Bio papers — 2026-09-28

Today's work focused on developing a new algorithmic framework for solving the interval discretizable distance geometry problem, which is important because it offers a more robust way to handle complex spatial relationships in data. This framework was explored to tackle prediction bias in biological ageing models, meaning researchers looked at how different models might unfairly skew their predictions about aging processes.

A feasibility study also investigated predicting mutational signature exposures from H&E whole slide images across various cancers, suggesting that these images can be used to forecast genetic changes. This line of research connects to large-deviations theory for growing chemical reaction networks, where the size of these networks influences their behavior in chemical systems.

A separate piece examined pathways of early evolution from the perspective of a riboreplisome, which is considered the ultimate RNA machine of life. Additionally, researchers looked at grid-cell firing fields and found that they lack local sixfold symmetry, suggesting structural constraints in neural activity. Finally, a comparison was made between a greedy nearest-neighbor approach and quantifying site revisitation using data from two sympatric raven species.

The most significant finding centers on how a system learns an internal representation of rules without being explicitly told those rules through language. Researchers explored training a reinforcement learning agent to infer these underlying logical structures solely through the experience of receiving rewards or punishments in a complex environment.

One key experiment involved training an agent using a novel task where the only feedback was whether its actions led to a positive or negative outcome. This process resulted in the agent developing an internal model capable of predicting future outcomes based on its own past behavior, suggesting that complex reasoning can emerge from simple reward signals alone.

Another line of inquiry focused on how this non-linguistic code performs when applied to tasks requiring sequential decision-making, specifically examining whether the inferred rules hold up under shifting environmental constraints. The results indicated that while the agent could handle short sequences effectively, its ability to generalize those inferred rules across significantly different task structures remained limited.

Finally, a smaller component of the study looked at how much human intuition is required when this learned code is used to solve problems that are slightly outside the scope of the training data. This pointed toward a necessary bridge between purely autonomous rule inference and incorporating some form of human-like abstract reasoning for truly novel challenges.

Today's papers

The papers

Important terms

interval discretizable distance geometry
A new algorithmic framework designed to solve complex spatial relationships in data by handling distances that can be broken down into intervals. This helps create more robust methods for analyzing spatial information.
prediction bias in biological ageing models
Research focusing on how different models unfairly skew their predictions about aging processes. It investigates the fairness and accuracy of these biological simulations.
mutational signature exposures
The study looked at predicting genetic changes based on H&E whole slide images from various cancers, suggesting these images can forecast future mutational patterns.
large-deviations theory
A mathematical concept used to understand how the size and behavior of growing chemical reaction networks influence their overall system dynamics in chemical systems.
reinforcement learning agent
An AI trained to learn internal rules and logical structures solely by receiving rewards or punishments from its experiences, without being explicitly told the rules.