An official website of the United States government
Abstract BackgroundThe number of applications of deep learning algorithms in bioinformatics is increasing as they usually achieve superior performance over classical approaches, especially, when bigger training datasets are available. In deep learning applications, discrete data, e.g. words or n-grams in language, or amino acids or nucleotides in bioinformatics, are generally represented as a continuous vector through an embedding matrix. Recently, learning this embedding matrix directly from the data as part of the continuous iteration of the model to optimize the target prediction – a process called ‘end-to-end learning’ – has led to state-of-the-art results in many fields. Although usage of embeddings is well described in the bioinformatics literature, the potential of end-to-end learning for single amino acids, as compared to more classical manually-curated encoding strategies, has not been systematically addressed. To this end, we compared classical encoding matrices, namely one-hot, VHSE8 and BLOSUM62, to end-to-end learning of amino acid embeddings for two different prediction tasks using three widely used architectures, namely recurrent neural networks (RNN), convolutional neural networks (CNN), and the hybrid CNN-RNN. ResultsBy using different deep learning architectures, we show that end-to-end learning is on par with classical encodings for embeddings of the same dimension even when limited training data is available, and might allow for a reduction in the embedding dimension without performance loss, which is critical when deploying the models to devices with limited computational capacities. We found that the embedding dimension is a major factor in controlling the model performance. Surprisingly, we observed that deep learning models are capable of learning from random vectors of appropriate dimension. ConclusionOur study shows that end-to-end learning is a flexible and powerful method for amino acid encoding. Further, due to the flexibility of deep learning systems, amino acid encoding schemes should be benchmarked against random vectors of the same dimension to disentangle the information content provided by the encoding scheme from the distinguishability effect provided by the scheme.
more »
« less
Numerical Encodings of Amino Acids in Multivariate Gaussian Modeling of Protein Multiple Sequence Alignments
Residues in proteins that are in close spatial proximity are more prone to covariate as their interactions are likely to be preserved due to structural and evolutionary constraints. If we can detect and quantify such covariation, physical contacts may then be predicted in the structure of a protein solely from the sequences that decorate it. To carry out such predictions, and following the work of others, we have implemented a multivariate Gaussian model to analyze correlation in multiple sequence alignments. We have explored and tested several numerical encodings of amino acids within this model. We have shown that 1D encodings based on amino acid biochemical and biophysical properties, as well as higher dimensional encodings computed from the principal components of experimentally derived mutation/substitution matrices, do not perform as well as a simple twenty dimensional encoding with each amino acid represented with a vector of one along its own dimension and zero elsewhere. The optimum obtained from representations based on substitution matrices is reached by using 10 to 12 principal components; the corresponding performance is less than the performance obtained with the 20-dimensional binary encoding. We highlight also the importance of the prior when constructing the multivariate Gaussian model of a multiple sequence alignment.
more »
« less
- Award ID(s):
- 1760485
- PAR ID:
- 10096308
- Date Published:
- Journal Name:
- Molecules
- Volume:
- 24
- Issue:
- 1
- ISSN:
- 1420-3049
- Page Range / eLocation ID:
- 104
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
Abstract In this work, we expand on a dataset recently introduced for protein interface prediction (PIP), the Database of Interacting Protein Structures (DIPS), to present DIPS-Plus, an enhanced, feature-rich dataset of 42,112 complexes for machine learning of protein interfaces. While the original DIPS dataset contains only the Cartesian coordinates for atoms contained in the protein complex along with their types, DIPS-Plus contains multiple residue-level features including surface proximities, half-sphere amino acid compositions, and new profile hidden Markov model (HMM)-based sequence features for each amino acid, providing researchers a curated feature bank for training protein interface prediction methods. We demonstrate through rigorous benchmarks that training an existing state-of-the-art (SOTA) model for PIP on DIPS-Plus yields new SOTA results, surpassing the performance of some of the latest models trained on residue-level and atom-level encodings of protein complexes to date.more » « less
-
Abstract Profile mixture models capture distinct biochemical constraints on the amino acid substitution process at different sites in proteins. These models feature a mixture of time-reversible models with a common matrix of exchangeabilities and distinct sets of equilibrium amino acid frequencies known as profiles. Combining the exchangeability matrix with each profile generates the matrix of instantaneous rates of amino acid exchange for that profile. Currently, empirically estimated exchangeability matrices (e.g. the LG matrix) are widely used for phylogenetic inference under profile mixture models. However, these were estimated using a single profile and are unlikely optimal for profile mixture models. Here, we describe the GTRpmix model that allows maximum likelihood estimation of a common exchangeability matrix under any profile mixture model. We show that exchangeability matrices estimated under profile mixture models differ from the LG matrix, dramatically improving model fit and topological estimation accuracy for empirical test cases. Because the GTRpmix model is computationally expensive, we provide two exchangeability matrices estimated from large concatenated phylogenomic-supermatrices to be used for phylogenetic analyses. One, called Eukaryotic Linked Mixture (ELM), is designed for phylogenetic analysis of proteins encoded by nuclear genomes of eukaryotes, and the other, Eukaryotic and Archaeal Linked mixture (EAL), for reconstructing relationships between eukaryotes and Archaea. These matrices, combined with profile mixture models, fit data better and have improved topology estimation relative to the LG matrix combined with the same mixture models. Starting with version 2.3.1, IQ-TREE2 allows users to estimate linked exchangeabilities (i.e. amino acid exchange rates) under profile mixture models.more » « less
-
Summary The covariance structure of multivariate functional data can be highly complex, especially if the multivariate dimension is large, making extensions of statistical methods for standard multivariate data to the functional data setting challenging. For example, Gaussian graphical models have recently been extended to the setting of multivariate functional data by applying multivariate methods to the coefficients of truncated basis expansions. However, compared with multivariate data, a key difficulty is that the covariance operator is compact and thus not invertible. This paper addresses the general problem of covariance modelling for multivariate functional data, and functional Gaussian graphical models in particular. As a first step, a new notion of separability for the covariance operator of multivariate functional data is proposed, termed partial separability, leading to a novel Karhunen–Loève-type expansion for such data. Next, the partial separability structure is shown to be particularly useful in providing a well-defined functional Gaussian graphical model that can be identified with a sequence of finite-dimensional graphical models, each of identical fixed dimension. This motivates a simple and efficient estimation procedure through application of the joint graphical lasso. Empirical performance of the proposed method for graphical model estimation is assessed through simulation and analysis of functional brain connectivity during a motor task.more » « less
-
Abstract ObjectiveThe COVID-19 pandemic emphasized the value of geospatial visual analytics for both epidemiologists and the general public. However, systems struggled to encode temporal and geospatial trends of multiple, potentially interacting variables, such as active cases, deaths, and vaccinations. We sought to ask (1) how epidemiologists interact with visual analytics tools, (2) how multiple, time-varying, geospatial variables can be conveyed in a unified view, and (3) how complex spatiotemporal encodings affect utility for both experts and non-experts. Materials and MethodsWe propose encoding variables with animated, concentric, hollow circles, allowing multiple variables via color encoding and avoiding occlusion problems, and we implement this method in a browser-based tool called CoronaViz. We conduct task-based evaluations with non-experts, as well as in-depth interviews and observational sessions with epidemiologists, covering a range of tools and encodings. ResultsSessions with epidemiologists confirmed the importance of multivariate, spatiotemporal queries and the utility of CoronaViz for answering them, while providing direction for future development. Non-experts tasked with performing spatiotemporal queries unanimously preferred animation to multi-view dashboards. DiscussionWe find that conveying complex, multivariate data necessarily involves trade-offs. Yet, our studies suggest the importance of complementary visualization strategies, with our animated multivariate spatiotemporal encoding filling important needs for exploration and presentation. ConclusionCoronaViz’s unique ability to convey multiple, time-varying, geospatial variables makes it both a valuable addition to interactive COVID-19 dashboards and a platform for empowering experts and the public during future disease outbreaks. CoronaViz is open-source and a live instance is freely hosted at http://coronaviz.umiacs.io.more » « less
- Free Publicly Accessible Full Text
-
Accepted Manuscript1.0
- Journal Article:
- https://doi.org/10.3390/molecules24010104
