from published studies, achieving accuracy in under two hours for models with descriptions.
Ines: The performance was strong, hitting eighty-three to eighty-six percent of study criteria in less than two hours for those existing models.
Marcus: Furthermore, the agentic loop proved robust because a separate LLM scorer agent agreed with manual expert scores in nearly ninety-four percent of runs.
Yuki: Testing five different tool setups showed that GPT-six point one Sol outperformed other models across all tested configurations, suggesting the LLM choice is critical for this task.
Ines: That points to the choice of the large language model being a critical factor when using agents to build these complex virtual tissue models.
Marcus: The paper acknowledges funding from several sources including the ARC Centre of Excellence for the Mathematical Analysis of Cellular Systems and an Australian Research Council Laureate Fellowship.
Yuki: It’s interesting how they frame fitting mechanistic models to cell-cycle resolved data as a way to understand gene regulation at a genome-wide level, even if deriving exact likelihoods is hard.
Lucky paper: 2611.01374: Ines: We're looking at Mathematical modeling to study combination strategies for neural stem cell therapy in glioblastoma today.
Marcus: It shows that IL-ten pretreatment improves neural stem cell arrival numbers at tumors only under the Gaussian chemoattractant profile.
Yuki: That means the effect of IL-ten is not universal; it really depends on how those chemoattractants are distributed spatially around the tumor.
Ines: Exactly, and CXCL12 pretreatment shows a moderate improvement when those tumors are located in the anterior brain region under a constant chemoattractant profile.
Marcus: But radiotherapy seems to significantly enhance tumor arrival when you use uniform profiles for those treatments.
Yuki: The researchers developed a hybrid mathematical model that couples an agent-based model with a partial differential equation to capture the migration dynamics in the brain with cancer.
Ines: That hybrid approach uses an agent-based framework for individual NSCs and an anisotropic reaction diffusion partial differential equation to manage chemoattractant concentration evolution.
Marcus: The agent-based model assumes that NSCs move preferentially along white matter tracts while showing random motion in gray matter, using fractional anisotropy to distinguish those regions.
Yuki: And the stochastic differential equation of the k-th cell includes tissue-guided motility, chemotaxis, and random motility terms.
Ines: They incorporated two chemoattractants, VEGF as C1(t,x) and MCP1 as C2(t,x), describing their evolution with time as they diffuse and diminish due to interaction with the NSCs.
Marcus: The reaction diffusion equation describes how the concentration of each chemoattractant evolves with time under the Fokker-Planck diffusion operator for anisotropic diffusion.
Yuki: Chemoattractant production is modeled as a sum of a tumor-localized source and a spatially distributed background source in logistic form, motivated by observations of elevated levels near brain tumors.
Ines: The study investigated three combination strategies: IL10 pretreatment, CXCL12 pretreatment, and radiotherapy.
Marcus: IL10 pretreatment is incorporated by increasing the stochastic chemotactic sensitivity parameters given as lambda times eta of w.
Yuki: It was found to be ineffective under a constant chemoattractant profile but moderately beneficial under a Gaussian profile when that chemosensitivity parameter is low.
Ines: CXCL12 pretreatment increases the migration step size dw on white matter, which can improve tumor arrival, especially for tumors located in the posterior region of the brain.
Marcus: Radiotherapy is investigated as an intervention to actively reshape the chemoattractant environment by inducing inflammatory and stress responses.
Yuki: The simulation results suggest that both treatment strategies and chemoattractant distribution influence intranasal NSC delivery.
Ines: Specifically, the Gaussian chemoattractant case shows successful tumor-directed migration only when the chemosensitivity parameter is large enough, but in the constant profile case, variations in chemosensitivity have little to no effect on NSC localization.
Marcus: The study concludes that treatment efficacy is strongly dependent on both the underlying chemoattractants profile and NSC responsiveness to chemotactic cues.
Yuki: Future work should focus on relaxing idealized assumptions imposed on those chemoattractant spatial profiles as more detailed experimental data becomes available, and the model can be extended to human-scale applications using patient-derived imaging data.
More episodes
- 2610.33881-Sensorimotor behavior from an untrained brain-and-nerve-cord connectome
- 2610.40344-Morphological Features of Murine Pulmonary Arterial Networks for Diagnosing Vascular Disease
- Daily Summary for 2026-11-19
- 2609.10193-Are You Learning Biological Signal or Shortcuts? Auditing and Mitigating Bias in Protein-Protein Interaction Datasets
- 2610.38236-Generative discovery of chemical reaction networks
- 2507.11176-An Interpretable AI framework Quantifying Traditional Chinese Medicine Principles Towards Enhancing and Integrating with Modern Biomedicine
- The day in conversation for 2026-11-19
- Daily Summary for 2026-11-19
- 2504.03590-Towards a Unified Framework for Determining Conformational Ensembles of Disordered Proteins
- 2603.15217-A multiscale discrete-to-continuum framework for structured population models