High precision prediction of functional sites in protein structures

Explainable protein function annotation using local structure embeddings

The rapid expansion of protein sequence and structure databases has resulted in a significant number of proteins with ambiguous or unknown function. While advances in machine learning techniques hold great potential to fill this annotation gap, current methods for function prediction are unable to associate global function reliably to the specific residues responsible for that function. We address this issue by introducing PARSE (Protein Annotation by Residue-Specific Enrichment), a knowledge-based method which combines pre-trained embeddings of local structural environments with traditional statistical techniques to identify enriched functions with residue-level explainability. For the task of predicting the catalytic function of enzymes, PARSE achieves comparable or superior global performance to state-of-the-art machine learning methods (F1 score > 85%) while simultaneously annotating the specific residues involved in each function with much greater precision. Since it does not require supervised training, our method can make one-shot predictions for very rare functions and is not limited to a particular type of functional label (e.g. Enzyme Commission numbers or Gene Ontology codes). Finally, we leverage the AlphaFold Structure Database to perform functional annotation at a proteome scale. By applying PARSE to the dark proteome-predicted structures which cannot be classified into known structural families-we predict several novel bacterial metalloproteases. Each of these proteins shares a strongly conserved catalytic site despite highly divergent sequences and global folds, illustrating the value of local structure representations for new function discovery.

COLLAPSE: A representation learning framework for identification and characterization of protein structural sites

The identification and characterization of the structural sites which contribute to protein function are crucial for understanding biological mechanisms, evaluating disease risk, and developing targeted therapies. However, the quantity of known protein structures is rapidly outpacing our ability to functionally annotate them. Existing methods for function prediction either do not operate on local sites, suffer from high false positive or false negative rates, or require large site-specific training datasets, necessitating the development of new computational methods for annotating functional sites at scale. We present COLLAPSE (Compressed Latents Learned from Aligned Protein Structural Environments), a framework for learning deep representations of protein sites. COLLAPSE operates directly on the 3D positions of atoms surrounding a site and uses evolutionary relationships between homologous proteins as a self-supervision signal, enabling learned embeddings to implicitly capture structure-function relationships within each site. Our representations generalize across disparate tasks in a transfer learning context, achieving state-of-the-art performance on standardized benchmarks (protein-protein interactions and mutation stability) and on the prediction of functional sites from the Prosite database. We use COLLAPSE to search for similar sites across large protein datasets and to annotate proteins based on a database of known functional sites. These methods demonstrate that COLLAPSE is computationally efficient, tunable, and interpretable, providing a general-purpose platform for computational protein analysis.

Machine learning differentiates enzymatic and non-enzymatic metals in proteins

Metalloenzymes are 40% of all enzymes and can perform all seven classes of enzyme reactions. Because of the physicochemical similarities between the active sites of metalloenzymes and inactive metal binding sites, it is challenging to differentiate between them. Yet distinguishing these two classes is critical for the identification of both native and designed enzymes. Because of similarities between catalytic and non-catalytic  metal binding sites, finding physicochemical features that distinguish these two types of metal sites can indicate aspects that are critical to enzyme function. In this work, we develop the largest structural dataset of enzymatic and non-enzymatic metalloprotein sites to date. We then use a decision-tree ensemble machine learning model to classify metals bound to proteins as enzymatic or non-enzymatic with 92.2% precision and 90.1% recall. Our model scores electrostatic and pocket lining features as more important than pocket volume, despite the fact that volume is the most quantitatively different feature between enzyme and non-enzymatic sites. Finally, we find our model has overall better performance in a side-to-side comparison against other methods that differentiate enzymatic from non-enzymatic sequences. We anticipate that our model's ability to correctly identify which metal sites are responsible for enzymatic activity could enable identification of new enzymatic mechanisms and de novo enzyme design.

ResiRole: residue-level functional site predictions to gauge the accuracies of protein structure prediction techniques

Motivation

Methods to assess the quality of protein structure models are needed for user applications. To aid with the selection of structure models and further inform the development of structure prediction techniques, we describe the ResiRole method for the assessment of the quality of structure models.

Results

Structure prediction techniques are ranked according to the results of round-robin, head-to-head comparisons using difference scores. Each difference score was defined as the absolute value of the cumulative probability for a functional site prediction made with the FEATURE program for the reference structure minus that for the structure model. Overall, the difference scores correlate well with other model quality metrics; and based on benchmarking studies with NaïveBLAST, they are found to detect additional local structural similarities between the structure models and reference structures.

Availabilityand implementation

Automated analyses of models addressed in CAMEO are available via the ResiRole server, URL http://protein.som.geisinger.edu/ResiRole/. Interactive analyses with user-provided models and reference structures are also enabled. Code is available at github.com/wamclaughlin/ResiRole.

Supplementary information

Supplementary data are available at Bioinformatics online.

High precision protein functional site detection using 3D convolutional neural networks

Motivation

Accurate annotation of protein functions is fundamental for understanding molecular and cellular physiology. Data-driven methods hold promise for systematically deriving rules underlying the relationship between protein structure and function. However, the choice of protein structural representation is critical. Pre-defined biochemical features emphasize certain aspects of protein properties while ignoring others, and therefore may fail to capture critical information in complex protein sites.

Results

In this paper, we present a general framework that applies 3D convolutional neural networks (3DCNNs) to structure-based protein functional site detection. The framework can extract task-dependent features automatically from the raw atom distributions. We benchmarked our method against other methods and demonstrate better or comparable performance for site detection. Our deep 3DCNNs achieved an average recall of 0.955 at a precision threshold of 0.99 on PROSITE families, detected 98.89 and 92.88% of nitric oxide synthase and TRYPSIN-like enzyme sites in Catalytic Site Atlas, and showed good performance on challenging cases where sequence motifs are absent but a function is known to exist. Finally, we inspected the individual contributions of each atom to the classification decisions and show that our models successfully recapitulate known 3D features within protein functional sites.

Availability and implementation

The 3DCNN models described in this paper are available at https://simtk.org/projects/fscnn.

Supplementary information

Supplementary data are available at Bioinformatics online.

Improving the explainability of Random Forest classifier - user centered approach

Machine Learning (ML) methods are now influencing major decisions about patient care, new medical methods, drug development and their use and importance are rapidly increasing in all areas. However, these ML methods are inherently complex and often difficult to understand and explain resulting in barriers to their adoption and validation. Our work (RFEX) focuses on enhancing Random Forest (RF) classifier explainability by developing easy to interpret explainability summary reports from trained RF classifiers as a way to improve the explainability for (often non-expert) users. RFEX is implemented and extensively tested on Stanford FEATURE data where RF is tasked with predicting functional sites in 3D molecules based on their electrochemical signatures (features). In developing RFEX method we apply user-centered approach driven by explainability questions and requirements collected by discussions with interested practitioners. We performed formal usability testing with 13 expert and non-expert users to verify RFEX usefulness. Analysis of RFEX explainability report and user feedback indicates its usefulness in significantly increasing explainability and user confidence in RF classification on FEATURE data. Notably, RFEX summary reports easily reveal that one needs very few (from 2-6 depending on a model) top ranked features to achieve 90% or better of the accuracy when all 480 features are used.

3D deep convolutional neural networks for amino acid environment similarity analysis

Background

Central to protein biology is the understanding of how structural elements give rise to observed function. The surfeit of protein structural data enables development of computational methods to systematically derive rules governing structural-functional relationships. However, performance of these methods depends critically on the choice of protein structural representation. Most current methods rely on features that are manually selected based on knowledge about protein structures. These are often general-purpose but not optimized for the specific application of interest.

In this paper, we present a general framework that applies 3D convolutional neural network (3DCNN) technology to structure-based protein analysis. The framework automatically extracts task-specific features from the raw atom distribution, driven by supervised labels. As a pilot study, we use our network to analyze local protein microenvironments surrounding the 20 amino acids, and predict the amino acids most compatible with environments within a protein structure. To further validate the power of our method, we construct two amino acid substitution matrices from the prediction statistics and use them to predict effects of mutations in T4 lysozyme structures.

Results

Our deep 3DCNN achieves a two-fold increase in prediction accuracy compared to models that employ conventional hand-engineered features and successfully recapitulates known information about similar and different microenvironments. Models built from our predictions and substitution matrices achieve an 85% accuracy predicting outcomes of the T4 lysozyme mutation variants. Our substitution matrices contain rich information relevant to mutation analysis compared to well-established substitution matrices. Finally, we present a visualization method to inspect the individual contributions of each atom to the classification decisions.

Conclusions

End-to-end trained deep learning networks consistently outperform methods using hand-engineered features, suggesting that the 3DCNN framework is well suited for analysis of protein microenvironments and may be useful for other protein structural analyses.

Electronic supplementary material

The online version of this article (doi:10.1186/s12859-017-1702-0) contains supplementary material, which is available to authorized users.

Using support vector machines to improve elemental ion identification in macromolecular crystal structures

In the process of macromolecular model building, crystallographers must examine electron density for isolated atoms and differentiate sites containing structured solvent molecules from those containing elemental ions. This task requires specific knowledge of metal-binding chemistry and scattering properties and is prone to error. A method has previously been described to identify ions based on manually chosen criteria for a number of elements. Here, the use of support vector machines (SVMs) to automatically classify isolated atoms as either solvent or one of various ions is described. Two data sets of protein crystal structures, one containing manually curated structures deposited with anomalous diffraction data and another with automatically filtered, high-resolution structures, were constructed. On the manually curated data set, an SVM classifier was able to distinguish calcium from manganese, zinc, iron and nickel, as well as all five of these ions from water molecules, with a high degree of accuracy. Additionally, SVMs trained on the automatically curated set of high-resolution structures were able to successfully classify most common elemental ions in an independent validation test set. This method is readily extensible to other elemental ions and can also be used in conjunction with previous methods based on a priori expectations of the chemical environment and X-ray scattering.