PoET-2 Introduces Retrieval-Augmented Multimodal Protein Foundation Model

On Feb 26, 2026, researchers published PoET-2, a multimodal, retrieval-augmented protein foundation model that integrates family-specific in-context evolutionary retrieval with optional structure conditioning. PoET-2 employs a hierarchical equivariant transformer encoder and dual decoders supporting generative and bidirectional modes, achieving state-of-the-art zero-shot variant effect prediction and improved supervised embeddings, notably for multi-mutation and indel scoring on small datasets.