Protein Function Prediction: Smart Decoding from Sequence to Function

As of June 2026, UniProtKB/Swiss-Prot (manually reviewed) contains 575,503 sequence entries, while UniProtKB/TrEMBL (computationally derived) contains 149,234,636 sequence entries. Manually reviewed entries account for only approximately 0.38% of the total. This means that over 99% of proteins lack experimentally validated complete functional annotations, while the Gene Ontology (GO) continues to add hundreds of new functional categories each year. The biological functions of the vast majority of known proteins remain unknown.

Bridging this sequence-function gap is the core mission of protein function prediction.

I. What Is Protein Function Prediction

Three Generations of Protein Function Prediction

Protein function prediction is a core task in bioinformatics, aiming to assign standardized functional labels to an amino acid sequence: which reaction it catalyzes (EC number), which biological processes it participates in (GO biological process), where it is localized within the cell (GO cellular component), and what molecular function it possesses (GO molecular function).

Protein function prediction holds foundational value for elucidating disease mechanisms, discovering drug targets, and empowering biotechnology development. Traditional functional annotation relies heavily on wet-lab experiments, which are time-consuming, costly, and limited in throughput. With the rapid advancement of high-throughput sequencing, newly discovered protein sequences are growing exponentially, and experimental validation lags far behind sequence accumulation.

This supply-demand imbalance continues to drive the iterative evolution of computational prediction methods.

II. Technological Evolution: From Homology Alignment to Multimodal Intelligent Reasoning

2.1 First Generation: Sequence Homology Alignment

Early function prediction relied heavily on sequence alignment tools such as BLAST. Based on the assumption that "sequence similarity often implies homologous evolution and thus functional conservation," unknown sequences were matched against annotated homologous sequences in databases to transfer functional annotations. This approach is reliable when sequence identity exceeds 40%, but accuracy drops significantly when dealing with distantly related homologs (sequence identity < 25%) or orphan proteins.

2.2 Second Generation: Deep Learning Classifiers

With the rise of deep learning, models such as DeepGO framed GO function prediction as a multi-label classification task, leveraging deep neural networks to learn the mapping between sequence features and functional labels. Subsequent DeepGO-SE further integrated protein language models and protein-protein interaction networks. Some studies proposed POSA-GO, which incorporates GO term topological structures into self-attention mechanisms to model GO hierarchical relationships.

However, the classifier paradigm has inherent limitations. Traditional classifier approaches often treat GO terms as independent multi-labels; although some models incorporate GO hierarchical structures, the closed-set supervised learning paradigm fundamentally struggles to adapt to the hundreds of new functional categories added each year. Once the GO ontology introduces novel functional labels not present in the training set, the traditional closed-set supervised learning paradigm becomes ineffective.

2.3 Third Generation: Multimodal Integrated Intelligent Reasoning

Protein Language Model & Structure Integration

Current frontiers have entered the stage of multimodal intelligent reasoning, with a growing consensus that the synergistic integration of sequence information and three-dimensional structural information is key to improving prediction accuracy. HiFusionPro employs a hierarchical fusion architecture to deeply integrate sequence and structural information, using protein language models to initialize structural encoders, enabling early feature sharing and aggregating multi-level intermediate representations to overcome the limitations of traditional late-stage feature concatenation.

RIMGOGraph converts three-dimensional structures predicted by structural prediction models into residue interaction graphs, jointly trained with protein language model embeddings to achieve structure-aware function prediction. StructSeq2GO similarly employs graph representation learning to extract structural features and integrates them with ProteinBERT sequence embeddings. Other studies have constructed multi-channel deep learning frameworks that achieve robust protein function prediction under extremely sparse functional annotation data.

ProtDML takes a different approach, leveraging distance metric learning to explicitly model relationships between functional labels, constructing annotation-aware feature spaces through similarity-weighted push-pull losses, achieving the lowest semantic distance across all ontologies on the CAFA3 benchmark.

III. Zero-Shot Prediction: Breaking the "Unseen Function" Challenge

Zero-Shot Prediction Workflow

Protein function prediction faces a major challenge: the long-tail distribution and continuous expansion of GO categories, with over 20,000 GO terms having fewer than 100 proteins with experimental annotations. Traditional supervised learning cannot handle entirely new functional categories never seen during training, making zero-shot function prediction an important paradigm for breaking this bottleneck.

COSMOS (Context-aware GO Subgraph Mining System) constructs a knowledge graph containing nearly 8 million functional semantic relationships, integrating protein functional annotations, GO hierarchical relationships, protein-protein interactions, and structural similarity information. Rather than using global graph embedding approaches, COSMOS extracts local subgraphs for the target protein's candidate functions, leveraging multi-relational message-passing neural networks to precisely capture subgraph topological logic.

In zero-shot prediction tasks (where training and test GO categories have no overlap), COSMOS achieves an Fmax of 0.915 on the molecular function branch, representing a 21.8% improvement over DeepGOGATSE; under few-shot and low-homology conditions, Fmax outperforms the optimal baseline by 13.8% and 19.9%, respectively, with similarly significant improvements on the more complex biological process branch. Ablation studies demonstrate that integrating protein-protein interaction and structural similarity information yields over 90% performance improvement for the model (data from COSMOS original paper's internal control experiments).

COSMOS demonstrates strong reasoning capabilities in real-world cases: for the SIRT6 protein, the model not only accurately predicted its canonical histone deacetylase activity but also identified the long-chain fatty acyl lysine deacylase activity that was experimentally confirmed after the training set's temporal cutoff, providing traceable subgraph reasoning paths as supporting evidence.

MZSGO proposes a multimodal zero-shot functional annotation framework centered on "semantic alignment." The model simultaneously encodes protein sequence features and domain-level textual descriptions, leveraging large language models to construct a unified semantic space for domain descriptions and GO labels, embedding disparate biological entities and functional concepts into a comparable vector space. In zero-shot scenarios, MZSGO achieves an overall harmonic mean H of 0.6669 on unseen labels, significantly outperforming comparable methods.

ProtNote leverages free-form text to support both supervised and zero-shot protein function prediction, maintaining high prediction accuracy on training set annotations while generalizing to entirely new functions not covered during training. STAR-GO employs a Transformer framework to jointly model GO term semantics and hierarchical structural features, further enhancing zero-shot reasoning capabilities.

IV. Evaluation Benchmarks: CAFA and ProteinGym

Protein function prediction models require standardized benchmarks for objective performance comparison.

CAFA (Critical Assessment of Function Annotation) is an authoritative international competition in the field of protein GO function annotation, specifically designed for protein function prediction tasks. InterLabelGO⁺ has demonstrated excellent performance in the CAFA series of challenges; ProtDML achieved the lowest semantic distance across all ontologies on the CAFA3 benchmark.

ProteinGym is a large-scale standardized benchmark suite primarily used for protein mutation fitness and mutation effect prediction, covering over 250 deep mutational scanning (DMS) experiments with millions of mutant sequences across diverse protein families and variant types. This benchmark does not directly address GO function annotation tasks but is primarily used to evaluate models for predicting functional changes caused by mutations.

These two benchmarks serve different task scenarios, providing a unified yardstick for different directions of AI protein models and continuously driving technological iteration in the field.

V. Industrial Applications: From Laboratory to Production Line

From Laboratory to Production

Protein function prediction technologies are accelerating their transition from academic research to industrial deployment.

In March 2025, Professor Hong Liang's team at Shanghai Jiao Tong University released the Venus series models. Unlike models focused on protein structure prediction, the Venus series targets the core objective of sequence-function prediction, learning the organizational rules of natural protein sequences and their intrinsic relationship with function, achieving top-tier accuracy in industrial-scale mutation function prediction.

The team established the protein sequence dataset Venus-Pod, containing nearly 9 billion protein sequences and billions of functional labels, making it the world's largest dataset in terms of both data volume and functional annotations, and four times the size of the 2.1 billion protein sequences used to train another industry-recognized model. Its composition includes: 3.62 billion terrestrial microbial protein sequences, 2.64 billion marine microbial protein sequences, 2.43 billion antibody protein sequences, and 0.06 billion viral protein sequences. The Venus series possesses two core capabilities: AI-directed evolution and AI enzyme mining. They have also developed the world's first high-throughput large-volume automated protein expression, purification, and functional testing system, capable of completing expression, purification, and testing of over 100 proteins continuously within 24 hours, achieving nearly a 10-fold improvement in overall efficiency compared to manual operation.

To date, multiple proteins designed by the Venus series models have achieved industrial deployment. In the single-domain antibody alkali resistance engineering project with Genescience Pharmaceutical, guided by the model, the alkali resistance of a conventional single-domain antibody was increased by 4-fold in less than a year. The platform has now developed dozens of protein products for numerous domestic and international companies, covering biopharmaceuticals, bioenergy, food processing, and other sectors, with multiple products achieving industrial application.

VI. Challenges and Outlook

Despite rapid progress, protein function prediction still faces multiple challenges:

Data sparsity and long-tail issues. Hundreds of new GO functional categories are added each year, and traditional supervised learning struggles to adapt to entirely new categories, making zero-shot generalization capability a key metric for assessing long-term model value.

Insufficient model interpretability. Deep learning models are often regarded as "black boxes," yet biological research requires that prediction results be traceable and verifiable. COSMOS's subgraph reasoning paths and MZSGO's semantic alignment mechanism represent frontier directions in interpretable functional annotation.

Computational predictions cannot replace wet-lab validation. All AI prediction results ultimately require experimental verification. However, the core value of AI lies in rapidly converging massive candidate sequences to the most promising targets, dramatically reducing the trial-and-error space in R&D.

Future technological trends are clear:

Multimodal fusion will continue to deepen—the deep integration of sequence, three-dimensional structure, protein-protein interactions, and literature text information will continuously enhance prediction accuracy;

Zero-shot reasoning capabilities will continue to push boundaries—models will possess the ability to recognize entirely new functional categories never seen during training;

Prediction and design integration—platforms will not only answer "what function does this protein have," but further guide "how to engineer the protein to achieve target performance."

From homology sequence alignment to multimodal intelligent reasoning, from passive annotation to active inference, protein function prediction is systematically decoding the vast "dark matter" of proteins. It builds a rapid pathway from sequence to function for billions of functionally unknown protein sequences, and lays a solid digital foundation for drug discovery, enzyme resource mining, and synthetic biology innovation.