Bio papers — 2026-09-18

Today's focus is on the limitations we found when trying to predict how patients will respond to immune checkpoint inhibitors. We tested nine different transcriptomic predictors, including five bulk RNA-seq models and four single-cell RNA sequencing based models, using independent datasets that hadn't been used during their development. The results showed that the predictive power was quite modest; the bulk RNA-seq models performed at or near chance level across most groups, and the scRNA-seq models only showed slight gains.

When looking closer at what worked, we saw that single-cell methods did manage to find some immune-related themes, specifically around allograft rejection. However, the bulk RNA-seq models did not show any of that consistent overlap in their findings. PRECISE and NetBio were the two scRNA-seq predictors that identified the most coherent immune programs. IRNet was mainly focused on metabolic pathways that were not strongly linked to ICI biology.

This suggests that current models lack cross-cohort robustness and biological consistency. This lack of consistency points toward a real need for better domain adaptation and more standardized preprocessing steps in these types of predictive modeling efforts.

Today's papers

Transcriptomic Models for Immunotherapy Response Prediction Show Limited Cross-cohort Generalisability Immune checkpoint inhibitors have limited predictive power across different patient groups. [paper]
Large Language Model Agents for Evidence Based Genetic Disease Severity Classification An AI agent uses reasoning and retrieval to classify genetic disease severity based on medical guidelines. [paper]
Self-Replicating Neural Cellular Automata Quantifying Emergent Phenotypic and Genotypic Diversity in an OpenEnded Substrate This study models how self-replicating agents create diversity based on their internal network weights. [paper]
Neural Langevin Machine a local asymmetric learning rule can be creative This paper introduces a generative model that uses local neural signals to learn data by relaxing to fixed points of recurrent networks. [paper]
RAG-GNN Integrating Retrieved Knowledge with Graph Neural Networks for Precision Medicine This framework combines knowledge retrieval with graph neural networks to improve functional clustering in cancer signaling studies. [paper]
A Mathematical Model of Motivated Emotional Mind Cognitive Embodied System This paper presents a mathematical model describing how embodied systems maintain homeostasis through motivated learning based on internal needs and affect. [paper]

The papers

Important terms

Immune Checkpoint Inhibitors (ICIs): These are drugs used to treat certain cancers by blocking proteins that help immune cells evade cancer detection. The research focuses on predicting patient response to these treatments.
Transcriptomic Predictors: These are models, like RNA-seq analyses, used to predict how a patient will react to an ICI based on their gene expression data. They were tested using independent datasets.
Bulk RNA-seq vs. scRNA-seq: Bulk RNA-seq looks at the average gene activity across many cells in a tissue sample, while single-cell (scRNA-seq) looks at individual cells. The study compared the predictive power of both methods.
Cross-cohort Robustness: This refers to whether a predictive model works consistently across different patient groups or datasets. The findings suggest current models lack this consistency.
Domain Adaptation: This is a technique needed to improve models by making them work well when applied to new, unseen data environments. It's suggested as a necessary next step for these predictors.