ProteInfer Info#

Article Information

Oct 24, 2023

5 min read

On this page, we will show and explain the use of ProteInfer for enzyme function prediction. As well as document the BioLM API, and demonstrate no-code and code interfaces to enzyme function prediction.

Model Background#

Proteins exhibit vast diversity in sequences and functions. Homology-based approaches for functional prediction are inherently limited by availability of closely related sequences. ProteInfer, on the other hand, is able to learn patterns and relationships in protein sequences that are not based on homology, and it has been shown to be effective in predicting the function of proteins with limited homology to known sequences.

“Here we introduce ProteInfer, which instead employs deep convolutional neural networks to directly predict a variety of protein functions; Enzyme Commission (EC) numbers and Gene Ontology (GO) terms – directly from an unaligned amino acid sequence.” -Sanderson et al., 2023

The model uses a deep neural network with special convolutional layers (dilated convolutions) to process one-hot encoded protein sequences. The architecture allows the model to capture both local and global hierarchical features of the sequences, and through a series of transformations, including mean-pooling and passing through a fully connected layer, the model outputs probabilities for different functional classifications of the proteins. This architecture enables the model to make nuanced predictions about protein functions based on their amino acid sequences.

ProteInfer implements dilated convolutional layers to extract hierarchical local and global features from one-hot encoded input sequences. Through progressive transformations, including mean-pooling and fully connected layers, ProteInfer produces probabilistic predictions for enzyme commission numbers and gene ontology terms. This architecture enables nuanced modeling of sequence-function relationships beyond homology.

A key component is ProteInfer EC, which predicts enzyme commission numbers from sequence. These standard codes classify enzyme-catalyzed reactions, enabling systematic identification of enzymes and functions. By predicting EC numbers, ProteInfer provides insights into the catalytic reactions and enzymatic roles of proteins, which is crucial for elucidating biological systems.

Comparisons reveal ProteInfer has higher precision while BLASTp alignment shows greater recall. An ensemble approach combining both methods improves overall performance, synergistically integrating strengths of alignment-based homology detection and deep neural network sequence modeling, particularly for challenging remote homology scenarios (for example, the dataset clustered based on UniRef50).

ProteInfer utilizes deep dilated convolutional neural networks to model mappings between full-length protein sequences and functional labels. As described by Sanderson et al. (2023), ProteInfer models were trained on high-quality Swiss-Prot entries within UniProtKB (UniProt Consortium:, representing a well-curated subset of the known protein universe. Swiss-Prot contains 570,157 expertly annotated sequences from 294,587 unique references, totaling 206 million amino acids. Within UniProtKB, protein functions are captured via cross-references to ontologies like Enzyme Commission (EC) numbers, denoting enzymatic activity, and Gene Ontology (GO) terms, describing molecular function, biological process, and subcellular localization. By linking to standardized ontologies, UniProt systematically associates proteins with functional descriptors.

The ProteInfer enzyme function predictor uses a deep neural network to predict EC numbers from sequence. Proteins may have multiple EC numbers mapping to over 8,000 classified reactions (EC-IUBMB, ExplorEnz, BRENDA databases). The optimized 5-block convolutional model achieves a maximum F1 score of 0.977 on randomly split test data, correctly predicting 96.7% of EC labels with a 1.4% false positive rate, indicating reliable EC number prediction from sequence alone. Performance was relatively consistent across EC classes, with minor variations in F1 scores between categories like ligases and oxidoreductases. Precision exceeded recall at optimal thresholds, suggesting accurate positive predictions but difficulty capturing all functional associations. Varying confidence thresholds enables balancing precision and recall based on use case. While improvements remain possible, ProteInfer EC exhibits robust sequence-based EC prediction that could enable high-throughput annotation of uncharacterized proteins.

Applications of ProteInfer EC#

By linking protein sequence to catalytic function, ProteInfer EC can provide useful insights to guide rational design and accelerate characterization of engineered enzymes.

  • Predicting function of engineered enzymes

  • Guiding site-directed mutagenesis

  • Assessing fitness landscapes

  • Drug discovery

  • Systems and synthetic Biology

ProteInfer is adept at identifying regions within a protein sequence that are pivotal for specific reactions. This facilitates the understanding of functional correlations in multi-domain enzymes by bridging sequence attributes to functional outcomes. A specific protein, “fol1” from Saccharomyces cerevisiae, which is not included in the training data, is highlighted as an important example due to its multiple domains that each perform different roles in tetrahydrofolate synthesis. The model predicts these regions as being highly involved or essential in carrying out certain reactions or functions of the protein. These predicted regions align with existing scientific knowledge.

The ProteInfer EC model enables prediction of conditional enzyme activity by identifying sequence motifs and features associated with activity under different conditions. For example, motifs present in thermophilic enzymes may indicate thermostability if also found in the query sequence. Identified similarities and differences in sequence could reveal structural factors modulating activity. By leveraging ProteInfer EC’s learned sequence representations, researchers can elucidate sequence-function relationships and patterns that determine an enzyme’s conditional activity in varying contexts.

Applications of ProteInfer GO#

ProteInfer GO effectively predicts Gene Ontology terms directly from protein sequence using a powerful multi-label transformer classifier tailored for functional annotation applications. ProteInfer GO predictions can potentially aid genome annotation, protein characterization, system biology, engineering, and biomedical applications involving analyzing protein and gene function.

  • Protein function prediction - Predicting GO terms can help annotate the molecular functions, biological processes, and cellular locations of uncharacterized proteins. This can aid discovery of the roles proteins play in biological systems.

  • Protein interaction prediction - Knowing the GO terms for proteins can help predict if proteins may interact based on if they share similar functions and processes. This can guide experiments to validate interactions.

  • Functional Annotation: ProteInfer GO can be used to annotate novel or uncharacterized protein sequences with functional labels. This can be particularly helpful in large-scale genomics projects where a myriad of new sequences are generated.


  • The BioLM API allows scientists to programmatically interact with ProteInfer EC, making it easier to integrate the model into their scientific workflows. The API accelerates workflow, allows for customization, and is designed to be highly scalable.

  • Our unique API UI Chat allows users to interact with our API and access multiple language models without the need to code!

  • The benefit of having access to multiple GPUs is parallel processing. Each GPU can handle a different protein folding simulation, allowing for folding dozens of proteins in parallel!