Abstract

Large-scale protein analysis requires reliable pipelines capable of ingesting, enriching, and organizing data from public repositories such as UniProt. Existing systems like the Protein Information System (PIS) and FANTASIA have demonstrated the feasibility of these workflows at scale, but their monolithic worker design conflates infrastructure concerns (database sessions, queue management) with domain logic (sequence deduplication, metadata enrichment), making the codebase difficult to extend, test, and maintain.

PROTEA is a platform designed to address this structural debt through an incremental migration strategy. Rather than a complete rewrite, it introduces a clean separation of concerns: a typed Operation protocol encapsulates domain logic, a job queue (RabbitMQ) decouples HTTP ingestion from computation, and a two-session worker pattern ensures robust, auditable state transitions. A React/Next.js frontend provides real-time visibility into job progress through structured event logs.

The platform implements the full protein functional annotation pipeline: UniProt sequence ingestion, GO ontology and annotation loading, GPU-accelerated embedding computation (ESM-2, ESM-C, T5/ProstT5, Ankh), KNN-based GO term prediction with optional pairwise alignment and taxonomic features, CAFA-style temporal holdout evaluation (NK/LK/PK), and LightGBM re-ranking. A scoring engine and one-click annotation endpoint make the system accessible to researchers without machine-learning infrastructure expertise.

The platform is designed to accommodate continuous extension (new operations, new data sources, new models) without architectural regression. Computational efficiency is preserved at each migration step, with sequence deduplication by MD5 hash, cursor-based pagination, and exponential backoff against upstream rate limits.