LucaPCycle: A Dual-Channel Architecture for Phosphorus-Solubilizing Function Prediction Based on Protein Language Models

LucaPCycle is a dual-channel architecture for phosphorus-solubilizing function prediction based on protein language models. It combines raw sequences with the representation capabilities of large protein language models to identify the phosphorus-solubilizing function of protein sequences and subdivide them into 31 specific types. It is applied to large-scale metagenomic data mining, addressing the limitations of traditional culture methods and the challenges of metagenomic function identification.

Phosphorus is an essential element for life, but most phosphorus in soil is insoluble. Phosphorus-solubilizing microorganisms convert phosphorus by secreting organic acids, playing a key role in agriculture and ecological cycles. Traditional screening relies on culture methods, which are time-consuming and have limited coverage. Metagenomic technology generates massive data, but quickly and accurately identifying phosphorus-solubilizing functional proteins has become a new challenge—LucaPCycle was developed to address this.

The core is a dual-channel design: the raw sequence channel captures local patterns, and the protein language model channel (e.g., LucaProt) extracts deep semantic representations. The model type is LucaProt, with input as seq_matrix and a default truncation length of 4096. It includes an identification model (binary classification) and a fine-grained classification model (31 categories).

It uses a two-stage prediction process: the first stage uses binary classification to determine phosphorus-solubilizing function (dataset: extra_p_2_class_v2, default threshold: 0.2); the second stage performs 31-category classification on positive sequences (dataset: extra_p_31_class_v2, supporting topk output). The 31 types cover different phosphorus-solubilizing mechanisms such as organic acid secretion and phosphatases.

Applied to 164 metagenomes and 33 metatranscriptomes (total of over 150 million sequences), with samples from 16 global cold seep sites (depth: 0-68.55 meters, water depth: 860-3005 meters). It predicted over 1.48 million positive sequences and more than 130,000 potentially interesting findings, verified by ECOD, DeepFRI, and CLEAN to ensure reliability.

Three independent validation methods are used: ECOD domain analysis to check domain composition; DeepFRI v1.0.0 to identify functional residues; CLEAN v1.0.1 for enzyme annotation. Only results passing all three validations are marked as 'verified' to ensure annotation quality.

It facilitates research on the functions of uncultured microorganisms, accelerates the development of biofertilizers, helps understand phosphorus cycle mechanisms in extreme environments, provides a technical route reference for other functional annotation tasks, and demonstrates the application value of AI for Science in the field of microbial function prediction.

Applicable scenarios: Metagenomic/transcriptomic phosphorus-solubilizing gene mining, new genome annotation, preliminary filtering for microbial screening, etc. Notes: Combine with experimental validation, adjust thresholds to balance recall and precision, manage GPU memory. Summary: LucaPCycle achieves efficient and accurate phosphorus-solubilizing function prediction, addresses large-scale data challenges, and promotes the application of AI in life sciences.