Back

Dynamically Assembling Biological Intelligence to Predict Novel Cellular Phenotypes

Pu, H.; Long, Y.

2025-10-30 bioengineering
10.1101/2025.10.28.685160 bioRxiv
Show abstract

In this work, we introduce Bio-AMLM (Biological Adaptive Modular Learning Model), a new framework designed to address out-of-distribution (OOD) generalization challenges in predicting cellular responses. Unlike monolithic deep learning models or simple data retrieval methods, which struggle to predict the effects of novel genetic or chemical perturbations, Bio-AMLM dynamically constructs a bespoke analytical pipeline for each biological query. It leverages a library of pre-trained, functionally specialized biological modules (e.g., for genomic, proteomic, and metabolic analysis). Guided by a biological context encoder, an adaptive inference planner selects, configures, and links these modules to form an optimal analysis chain. In experiments on several challenging bio-simulation benchmarks, including Gene-Edit-Bench, Drug-Response-Bench, and Toxicity-Bench, Bio-AMLM consistently outperformed state-of-the-art approaches, producing more reliable, robust, and interpretable predictions of cellular behavior in complex OOD scenarios.

Matching journals

The top 7 journals account for 50% of the predicted probability mass.

50% of probability mass above

"Similar papers" are the closest papers from that journal in the model's embedding space. They show what the match is built on, but the ranking comes mostly from a classifier over the whole training set, not from these examples alone.