Transformers in protein informatics: A survey of structure prediction, function annotation, interactions, and drug discovery

Protein informatics increasingly requires artificial intelligence models for analyzing biological sequences, structures, interactions, and drug-related targets. Transformer architectures, built on self-attention mechanisms, have become widely used in this field, but their reliability remains dependent on data coverage, biological context, and evaluation protocol. This survey critically reviews over 100 representative studies on Transformer-based methods for protein research. We organize the literature through a two-level taxonomy that separates model-class principles from problem-space applications, covering sequence-only language modeling, geometry-aware prediction, multimodal alignment, and generative design, together with structure prediction, function annotation, interaction analysis, drug discovery, and drug-target identification. We examine how pre-trained protein language models, structure-aware Transformers, and multimodal variants incorporate sequences, evolutionary profiles, and three-dimensional coordinates to address protein-specific constraints. We further synthesize key methodological issues, including geometric inductive bias, physical plausibility, scalability, uncertainty calibration, dataset bias, reproducibility, and generative-model evaluation. Through comparative tables and cross-domain analysis, this survey connects protein Transformer architectures to automated protein engineering, high-throughput drug screening, proteome-scale annotation, and laboratory-in-the-loop design. It evaluates these systems not only by predictive accuracy, but also by computational cost, latency, calibration, reproducibility, physical validation, and experimental utility, and concludes with a prioritized roadmap for deployable protein artificial intelligence.