Adapting TabPFN for Zero-Inflated Metagenomic Data

This paper introduces a novel prior assumption for TabPFN—a meta-learning method designed to approximate Bayesian inference on synthetic datasets generated from a predefined prior—aimed at better accommodating the unique zero-inflated distributions characteristic of metagenomic data. We modify the model's prior assumptions without changing its architecture by generating synthetic training data replicating the sparsity and variability inherent in these datasets. Preliminary results from metagenomic classification tasks show significant improvements in predictive performance, exceeding that of the original TabPFN and competing with state-of-the-art methods. This work emphasizes the necessity of tailoring PFN priors to align with the specific statistical properties of biomedical data, thereby enhancing their effectiveness in precision medicine.