Deep learning tools such as convolutional neural networks (CNNs) and transformers have spurred great advancements in computational biology. However, existing methods are constrained architecturally in terms of context length, computational complexity, and model size. We present Janus, a novel subquadratic architecture for sequence modeling, and discover for the first time that state space models can efficiently learn polynomial functions—a property that makes them inherently suited for modeling epistatic interactions in biological sequences. By integrating this mathematical insight with projected gated convolutions, Janus achieves state-of-the-art performance across diverse tasks including protein fitness landscape prediction, RNA structure analysis, and CRISPR guide design. Notably, Janus does this with significant performance improvements, both in speed and orders-of-magnitude reduction in parameters compared to recent biology foundation models. This dramatic reduction in computational requirements enables training on consumer-grade hardware in hours rather than requiring specialized GPU clusters for weeks, democratizing access to sophisticated sequence analysis and positioning Janus as a promising tool for biological sequence modeling.

Not everyone has the computational resources to train large language models, especially biologists. My work is driven by the belief that creating efficient architectures rooted in biological principles can democratize AI insights for all.