Source-linked AI summary
DeepGO: Predicting protein functions from sequence and interactions using a deep ontology-aware classifier
Maxat Kulmanov, Mohammed Asif Khan, Robert Hoehndorf
TL;DR
Protein-function annotation must scale to rapidly growing sequence collections despite expensive experiments and a large, hierarchically structured Gene Ontology. DeepGO combines learned sequence and interaction-network representations with ontology-aware deep classification, improving over BLAST, particularly for cellular-location prediction.
Problem
Protein-function annotation is difficult to scale because experiments are costly, while the Gene Ontology presents over 40,000 interrelated functions and locations.
Method
DeepGO learns representations from protein sequences and interaction networks, then refines predictions through a deep neuro-symbolic model structured around Gene Ontology dependencies.
Results
DeepGO improves protein-function prediction over BLAST and performs particularly well for predicting cellular locations.
Takeaways & Limitations
The approach provides a data-driven alternative to manually crafted features for ontology-structured protein-function prediction.
Takeaways & Limitations
The model requires substantial computational resources for training and may not be applicable in all settings.
Abstract
from arXiv · showhide
A large number of protein sequences are becoming available through the application of novel high-throughput sequencing technologies. Experimental functional characterization of these proteins is time-consuming and expensive, and is often only done rigorously for few selected model organisms. Computational function prediction approaches have been suggested to fill this gap. The functions of proteins are classified using the Gene Ontology (GO), which contains over 40,000 classes. Additionally, proteins have multiple functions, making function prediction a large-scale, multi-class, multi-label problem. We have developed a novel method to predict protein function from sequence. We use deep learning to learn features from protein sequences as well as a cross-species protein-protein interaction network. Our approach specifically outputs information in the structure of the GO and utilizes the dependencies between GO classes as background information to construct a deep learning model. We evaluate our method using the standards established by the Computational Assessment of Function Annotation (CAFA) and demonstrate a significant improvement over baseline methods such as BLAST, with significant improvement for predicting cellular locations.
1 Introduction
The paper addresses protein function prediction amid rapidly growing sequence data and a complex output space. It presents a data-driven method combining sequence and interaction-based representations, improving over BLAST and performing particularly well for cellular-location prediction.
- Motivation: Rapidly increasing protein and genetic sequence data creates a need to identify protein functions computationally.Sequencing additional organisms and metagenomics are expected to increase available sequence data further.
- Challenges: Protein function prediction faces challenges from complex sequence–structure–function relations and a large, complex classification output space.Predicting protein structures with sufficient quality for function prediction still requires substantial effort.
- Approach: The proposed method learns protein features separately from sequence and interaction-network location using multiple layers of neural networks.These learned representations are used for protein function prediction.
- Results: The model improves function-prediction performance over a BLAST baseline and performs particularly well for predicting cellular locations.Its main advantage is an entirely data-driven approach that does not rely on manually crafted features.
2 Materials and Methods · 2.1 Datasets · 2.2 Training
The study uses a filtered, experimentally supported SwissProt dataset and the January 2016 Gene Ontology release. Training propagates annotations, splits proteins into training and testing sets, selects frequent GO terms, and predicts three ontology branches with binary labels.
- 2.1 Datasets: The January 2016 Gene Ontology release contains 44,683 classes, including 1,968 obsolete classes, across BP, MF, and CC branches.The branches contain 28,647 BP, 10,161 MF, and 3,907 CC classes.
- 2.1 Datasets: The starting SwissProt dataset contains 553,232 reviewed, manually annotated proteins, of which 525,931 have function annotations.Sequences and GO annotations were downloaded on 05 January 2016.
- 2.1 Datasets: Proteins were retained using experimental evidence codes, a maximum length of 1,002, and exclusion of ambiguous amino acid codes.The evidence codes were EXP, IDA, IPI, IMP, IGI, IEP, TAS, and IC.
- 2.1 Datasets: The final dataset contains 60,710 proteins annotated with 27,760 GO classes: 19,181 BP, 6,221 MF, and 2,358 CC.The filtering omitted proteins with ambiguous amino acid codes.
- 2.2 Training: Three models were trained, one for each GO sub-ontology, after propagating annotations through the GO ontology structure.Proteins were randomly split into 80% training and 20% testing sets.
- 2.2 Training: The models selected the top 932 BP, 589 MF, and 436 CC terms by annotation count because of computational limits and sparse specific-class annotations.These cutoffs correspond to minimum annotation counts of 250, 50, and 50 for BP, MF, and CC, respectively.
- 2.2 Training: Each protein received three binary label vectors, with 1 for an annotated selected term and 0 otherwise.Training and testing used proteins annotated with at least one term in the model’s selected set.
2.3 Data Representation · 2.4 Convolutional Neural Network
The model represents proteins as amino-acid trigram sequences mapped to dense embeddings, then applies one-dimensional convolution and temporal max-pooling to learn compact sequence features for classification. The CNN exploits sequential and local correlations, with the reported architecture using 32 filters and producing an 832-length pooled vector.
- 2.3 Data Representation: Protein inputs are amino-acid sequences whose trigrams are represented with dense embeddings to overcome the limited generalization of sparse one-hot encodings.A protein sequence uses 20 unique amino-acid codes; trigram one-hot vectors would have length 8,000.
- 2.3 Data Representation: A vocabulary assigns each unique amino-acid trigram a 1-based index, encoding sequences of length 1002 as 1000 indices.Shorter sequences are padded with zeros, while proteins longer than 1002 are ignored.
- 2.4 Convolutional Neural Network: CNNs use local connections and multiple filters to capture local correlations and learn multiple features from adjacent input regions.The method applies this convolutional approach to protein sequence data.
- 2.4 Convolutional Neural Network: The model uses one-dimensional convolution over protein sequence data to exploit sequential correlation.The convolution is defined for an input sequence and kernel with stride d.
- 2.4 Convolutional Neural Network: Convolution computes feature maps by summing convolutions between input components and filter components.The resulting output vector h represents the features learned through convolution.
- 2.4 Convolutional Neural Network: Temporal max-pooling discards redundant information by selecting maximum values over windows, producing higher-level sequence representations for classification.These pooled features are passed to fully connected layers.
- 2.4 Convolutional Neural Network: The experiments use one 1D convolution layer with 32 filters of size 128 and a 1D max-pooling layer with pool length 64 and stride 32.The max-pooling output is a vector with length 832, while filters learn different aspects of the underlying data.
2.5 Protein-protein interaction (PPI) network features
The method incorporates cross-species PPI networks from STRING, connects them through eggNOG orthology relations, and represents network information as knowledge graph embeddings combined with sequence features. Missing network representations are replaced with zero vectors.
- Network construction: The PPI network combines multiple species from STRING, confidence-filtered at 300, with symmetric ortholog-of and orthology-group relations from eggNOG.The resulting network contains 8,478,935 proteins.
- Feature representation: Network nodes are mapped to UniProt identifiers, covering 6,960,395 UniProt proteins and producing knowledge graph embeddings.Proteins without network representations receive zero vectors.
- Feature representation: Knowledge graph embeddings are combined with the 832-length max-pooling output into a combined feature vector.This integrates network-derived information with sequence-based features.
2.6 Hierarchical classification layout
DeepGO uses a hierarchical neural network with a separate two-layer classifier for each selected Gene Ontology class, connected through subclass relations. The model combines sequence-derived and protein-interaction representations, then aggregates leaf and internal-node outputs.
- Hierarchical classification layout: The model uses subclass relations to create a small neural network for each selected Gene Ontology class.Each class-specific network is fully connected and contains two layers.
- Hierarchical classification layout: Each class-specific network begins with a 256-neuron fully connected layer using a Rectified linear unit activation.Its input is the protein representation concatenated with a preceding representation.
- Hierarchical classification layout: The model output concatenates classification layers from leaf nodes with maximum layers from internal nodes.This aggregation follows the hierarchy shown in the neural network architecture.
- Hierarchical classification layout: The architecture receives protein-sequence trigram indexes and a 256-dimensional protein interaction-network representation.Trigrams are embedded as 128-dimensional vectors, producing a 1000x128 matrix for convolution and max-pooling.
2.7 Model implementation and optimization
The model is trained with multi-output binary cross entropy optimized by RMSprop, using mini-batches of 128 and a learning rate of 0.01. Training uses validation-based model selection, dropout regularization, and a Keras implementation with TensorFlow.
- The model minimizes multi-output binary cross entropy with RMSprop, using mini-batches of 128 and a learning rate of 0.01.
- Weights are uniformly initialized, and 80% of the training set is used for fitting while 20% is reserved for validation.
- The best-performing epoch’s weights are retained based on validation convergence, while dropout layers regularize the model against overfitting.
- The model is implemented using Keras with TensorFlow.
2.8 BLAST baseline
BLAST serves as the sequence-alignment baseline for performance comparison. For each test protein, BLASTP identifies the highest-scoring training-set match and transfers its functional annotations.
- BLAST is used as a sequence-alignment baseline to compare model performance.
- BLAST searches for the most similar experimentally annotated protein and assigns its annotations to the query sequence.
- For each ontology, the database contains training proteins annotated with at least one term from that ontology.
- For each test protein, BLASTP selects the training protein with the highest alignment score and transfers all its functional terms.
2.9 Evaluation
The model is evaluated using CAFA measures that assess protein-centric and term-centric function-prediction performance. These include maximum F-measure calculations and term-centric ROC-AUC and Fmax measures across prediction thresholds.
- Evaluation uses two measures established for the CAFA challenge.
- The protein-centric measure computes F-measure across thresholds and selects the maximum value, Fmax.Precision is averaged over proteins with at least one predicted term, while recall is averaged over all proteins.
- The term-centric measure computes ROC-curve AUC for each GO term using sensitivity across false-positive rates.Sensitivity corresponds to recall, and false-positive rate equals 1 − specificity.
- A term-centric Fmax is also reported by calculating each term’s F-measure across thresholds and selecting the maximum.This calculation uses all proteins in the test set.
3 Results
DeepGO was evaluated against BLAST and across organisms and protein-function groupings using CAFA metrics. Sequence-only DeepGOSeq improved cellular-location prediction, while performance varied by ontology, organism, and function specificity.
- Overall evaluation: DeepGOSeq, which uses only protein sequences, outperformed BLAST for cellular-location prediction but not for molecular function or biological process.This pattern held when evaluating either all GO functions or the subset of terms the model could predict.
- Organism-level evaluation: DeepGO performance varied significantly across organisms, with high performance reported for well-characterized model organisms.The evaluation also distinguished single-cell from multicellular organisms and eukaryotic from prokaryotic organisms.
- Term- and protein-level evaluation: Term-centric evaluation found that broadly annotated, high-level GO functions generally performed better than more specific functions.Because GO is hierarchical, more general functions always have more available training instances.
- Term- and protein-level evaluation: The authors propose weighting GO classes by information content during training to improve prediction of specific functions.They identify this as a future improvement motivated by the greater information carried by specific classes.
4 Discussion
DeepGO combines end-to-end feature learning with hierarchical classification that models dependencies across the GO hierarchy. Its benefits are balanced by substantial data and computational requirements, while future work will add richer GO relations and feature sources.
- Contributions: Deep learning extends function prediction beyond approaches that rely on hand-crafted features, building on feature-learning successes in related genomic prediction tasks.The passage identifies transcription-factor binding-site and genomic-variant impact prediction as examples of prior improvements from feature learning.
- Contributions: DeepGO’s hierarchical classifier optimizes whole hierarchies, accounts for class dependencies, learns hierarchically, and jointly trains with feature learning end-to-end.The model is also intended for other applications with structured output spaces.
- Advantages: End-to-end learning and global optimization allow the model to predict any class given sufficient training data, unlike approaches relying on hand-crafted feature vectors.The passage specifically contrasts this benefit with structured support vector machines.
- Limitations: The model requires large amounts of training data for each class, which are available for manual GO annotations but may be scarce for phenotype or variant-effect prediction.This data requirement limits straightforward application beyond settings with extensive annotations.
- Limitations: The model is complex and computationally demanding to train, so it may not be applicable in all settings.The limitation concerns the computational resources required for training.
- Future work: Future extensions will incorporate GO parthood and regulatory relations, additional interaction types, co-expression networks, and information extracted from text.The planned interaction sources include genetic interactions and co-expression networks.
Supplementary materials
The supplementary materials specify an InterPro-domain evaluation restricted to domains represented by at least 50 proteins and report cellular-component results for DeepGO and DeepGOSeq.
- Table 1 evaluates DeepGO performance by InterPro domain, including only domains represented by at least 50 proteins in the evaluation dataset.