How can we effectively decode and understand the complex molecular language of proteins to unlock their functional secrets at scale?@biorxivpreprint @Westlake_Uni "Decoding the Molecular Language of Proteins with Evola" • Scientists have developed Evola, an 80 billion parameter frontier protein-language generative model that combines information from protein sequences, structures, and user queries to decode protein function. It's trained on an unprecedented AI-generated dataset of 546 million protein question-answer pairs and 150 billion word tokens, integrating Direct Preference Optimization (DPO) and Retrieval-Augmented Generation (RAG) to improve response quality. • While protein structure prediction has seen major breakthroughs with tools like AlphaFold, there remains a critical gap between structural determination and functional understanding, with fewer than half a million proteins having expert-curated functional annotations despite hundreds of millions of known sequences. Previous attempts at protein language models have been limited by training data of only 500,000 to 3 million protein-text pairs. • The model architecture includes SaProt-650M or SaProt-1.3B as protein encoders, sequence compressors with 6-8 layers, and Llama3 decoders (8B or 70B parameters). Training used DeepSpeed on 32 NVIDIA A100 GPUs for the 10B model and FSDP on 64 H800 GPUs for the 80B model. The training data combined Swiss-Prot annotations (16 million triples) and ProTrek-derived data (530 million triples). • Performance evaluation showed Evola-80B achieved a GPT score of 74.10 compared to 40.49 for Deepseek-v3 and 37.07 for GPT-4. On EC number prediction for novel proteins, Evola achieved 41.2% accuracy for exact matches and 66.8% for three-digit matches, comparable to specialized classification models. The model demonstrated strong performance across both general test sets and more challenging hard subsets with low sequence similarity to training data. Authors: Xibin Zhou , Chenchen Han , Yingqi Zhang , Jin Su , Kai Zhuang , Shiyu Jiang et. al @duguyuan Link:

@duguyuan That seems to work well. BLAST in FPbase and UniProt matches fluorescent proteins, but with 30-35% sequence identity.