Advancing generative large language models toward discriminative performance in protein function prediction

While generative large language models (LLMs) have revolutionized diverse research domains through their advanced semantic understanding capabilities, their applications to protein function prediction remain limited. Although significant efforts have been made to develop biological-knowledge-integrated LLMs, current approaches primarily focus on benchmarking their performances against general-purpose foundation models (e.g., ChatGPT-4o, DeepSeek-v3) rather than addressing their substantial performance gaps compared to specialized discriminative models (e.g., ESM2, ProtT5-based models).

We introduce OPUS-PLLM, a multitask generative LLM that establishes a sequence-to-function paradigm through natural language generation. The model integrates three components: modality encoding, modality refinement, and instruction tuning. To support its training, we construct two datasets, OPUS-InstructionCorpus and OPUS-InstructionCorpus-Evol, covering six protein functional annotations. The evaluations across five core protein function prediction tasks (spanning 18 benchmarks) demonstrate that OPUS-PLLM not only outperforms existing biological-knowledge-integrated LLMs but also surpasses specialized discriminative models in most cases.

Our results highlight the unexplored potential of LLMs in protein function prediction and provide a robust, scalable, and generalizable solution for developing biological LLMs.