BioWorldModel: A Multi-Kingdom Trajectory Architecture for Genomic Prediction with Evolutionary Curriculum Learning

Abstract

Genomic prediction models are trained on single species and ignore temporal dynamics-- assumptions that limit their biological scope. Here I present BioWorldModel, a unified architecture that predicts multi-trait phenotypic distributions across fungi, plants, and animals with a single set of parameters. The model introduces a scalable genotype encoder with organism-conditioned attention pooling, a four-channel biological memory system, and a Gaussian output head with diagonal variance parameterization. Trained jointly on five organisms spanning three kingdoms (S. cerevisiae, A. thaliana, D. melanogaster, O. sativa, Z. mays; 641 traits total), the model achieves organism-averaged R2 = 0.821 (trait-weighted R2 = 0.413) and substantially outperforms GBLUP, BayesB, Lasso, and Random Forest baselines trained per organism. These results demonstrate that genotype-to-phenotype mapping follows shared principles across kingdoms that a single model can exploit.