Daily Summary for 2026-11-07
daily
Transcript
Introduction to the show: ident: Genomics Radio. Generated commentary on the latest computational biology and genomics papers.
Ines: It's the seventh of November, twenty twenty-six, and this is the day's research.
Marcus: 13 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 Genome-Factory: A Library for Tuning, Deploying, and Interpreting Genomic Foundation Models.
The day's research: Ines: Genome-Factory simplifies the entire workflow from sequence collection through deployment.
Marcus: That library unifies data collection and fine-tuning for various model types.
Yuki: The paper Genome-Factory: A Library for Tuning, Deploying, and Interpreting Genomic Foundation Models covers this.
Ines: Emergent generalization suggests forcing a recurrent network to learn low-dimensional representations helps with out-of-distribution data.
Marcus: That connects to adaptation through cumulative selection, which goes beyond simple Darwinian evolution.
Yuki: Adaptation by Cumulative Selection proposes a framework where biological systems adapt via clonal organisms.
Ines: The Informational Model of the Holobiont partitions evolutionary change into host and microbial contributions using information theory.
Marcus: That paper develops a statistical model to test selection on host versus microbial components in holobionts.
Yuki: It offers a more granular view than some other modeling efforts for evolutionary change.
Ines: Shifts in environmental conditions alter trajectories when both optimal trait value and diffusion variance change simultaneously.
Marcus: That occurs under an Ornstein-Uhlenbeck model, moving beyond tracking just the ideal trait changes.
Yuki: Detection of evolutionary shifts in variance under an Ornstein-Uhlenbeck model uses an L1-penalized loss function for this.
Ines: That method shows improved predictive ability and reduces false positives when optimal value shifts occur with variance shifts.
Marcus: It performs comparably to existing tools when only the optimal value changes, which is a key comparison point.
Yuki: Another study explores how phenotype distributions control Turing instability by affecting the cross-diffusion coefficient.
Ines: On a phenotype-structured Shigesada--Kawasaki--Teramoto model: Turing instability and pattern selection under fast phenotype switching shows this effect.
Marcus: The low dimensional dynamics of a sparse balanced synaptic network uses circular cumulant reductions to model macroscopic dynamics.
Yuki: That explains why certain complex neural network regimes do not conform to simpler noise models.
Ines: One Method Does Not Fit All benchmarks 16 imputation methods across datasets, emphasizing task-specific evaluation is needed.
Marcus: That paper provides guidance on selecting imputation methods tailored to specific scRNA-seq analyses goals.
Lucky paper: 2509.12266: Ines: Okay, so we're looking at Genome-Factory, this library that aims to unify the entire genomic foundation model lifecycle from data collection to interpretation.
Marcus: It has six modular components—the Collector, Model Loader, Trainer, Inference Engine, Benchmarker, and Biological Interpreter—to handle all those steps.
Yuki: The trainer part is interesting because it supports full fine-tuning alongside parameter-efficient methods like low-rank adaptation or LoRA.
Ines: And the experiments show that the adapter method consistently uses less trainable parameters, which means it consumes less memory and gets higher throughput than full fine-tuning for most models.
Marcus: For instance, the DNABERT-two adapter uses only zero point two million trainable parameters and reaches a throughput of one hundred twenty-four thousand tokens per second.
Yuki: That efficiency gain is significant as the model size grows, especially with things like Nucleotide Transformer where LoRA tuning reduces the parameter ratio.
Ines: The Benchmarker module compares six genomic foundation models across three fine-tuning strategies on two benchmark suites to show how useful that module is for comparing performance.
Marcus: The results indicate that GenomeOcean-100M achieves the strongest performance under full fine-tuning on the GUE benchmark, while DNABERT-two performs best on Genomic Benchmarks.
Yuki: That's helpful context when you're looking at which model to use for a specific task.
Ines: The Biological Interpreter module lets us link internal model representations directly to biological features using a sparse autoencoder workflow.
Marcus: Analysis of the sparse autoencoder latent dimensions reveals that latent three hundred eighty-two latent five hundred nineteen and latent three thousand five hundred nineteen exhibit strong correlations with sequence length.
Yuki: So it moves beyond just black-box prediction by linking those specific latent features to measurable genomic properties.
Ines: The framework offers a zero-code command line interface and a Gradio based web user interface to lower the technical barrier for users in genomics research.
Marcus: They also provide task specific dataset generation tools for things like promoter region datasets and enhancer region datasets via YAML configuration files.
Yuki: I do have to point out, though, that the limitations mention the high computational cost of full fine-tuning for very large models like EVO-7B.
Ines: And they also flag that the context truncation mechanism might remove biologically meaningful elements in long range benchmarks, so future support for sliding window evaluation is needed.
Marcus: They also didn't systematically integrate classical methods like PWM scanning and HMM based motif finding into the framework yet, which is a gap for some researchers.
Yuki: That leaves an area where they haven't fully bridged the gap between modern AI and established computational biology techniques.
Lucky paper: 2602.20883: Ines: So we’re looking at Adaptation by Cumulative Selection, which is a framework that goes beyond traditional Darwinian evolution by including things like holobionts and long-lived clonal organisms.
Marcus: The paper defines cumulative selection as a random search process generating adaptation in a reasonable time through the accumulation of small changes.
Yuki: It sets three general conditions for this process to work: first, the features must vary randomly over time; second, different states need different rates of persistence; and third, the rate of variation cannot exceed an error threshold.
Ines: That sounds like a recipe for how selection happens without a full population structure.
Marcus: The paper formalizes this model using something based on Dawkins' work and Neander's work from 1995b.
Yuki: The core concept involves generating N new sequences by copying the current best state with probabilities related to mutation rates nu and mu.
Ines: How does that relate to the expected time to discovery T?
Marcus: The expected time T for the unordered case is approximately (ln L plus gamma) times p, where gamma is the Euler-Mascheroni constant.
Yuki: That gives us a concrete mathematical expression for how fast adaptation can occur under these assumptions.
Ines: If we look at biological systems that adapt without reproduction, like those holobionts mentioned in Adaptation by Cumulative Selection, how does this framework fit into their history?
Marcus: The authors use this cumulative selection recipe to provide an overarching theoretical framework that encompasses standard Darwinian evolution alongside other adaptation processes.
Yuki: It suggests that the conditions for cumulative selection are both necessary and sufficient for a system to adapt in a certain way.
Ines: So, it's not just about mutation and copying; it’s about those three specific conditions working together.
Marcus: The paper connects those three conditions directly to phenotypic variation, differential persistence, and the correlation between parents and offspring which is memory.
Yuki: That memory element is what links the system to past states in a way that drives adaptation forward.
Ines: What does this mean for us when we look at how complex biological interactions are modeled, like in The Informational Model of the Holobiont?
Marcus: That paper uses information theory to partition evolutionary change into contributions from host types and microbial components.
Yuki: It develops a statistical model specifically to test which parts of an evolutionary change are due to selection on the host versus selection on the microbial components.
Ines: So, it offers a way to get a more granular view than some other modeling efforts when studying these complex relationships.
Marcus: The paper suggests that this statistical partitioning helps clarify the selective pressures acting in different parts of an evolving holobiont.
Yuki: It’s providing a statistical test for selection and an extension to a theory of variable interactions.
Ines: This is really interesting because it moves away from just seeing one big force driving evolution.
Marcus: The paper shows that the model constructed in Section two performs cumulative selection specifically under those three stated assumptions.
Yuki: So, the framework itself is robust enough to cover both standard Darwinian evolution and these other adaptation processes.
Ines: It seems like this paper gives us a strong theoretical backbone for understanding how systems can evolve without relying on traditional population dynamics.
Lucky paper: 2608.23504: Ines: So we’re diving into The Informational Model of the Holobiont today. It reviews and generalizes an evolutionary information-theoretic model for holobionts.
Marcus: It uses Jeffreys divergence to partition informational change into contributions from the host, microbial components, and host-microbiome associations.
Yuki: The core formulation defines fitness using log W i j equals log W H i plus log W M j plus log a i j.
Ines: And selection changes the host and microbial marginal distributions through effective marginal fitnesses WH and WM, right?
Marcus: Exactly, and the association structure is captured by Jassoc through the post-selection association factor a i j.
Yuki: For an aggregate-microbiome model, the statistic n J T asymptotically follows a chi squared distribution with K H K M minus one degrees of freedom.
Ines: That gives us a concrete way to test whether observed informational change is due to selection on host types or microbial states.
Marcus: The structured multicomponent extension uses nu J equals K H K M minus one degrees of freedom, where nu J is the total number of possible microbiome configurations.
Yuki: It also introduces the Theory of Variable Interactions or TVI, which allows informational change to be partitioned recursively across arbitrary organizational levels.
Ines: TVI lets you see how units at one level can themselves consist of multicomponent units whose associations can change further up the hierarchy.
Marcus: The paper shows that this hierarchical formulation allows you to decompose realized relative fitness into changes in marginal frequencies and association factors at distinct organizational levels.
Yuki: It recovers previous models by setting L equal to one which means each higher-level observational unit contains a single multicomponent group.
Ines: So, the significance is that this framework provides a direct way to identify the hierarchical source of informational effects generated by selection.
Marcus: It separates the informational contribution into host and microbial marginal distributions and changes in their association structure, making them analytically separable and testable across levels.
Yuki: That separation means you can see exactly what part of the evolutionary change is due to host evolution versus microbial evolution or their interaction.
Lucky paper: 2411.10424: Ines: So we're looking at another paper today, and this one is about using sequence data to study spatial scales of interactions driving the spread of Highly Pathogenic Avian Influenza in Great Britain.
Marcus: It combines infection and bird population data using Random Forest Decision-Tree models to quantify relationships between spatially aggregated values and infection risk.
Yuki: The authors propose a different approach by directly exploiting the relationship between genetic distance, which they say is highly correlated with the transmission distance between infected individuals.
Ines: They determine the scales of aggregation by assuming the best scale is the one that provides the most important aggregated variable for genetic distance determination with highest importance score and results in the best model fit with highest R squared.
Marcus: The key finding there is that period dependent factors were most informative; for period one game bird abundance and wild bird abundance were strong predictors of genetic distance between samples.
Yuki: That suggests a bridging role between observed samples with significant undetected transmission among them during that first period.
Ines: But then they look at the later period, from June to October two thousand twenty-two where farm count played an important role, and areas of higher abundance of gulls and corvids indicated much shorter genetic connections.
Marcus: It seems like the relevance of the spatial scale changes depending on the time period you are looking at.
Yuki: The authors use variables describing HPAI H5N1 cases, farm counts as of December two thousand twenty-two and wild bird abundance from eBird for year two thousand twenty-one in this study.
Ines: They also integrate sequences from December 3rd, two thousand twenty to October 19th, two thousand twenty-two with all positive cases from that same period.
Marcus: The methodology involved assigning grid areas considering resolutions from H to XXL and using the two thousand eleven Census Output Areas for each case.
Yuki: They also implemented a temporal distance cutoff excluding pairs sampled over thirty days from each other because those are likely due to the time of emergence.
Ines: That selection threshold was determined by temporal distance optimization aiming for the lowest importance without compromising the model fit, which they show on Figure three.
Marcus: It’s interesting that they also used a Tamura and Nei one thousand nine hundred ninety-three model to calculate the genetic distance between paired sequences.
Yuki: The conclusion is that the variability of scale across risk factors and time periods illustrates why considering the relevant spatial scale is important for understanding transmission in complex disease ecosystems.
Ines: So, if you're looking at avian influenza spread, you can't just pick one fixed spatial unit; you need to consider what matters for that specific time and context.
Marcus: It gives researchers a method to study geographical scales of complex pathogen-host-environmental systems by combining sequence data with many variables reported in spatial resolution.
Yuki: That framework could be used to define targeted interventions or risk maps for HPAI spread, which is what the paper hopes for in the future.
Lucky paper: 2606.02937: Ines: We're looking at BEASTthree dee, which is this self-supervised pretraining framework that learns three dee visual representations from unlabeled multi-view animal behavior videos.
Marcus: The authors say it reconstructs three dee structure with as few as four views by conditioning directly on known camera parameters, which is a big deal compared to general models that have to estimate geometry.
Yuki: That's because BEASTthree dee uses image tokenization from a frozen DINOv3 ViT-B/sixteen as the image encoder and fuses those with camera tokens via a two layer multi layer perceptron before processing them in a geometry transformer.
Ines: So, the training loss combines three terms: a photometric loss for pixel differences, a perceptual loss for abstract feature differences, and a mask loss between the rendered alpha channel and SAM3 segmentation masks.
Marcus: That specific combination of losses is how they train it using the AdamW optimizer with a base learning rate of five times ten to the minus five and weight decay of zero point zero five.
Yuki: And during each forward pass, they randomly sample Vref equal to V minus one reference views and hold out Vtgt equal to one target view for training.
Ines: Now, when we look at the downstream tasks, they found BEASTthree dee produces sharp reconstructions of held-out viewpoints for novel view synthesis across all four datasets.
Marcus: For pose estimation, they claim it achieves the best performance across nearly all datasets, with Rat7M being the exception where VGGT performs better.
Yuki: In neural encoding, they found that BEASTthree dee features outperform three dee keypoints for predicting neural activity in both Cheesethree dee and Chickadee datasets.
Ines: That means dense three dee representations capture behavioral information that sparse keypoints just discard, which is a significant difference.
Marcus: They compared it against E-RayZer and Pose Splatter, and BEASTthree dee actually outperforms both of those in novel view synthesis on the InD Test set across all datasets.
Yuki: But the comparative analysis also showed that multi-view models like VGGT and E-RayZer do not consistently outperform single view DINOv3 or BEAST baselines for pose estimation.
Ines: The inference compute cost is comparable to E-RayZer, with the DINOv3 frontend adding roughly a constant per view memory and latency overhead.
Marcus: A key advantage they point out is the ability to reconstruct three dee structure using as few as four views by conditioning directly on known camera parameters instead of estimating them.
Yuki: But they do flag that a current limitation is the compute requirement, because pretraining on a single dataset takes roughly thirty-two hours on eight A100 GPUs, which makes it out of reach for many labs.
Ines: The broader implication is that BEASTthree dee could democratize advanced video analysis capabilities by reducing the need for extensive human labeling while improving accuracy.
Marcus: But there are concerns about surveillance capabilities, bias across demographic groups, and resource inequality because of those high pretraining compute requirements.
Yuki: The evaluation spans four datasets: Cheesethree dee, Rat7M, Chickadee, and Human3 point 6M; the dataset construction involves five stages: cut, downsample, segment, assemble and resize to get the calibrated tuples.
Ines: For pose estimation on Cheesethree dee and Human3 point 6M datasets specifically, they constructed a pseudo labeled dataset from six hundred sixty-five instances where each keypoint is labeled in a subset of views based on visibility.
Marcus: They use k-means clustering in the space of flattened three dee keypoints for each session to select representative frames for training and testing.
Yuki: For neural encoding, they segment spike trains into non overlapping two point zero s windows and partition them into train, val or test splits.
Ines: The encoding model is intentionally lightweight and input agnostic so that it doesn't learn three dee structure to compensate for a weak input.
Marcus: They use the foreground filtered BEASTthree dee point cloud as a baseline when contrasting it with multi view triangulated three dee keypoints produced by supervised pose estimation models.
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