Bio papers — 2026-10-01

Today's focus is on using deep learning to find specific patterns in ATAC-seq peaks that can tell us the difference between wild-type and knockout samples, which is crucial for understanding how genetic changes affect chromatin accessibility. We are trying to build a WTKO-CNN model to do this, and the core idea involves applying belief-based maximum occupancy principles and active inference to interpret these complex epigenetic signals.

The work that matters most is the WTKO-CNN because it directly addresses distinguishing genetic states through sequence motifs, which is a key biological question. This model attempts to learn these distinctions by analyzing ATAC-seq data, and this deep learning approach builds upon earlier work like MerKurio, which extracts sequence information by matching k-mers to annotate the peaks.

We also looked at how covariance eigenspaces can provide latent attribution for longitudinal effects in brain age gaps, offering a way to understand temporal changes in neural data. This is less central than the main model, but it provides context on how dynamic systems behave. Finally, there is some more theoretical work exploring attraction to hierarchical feature memory and orientation bias, which touches on how features are learned in general neural systems.

The work on kinetic models predicting tau-catalyzed A beta nucleation is particularly important because it suggests a mechanism for how tau pathology can progress without simply increasing the visible plaque volume. This kinetic model shows that the burden of these toxic oligomers can rise significantly even when the physical plaques remain unchanged.

This finding connects to research into Ca 2+ tunable mechanics and recoil in reconstituted Tcb2 networks, which explores how mechanical forces within protein structures might influence their behavior. Furthermore, the study on life finding a way through emergent cooperative structures in adaptive threshold networks offers insights into how complex behaviors arise from simple local rules.

A kinetic model predicting that tau-catalyzed A beta nucleation can raise oligomer burden without increasing plaque volume is supported by work on Ca 2+ tunable mechanics and recoil in reconstituted Tcb2 networks, which investigates how mechanical forces within protein structures might influence their behavior. This structural insight complements the findings from the study on life finding a way through emergent cooperative structures in adaptive threshold networks, as both look at how systems organize themselves under specific physical constraints.

The most significant piece of work today involves the stage structured deterministic model of fall armyworm infestation on maize farming, which attempts to predict pest outbreaks based on environmental factors. This model uses a specific set of parameters derived from field observations to simulate how the pest population evolves across different growth stages of the maize crop.

This simulation framework is built upon a foundation that incorporates context modeling for single-cell representation learning, suggesting that understanding individual insect states is key to predicting population dynamics. Furthermore, the work on neural spikes as rare events provides a mechanism for capturing sudden, unpredictable jumps in infestation levels within this deterministic structure.

A related effort explored association profile conditioning in a set-temporal transformer for cross-session intracortical motor decoding, which seems to offer insights into how complex sequential data can be processed when trying to interpret rapid changes in behavior. This contrasts with the pest model by focusing on decoding motor commands from neural activity, but both share an interest in modeling temporal dependencies.

The work on cellMSA, which focuses on context modeling for single-cell representation learning, suggests a way to create rich representations of individual cells that could feed into more complex predictive systems like the pest model. This approach is foundational because it deals with how to capture meaningful information from single data points before scaling up to population predictions.

Today's papers

The papers

Important terms

WTKO-CNN
A deep learning model designed to distinguish between wild-type and knockout samples by analyzing ATAC-seq data. It uses sequence motifs to learn how genetic changes affect chromatin accessibility.
ATAC-seq peaks
Specific regions in the genome where chromatin is accessible, which are analyzed using ATAC-sequencing. These peaks show how genetic changes alter the epigenetic landscape.
Active Inference
A theoretical framework used to interpret complex biological signals. It helps explain how systems actively maintain their states by minimizing prediction errors.
Kinetic Model
A mathematical model predicting how biological processes, like tau pathology, progress over time. This helps understand mechanisms beyond simple physical changes.
Context Modeling for Single-Cell Representation Learning
A technique used to create rich representations of individual cells from single data points. This is foundational for scaling up predictions to whole populations.