Daily Summary for 2026-10-31
daily
Transcript
Introduction to the show: ident: Genomics Radio. Generated commentary on the latest computational biology and genomics papers.
Ines: It's the thirty-first of October, twenty twenty-six, and this is the day's research.
Marcus: 12 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 Leveraging genomic deep learning models for the prediction of non-coding variant effects.
The day's research: Ines: We are using deep learning to sense genetic variations outside coding regions because non-coding effects matter for genetics.
Marcus: AKMPC reviews how supervised sequence-to-activity models and self-supervised language models predict phenotypes from DNA sequences.
Yuki: scTREAT uses a diffusion framework to predict single-cell drug response by factoring state and drug guidance into a shared latent space.
Ines: That model combines baseline cell characteristics with dose information to predict effects on unseen lines or drugs.
Marcus: scMRDR tackles integrating unpaired single-cell multi-omics data using a generative framework called disentangled representations.
Yuki: It separates cell representations into modality-shared and modality-specific components to scale well across different omics data types.
Ines: QUAREP-LiMi focuses on common quality control standards for light microscopy images to boost reproducibility in research.
Marcus: That effort seeks accepted guidelines and tools for assessing image quality across various scientific studies.
Yuki: pyAvalanches is a Python package that detects neuronal avalanches from recordings using avalanche transition matrices to map propagation patterns.
Ines: The most significant finding relates to how persistent homology reveals underlying topological structure in spontaneous brain-organoid activity.
Marcus: This matters because it shows complex neural activity has a structured topology even with modest node counts.
Yuki: This method reads structure directly from pairwise correlations without assuming which variables are important beforehand across human and mouse cortical organoids.
Ines: The integrated first Betti number summarizes the loop structure across the whole filtration in 14 of 18 datasets exceeding null model predictions at zero point zero five significance.
Marcus: This means loops, which are cycles in data patterns, are present even when local measures miss them due to a strong-correlation backbone.
Yuki: The loop structure is carried by an identifiable core of units rather than being spread redundantly across the population suggesting robustness against recording geometry artifacts.
Ines: Higher-order structure like enclosed voids only appears when network size reaches at least one hundred units indicating loops are the first level of complexity before voids emerge.
Marcus: This topological signature indicates loop structure is functional within self-organizing tissue connecting to schema-based active inference which suggests abstract relational structures underlie rapid generalization.
Yuki: Schema-based active inference focuses on cognitive frameworks rather than purely topological features in organoid activity.
Ines: Leveraging genomic deep learning models for the prediction of non-coding variant effects is a review on using deep learning to predict how changes in non-coding DNA affect biological functions.
Marcus: scTREAT is a new diffusion model that predicts single-cell drug response by combining shared latent expression with explicit state and drug conditioning.
Yuki: QUAREP-LiMi is an initiative to create common quality standards for light microscope images to improve reproducibility in scientific research.
Ines: Multimodal Knowledge Aggregation for Microbial Keratitis uses a multimodal model to classify bacterial versus fungal microbial keratitis from pooled patient data.
Marcus: Mechanistic bridges from receptors to whole-brain dynamics reviews how models connect molecular changes at the receptor level to large-scale brain activity.
Yuki: scMRDR is a scalable framework that integrates unpaired single-cell multi-omics data by disentangling modality-shared and modality-specific representations.
Ines: Inferring Inference proposes a mathematical framework to infer canonical distributed computations from large patterns of neural activity in the cortex.
Marcus: pyAvalanches is a Python package that provides tools for detecting and analyzing spatiotemporal propagation patterns or neuronal avalanches from electrophysiological recordings.
Yuki: Emergent topological structure in spontaneous brain-organoid activity shows that persistent homology can resolve real structural topology in developing brain tissue recordings.
Lucky paper: 2411.11158: Ines: So, taking a look at *Leveraging genomic deep learning models for the prediction of non-coding variant effects*, we see they are comparing sequence-based approaches, like supervised and self-supervised language models.
Marcus: The paper talks about supervised sequence-to-activity models predicting functional activity levels directly from DNA sequences using things like chromatin accessibility or histone modifications.
Yuki: And then you have the self-supervised genomic language models that learn representations by predicting a masked portion of the input DNA sequence based on surrounding context.
Ines: When they are making variant effect predictions with supervised models, they generate two separate functional activity predictions, one for the reference allele and one for the alternate allele.
Marcus: They compare these using different distance metrics like absolute differences in counts or relative log fold-change differences between the two predictions.
Yuki: They also use multi-task learning to make simultaneous predictions for hundreds or thousands of cell types and experimental assays, which then get aggregated into a single score per variant.
Ines: If you have multiple tasks relevant, they can average these effect predictions over many cell types or assays, take the maximum prediction, utilize a lower-dimensional representation of the predictions, or train an additional supervised model to learn a weighting.
Marcus: For purely self-supervised models without labeled functional data, they compare sequence embeddings for the reference and alternate allele-containing sequences.
Yuki: Another way they predict effects from genomic language models is by masking the variant location in the input sequence and getting the predicted probability of each allele being at that position based on surrounding sequence context.
Ines: These predicted probabilities for reference and alternate alleles are compared using simple distance metrics like the log-likelihood ratio.
Marcus: The paper discusses validation data sources, which include observational data from GWAS and QTL studies, as well as experimental data from massively parallel reporter assays or CRISPR-based perturbation experiments.
Yuki: A limitation they point out with association studies is that they can identify both causal and non-causal variant associations because of correlations between nearby variants, or linkage disequilibrium.
Ines: They also note that association studies have power to detect effects of only relatively common variants and are biased by the genetic ancestry of the studied cohorts.
Marcus: The experimental assays test functional activity levels of synthetic DNA sequences or make perturbations to native DNA to collect functional data from sequences not present in observed populations.
Yuki: When constructing evaluation tasks, they use binary classification for functional variant identification, effect direction prediction, and regression tasks for effect size prediction.
Ines: When using genome-wide summary statistics from GWAS or QTL studies to evaluate predicted effect sizes, they use approaches like stratified LD score regression or signed LD profile regression that explicitly account for linkage disequilibrium.
Marcus: They also mention that successful variant effect prediction at a given scale might require specific modeling choices because chromatin states tend to be modulated by proximal sequence.
Yuki: Predicting effects on epigenetic or chromatin state requires a causal understanding of relevant sequence features like transcription factor binding motif disruption.
Ines: A consistent finding is that predicted variant effects are more accurate when the magnitude of the predicted difference between reference and alternate alleles is larger, as shown by DeepSEA for chromatin accessibility and histone modification variants.
Marcus: Similarly, CpGenie for methylation QTLs and DeepFIGV for chromatin accessibility QTLs show this trend where larger magnitudes lead to better accuracy.
Yuki: The paper also looks at transcriptional regulation predictions, noting that supervised sequence-to-activity models perform better than non-deep-learning baselines for identifying gene-proximal expression-altering variants.
Ines: However, they admit models show room for improvement in predicting distal variant effects and personalized gene expression.
Lucky paper: 2510.11726: Ines: So we're looking at scTREAT now, which is the paper focusing on latent diffusion for single-cell drug response prediction.
Marcus: It introduces scPPDM, which is the first diffusion-based framework for predicting single-cell drug response from scRNA-seq data.
Yuki: The most informative sentence seems to be that scPPDM couples two condition channels, pre-perturbation state and drug with dose in a unified latent space via non-concatenative GD-Attn.
Ines: That means they’re using a shared latent space to handle both the starting cell condition and the drug treatment simultaneously.
Marcus: The model has this latent response schema with three coupled components: first, a fine-tuned SCimilarity VAE defines the shared latent geometry and learns response dynamics in section three point one.
Yuki: Second, baseline states and drug information are encoded, fused, and injected through Guided Decomposable Attention in section three point two.
Ines: And they use four-state independent channel dropout during training with factorized classifier-free guidance during inference using two knobs: sp for state and sd for drug.
Marcus: The training objective is defined as Ltotal equals Ldenoise plus Linject, where Ldenoise is the MSE on noise and Linject is the injection regularizer.
Yuki: During inference, factorized guidance uses those two knobs with a mapping dose to sd(dose) for intensity control in the denoising step.
Ines: The performance results show new state-of-the-art across log fold-change recovery, delta correlations, explained variance, and DE overlap on the Tahoe-100M benchmark under UC and UD regimes.
Marcus: Representative gains include a +thirty-six point one one percent plus a +thirty-four point two one percent on DEG logFC Spearman or Pearson in UD over the second-best model.
Yuki: Ablation studies confirm that conditioning is indispensable and GD-Attn improves fit and delta alignment over Concat plus MLP while preserving dose-rank behavior.
Ines: They also found that structure-aware priors are critical for UD generalization because replacing structure with Drug ID hurts most in that regime.
Marcus: The paper details extensive reproducibility steps including the latent VP diffusion in section three point one and the guidance decomposition analysis in appendix B.
Yuki: The protocol adopts plate-matched DMSO pairing with within-key averaging to avoid group mean degeneration and reduce uninformative gradient noise across all evaluations.
Ines: scPPDM is presented as an interpretable engine for rapid compound and dose scoring that shrinks wet-lab effort through its dual knobs enabling transparent what-if analyses.
Marcus: This supports early stage drug discovery and screening as well as precision medicine because of the transparency in how the model makes its decisions.
Yuki: Looking ahead, scPPDM is an interpretable engine for rapid compound and dose scoring that shrinks wet-lab effort supporting early stage drug discovery and screening.
Ines: The complete ablation experiments description and further analysis are given in appendix F with all hyperparameters listed in appendix D.
Marcus: Implementation details for tokenization and GD-Attn are in appendix G, but the dataset pairing protocol is described in section four point one.one with the rationale against random single-cell pairing in appendix E.
Yuki: All evaluation metrics including logFC delta correlations explained variance and DE overlap accuracy are precisely defined in appendix C for reference.
Lucky paper: 2608.00306: Ines: We’re looking at Mechanistic Bridges from Receptors to Whole-Brain Dynamics and the tension between fidelity and tractability here.
Marcus: So it seems like the core idea is that you can maintain interpretability by preserving certain microscopic elements, specifically receptor-dependent synaptic kinetics, conductance state dynamics, and spike-frequency adaptation.
Yuki: But those preserved mechanisms are only manipulable across scales conditionally on things like coarse-grained Markovianity and population homogeneity.
Ines: That means the whole framework isn't an exact derivation of whole-brain dynamics but a tractable approximation chain, right?
Marcus: Exactly, and the paper traces this lineage from finite-size population statistics up to connectome-coupled whole-brain dynamics through those transfer functions.
Yuki: And they are suggesting that effective parameter changes and transfer functions serve as the bridge for causal analysis when those specific parameters are calibrated within their bounded validity domains.
Ines: That shifts the focus from finding a single perfect model to building a sequence of models that allows for testable consequences at different levels.
Marcus: Speaking of tractability, did you see the computational cost analysis in Part B comparing master-equation mean-field models like MPR equations to microscopic spiking neural network simulations?
Yuki: The MPR equations offer an exact macroscopic reduction for a specific class of systems, but achieving that requires corresponding assumptions for every population and coupling term.
Ines: And structurally, the MPR model is efficient because its dynamics are closed in only two state variables per population and the Right-Hand Side is a low-order polynomial.
Marcus: It's significantly cheaper than the first-order adaptive node model because it avoids complex elements like erfc-based transfer functions and subthreshold moment blocks.
Yuki: The whole brain cost for a whole brain network of K nodes with connectome density rho scales as (2P squared + 11P)K plus two rho K squared.
Ines: That scaling means the cost over one biological second, W MPR 1s, doesn't depend on the number of microscopic neurons N represented by the node.
Marcus: That’s a big difference because an SNN cost scales explicitly with N: Tbio t cneur N plus Tbio csyna nu p N squared.
Yuki: The paper sets up a benchmark where you compare a regional SNN against a mean-field network built from the same regional circuit to ensure matched regional scale and structure.
Ines: So, while the whole-brain model might be computationally cheap in terms of K, the underlying complexity is still tied directly to N.
Marcus: The conclusion suggests that these models are powerful tools for causal analysis and prediction of large-scale propagation but they aren't exact derivations on their own.
Yuki: The research points toward a necessary comparative framework between regional SNNs and their mean-field surrogates as the most informative current benchmark.
Lucky paper: 2310.03186: Ines: Alright, let's talk about Inferring Inference. This paper proposes a mathematical framework to infer canonical distributed computations from large-scale neural activity patterns.
Marcus: It integrates normative and algorithmic theories of neural computation to find latent causes explaining sensory inputs and then uses message-passing as a nonlinear algorithm on a graph structure.
Yuki: The framework simultaneously tries to find the neural representation of latent variables, the interactions between those variables defining the brain's internal model, and the specific message functions that specify the inference algorithm.
Ines: So you're saying they are constructing a Hidden Markov Model where every part—encoding, interactions, message functions—affects how likely measured neural activity is given sensory observations.
Marcus: They collect all these parameters into one big vector theta which includes parameters for the neural encoding function, coupling parameters, and the message-passing functions.
Yuki: The inference procedure uses a Particle EM algorithm to get the maximum likelihood estimate of TAP brain parameters, involving an E-step with a particle filter to approximate the posterior distribution of latent variables.
Ines: After that comes an M-step using gradient ascent on the Q function, and they use greedy backward optimization to break degeneracies in estimating the coupling matrix J and message-passing parameters G by finding a sparse subset.
Marcus: The experimental design considerations suggest presenting stimuli with a wide range of intensities is best to identify both the embedding and the dynamics, specifically when input gain is in the range of twenty twenty-five.
Yuki: When that input gain is in that range, latent dynamics explore their full range resulting in a fairly uniform distribution of latent states which allows Independent Component Analysis to get a good initial estimate for the neural embedding matrix R hat.
Ines: But the authors acknowledge several assumptions they could relax, like assuming the brain computes via canonical message-passing on graphs instead of just that.
Marcus: They suggest generalizing this to richer classes of graphical models that account for continuous variables and directed acyclic graphs to capture causal structure instead of just canonical operations.
Yuki: They also propose augmenting the framework to include world dynamics by allowing an internal model based on a spatiotemporal graph rather than just a static one.
Ines: It's interesting how they are focusing on this canonical operation idea because it's central to understanding any kind of computation.
Marcus: They compare their model performance against other models that do not use canonical shared message functions to see what the gain is.
Lucky paper: 2609.11530: Ines: Let's talk about pyAvalanches now, the paper is titled pyAvalanches: A Python Package for Analyzing Spatiotemporal Propagation in Neuronal Avalanches.
Marcus: So this package is an open-source Python tool that aims to provide an end-to-end pipeline for detecting, characterizing, and analyzing neuronal avalanches from electrophysiological recordings like EEG.
Yuki: It standardizes the analysis workflow and incorporates the computation of Avalanche Transition Matrices or ATMs to map spatiotemporal propagation patterns.
Ines: The core innovation here is that it derives network-based metrics directly from those ATMs, bridging the study of topology with network neuroscience using the neuronal avalanches framework.
Marcus: It's built on a modular architecture where each stage, like avalanche detection or ATM computation, has its own pipeline object.
Yuki: The package supports both continuous resting-state data and epoched task-related data which can be given as NumPy arrays or MNE-Python objects.
Ines: They demonstrate the utility through a group-level analysis on a public resting-state EEG dataset comparing propagation patterns across different clinical populations.
Marcus: The primary output for group-level analysis is the group-averaged ATM for each condition, visualized as a heatmap for direct comparison of large-scale spatiotemporal propagation patterns.
Yuki: The authors mention that the package’s integration with scikit-learn allows those ATM derived features to be fed into machine learning classifiers.
Ines: They show they are enabling the development of novel avalanche-based biomarkers and facilitating reproducible research for a broader scientific community by making complex analysis accessible.
Marcus: What about the specific mechanics of the ATM computation? How are they characterizing propagation using both methods mentioned?
Yuki: They use two distinct methods for computing the ATMs, which are the standard conditional probability model and a weighted-stochastic formulation.
Ines: And then they use graph-theoretic metrics directly from those ATMs with libraries like Brain Connectivity Toolbox and NetworkX to study the topology.
Marcus: That sounds like a solid bridge between dynamics and network structure, but what kind of specific metrics are they extracting from those matrices?
Yuki: They focus on graph-theoretic metrics derived from the ATMs that characterize the propagation of activity during avalanches.
Ines: The authors show that this method resolves real structural topology in developing brain tissue recordings through persistent homology.
Marcus: How does that persistence homology connect back to what we discussed earlier about loop structures and voids?
Yuki: It shows that the integrated first Betti number summarizes the loop structure across the whole filtration, and in fourteen of eighteen datasets, this exceeds null model predictions at a significance level less than or equal to zero point zero five.
Ines: So even when local measures miss loops, they are present because of a strong-correlation backbone rather than dense connections.
Marcus: And the structure is carried by an identifiable core of units instead of being spread redundantly across the population, which suggests these topological features are robust.
Yuki: They note that higher-order structure like enclosed voids only appears when the network size reaches at least one hundred units before they emerge.
Ines: This indicates that loops are the first level of complexity before voids show up, and this topological signature is functional within the self-organizing tissue.
Marcus: That connects back to schema-based active inference which suggests abstract relational structures underlie rapid generalization in cognition.
Yuki: Schema-based active inference supports rapid generalization by combining predictive processing with schema-based mechanisms to model how humans learn and adapt.
Ines: So, this paper is providing a concrete, reproducible tool for analyzing neural activity patterns that reveal underlying structure without needing to assume variables beforehand.
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