ProtLingo: Efficient Protein Language Modeling via Conditional Memory and Expert Routing

On September 7, 2026, arXiv released a new efficient protein language model framework called ProtLingo. This model employs conditional local memory and sparse expert routing mechanisms to achieve performance comparable to existing methods while maintaining a backbone network of 150 million parameters. ProtLingo maps context residues into discrete codes of specific sequences, forms central local windows to generate potential N-gram contexts, and retrieves reusable residual signals related to repeated local sequence contexts; at the same time, some feedforward blocks are upgraded to shared and routed sparse hybrid expert layers, performing computation on dependent residues only when a subset of parameters is activated. In experiments such as protein adaptability prediction, FLIP benchmarks, and supervised contact prediction, the model demonstrated high parameter efficiency while retaining long-range structural representation capabilities, especially showing excellent performance in mutation effect prediction.