Trion-1: a diffusion language model that handles both natural language, sequence and protein structure

Folding, reverse design, understanding: one model covers six tasks and reaches a TM-score greater than 0.8 with one sample.

Researchers present Trion-1, described as the first natively multimodal diffusion language model for proteins. It puts amino acid sequence, three-dimensional structure and natural language in a single token vocabulary, and abandons strictly left-to-right generation in favor of bidirectional interactions between tokens, which are better suited to long-distance contacts that shape folding.

This choice allows him to cover six abilities with a single model: predict the structure from the sequence, find a sequence for a given structure, understand one or the other, and design new sequences or structures, while maintaining the conversational skills of a generalist LLM.

Competitive results

In folding, under strict homology control, Trion-1 exceeds a TM-score of 0.8 with a single sample, whereas the tri-modal reference model requires 200 for a comparable result. In sequence comprehension, it is on par with state-of-the-art LLMs such as Claude Opus 4.7 (62.37%).