Bio papers — 2026-10-05

The focus today is on building better models that predict how cells will react when nudged, which is crucial for controlling biology. Researchers are looking at using energy-guided flow matching to make these predictions more generalizable across different cell types. This approach tries to capture the underlying dynamics of a system by guiding a generative process toward physically meaningful states.

A related piece explored parameter uncertainty in dynamical models using an identifiability index, which helps us understand how well we can actually determine the parameters from the observed data. This is important because if we do not know our model parameters reliably, our predictions about cell response will be shaky.

Another approach looked at a likelihood-based framework that simultaneously learns both noise and growth dynamics using biologically-informed neural networks. This means training a network to understand not just how things grow, but also the inherent randomness in the biological process itself. This framework connects to reliable mechanistic operator recovery with biologically-informed neural networks, which provides principles for designing these architectures so they can accurately recover the underlying biological rules.

Finally, there is work on offline learning of prompt-conditioned interventions for cells and biobots using language models. This suggests a path toward controlling biology by using natural language prompts to guide experimental manipulations in living systems.

The causal discovery identifying pathways linking physical activity to dementia risk in the UK Biobank is particularly important because it provides tangible evidence connecting observable behaviors to long-term health outcomes. This work used causal discovery methods applied to UK Biobank data, specifically looking at how physical activity relates to dementia risk. It suggests that specific physical activities can be identified as pathways influencing cognitive decline.

Another significant piece of work involves contrastive neural embeddings which reveal individual traits beyond conversational role in language models. This method shows that these embeddings capture unique personal characteristics in a way that goes deeper than just what the model learns from dialogue. This finding complements the work on stimulus symmetries confounding representational similarity analyses because it suggests that even when looking at similarities between stimuli, inherent individual differences are still being captured by these embedding techniques.

On a more theoretical level, there is work on an infinite family of substrates satisfying all the postulates of integrated information theory with arbitrarily large integrated information. This means that the underlying physical systems capable of supporting complex computation are far more diverse than previously thought. This idea relates to clustering without clusters, which mistakes continuous dynamics for discrete states by proposing a meta-criterion and centroid reliability mistake to better handle these complex dynamics.

Today's papers

The papers

Important terms

Energy-guided flow matching
A method used to make predictions about cell reactions more generalizable across different cell types by guiding a generative process toward physically meaningful states.
Identifiability index
A tool that helps researchers understand how reliably they can determine the parameters of dynamical models from the data, ensuring prediction accuracy.
Likelihood-based framework
A learning approach that simultaneously trains neural networks to understand both the growth dynamics and the inherent randomness (noise) in biological processes.
Contrastive neural embeddings
Techniques that reveal unique personal traits beyond conversational roles in language models, capturing deep individual characteristics.
Causal discovery
Methods used to identify specific pathways linking physical activities to long-term health outcomes like dementia risk from large datasets.