Exploring the limitations of biophysical propensity scales coupled with machine learning for protein sequence analysis

AAontology: An Ontology of Amino Acid Scales for Interpretable Machine Learning

Amino acid scales are crucial for protein prediction tasks, many of them being curated in the AAindex database. Despite various clustering attempts to organize them and to better understand their relationships, these approaches lack the fine-grained classification necessary for satisfactory interpretability in many protein prediction problems. To address this issue, we developed AAontology-a two-level classification for 586 amino acid scales (mainly from AAindex) together with an in-depth analysis of their relations-using bag-of-word-based classification, clustering, and manual refinement over multiple iterations. AAontology organizes physicochemical scales into 8 categories and 67 subcategories, enhancing the interpretability of scale-based machine learning methods in protein bioinformatics. Thereby it enables researchers to gain a deeper biological insight. We anticipate that AAontology will be a building block to link amino acid properties with protein function and dysfunctions as well as aid informed decision-making in mutation analysis or protein drug design.

Genotype sampling for deep-learning assisted experimental mapping of a combinatorially complete fitness landscape

MOTIVATION

Experimental characterization of fitness landscapes, which map genotypes onto fitness, is important for both evolutionary biology and protein engineering. It faces a fundamental obstacle in the astronomical number of genotypes whose fitness needs to be measured for any one protein. Deep learning may help to predict the fitness of many genotypes from a smaller neural network training sample of genotypes with experimentally measured fitness. Here I use a recently published experimentally mapped fitness landscape of more than 260 000 protein genotypes to ask how such sampling is best performed.

RESULTS

I show that multilayer perceptrons, recurrent neural networks, convolutional networks, and transformers, can explain more than 90% of fitness variance in the data. In addition, 90% of this performance is reached with a training sample comprising merely ≈103 sequences. Generalization to unseen test data is best when training data is sampled randomly and uniformly, or sampled to minimize the number of synonymous sequences. In contrast, sampling to maximize sequence diversity or codon usage bias reduces performance substantially. These observations hold for more than one network architecture. Simple sampling strategies may perform best when training deep learning neural networks to map fitness landscapes from experimental data.

AVAILABILITY AND IMPLEMENTATION

The fitness landscape data analyzed here is publicly available as described previously (Papkou et al. 2023). All code used to analyze this landscape is publicly available at https://github.com/andreas-wagner-uzh/fitness_landscape_sampling.

Machine Learning-Guided Protein Engineering

Recent progress in engineering highly promising biocatalysts has increasingly involved machine learning methods. These methods leverage existing experimental and simulation data to aid in the discovery and annotation of promising enzymes, as well as in suggesting beneficial mutations for improving known targets. The field of machine learning for protein engineering is gathering steam, driven by recent success stories and notable progress in other areas. It already encompasses ambitious tasks such as understanding and predicting protein structure and function, catalytic efficiency, enantioselectivity, protein dynamics, stability, solubility, aggregation, and more. Nonetheless, the field is still evolving, with many challenges to overcome and questions to address. In this Perspective, we provide an overview of ongoing trends in this domain, highlight recent case studies, and examine the current limitations of machine learning-based methods. We emphasize the crucial importance of thorough experimental validation of emerging models before their use for rational protein design. We present our opinions on the fundamental problems and outline the potential directions for future research.

Prediction of B cell epitopes in proteins using a novel sequence similarity-based method

Prediction of B cell epitopes that can replace the antigen for antibody production and detection is of great interest for research and the biotech industry. Here, we developed a novel BLAST-based method to predict linear B cell epitopes. To that end, we generated a BLAST-formatted database upon a dataset of 62,730 known linear B cell epitope sequences and considered as a B cell epitope any peptide sequence producing ungapped BLAST hits to this database with identity ≥ 80% and length ≥ 8. We examined B cell epitope predictions by this method in tenfold cross-validations in which we considered various types of non-B cell epitopes, including 62,730 peptide sequences with verified negative B cell assays. As a result, we obtained values of accuracy, specificity and sensitivity of 72.54 ± 0.27%, 81.59 ± 0.37% and 63.49 ± 0.43%, respectively. In an independent dataset incorporating 503 B cell epitopes, this method reached accuracy, specificity and sensitivity of 74.85%, 99.20% and 50.50%, respectively, outperforming state-of-the-art methods to predict linear B cell epitopes. We implemented this BLAST-based approach to predict B cell epitopes at http://imath.med.ucm.es/bepiblast.

Machine Learning Based Predictive Model for the Analysis of Sequence Activity Relationships Using Protein Spectra and Protein Descriptors

Accurately establishing the connection between a protein sequence and its function remains a focal point within the field of protein engineering, especially in the context of predicting the effects of mutations. From this, there has been a continued drive to build accurate and reliable predictive models via machine learning that allow for the virtual screening of many protein mutant sequences, measuring the relationship between sequence and 'fitness' or 'activity', commonly known as a Sequence-Activity-Relationship (SAR). An important preliminary stage in the building of these predictive models is the encoding of the chosen sequences. Evaluated in this work is a plethora of encoding strategies using the Amino Acid Index database, where the indices are transformed into their spectral form via Digital Signal Processing (DSP) techniques, as well as numerous protein structural and physiochemical descriptors. The encoding strategies are explored on a dataset curated to measure the thermostability of various mutants from a recombination library, designed from parental cytochrome P450s. In this work it was concluded that the implementation of protein spectra in concatenation with protein descriptors, together with the Partial Least Squares Regression (PLS) algorithm, gave the most noteworthy increase in the quality of the predictive models (as described in Encoding Strategy C), highlighting their utility in identifying an SAR. The accompanying software produced for this paper is termed pySAR (Python Sequence-Activity-Relationship), which allows for a user to find the optimal arrangement of structural and or physiochemical properties to encode their specific mutant library dataset; the source code is available at: https://github.com/amckenna41/pySAR.

PARROT is a flexible recurrent neural network framework for analysis of large protein datasets

The rise of high-throughput experiments has transformed how scientists approach biological questions. The ubiquity of large-scale assays that can test thousands of samples in a day has necessitated the development of new computational approaches to interpret this data. Among these tools, machine learning approaches are increasingly being utilized due to their ability to infer complex nonlinear patterns from high-dimensional data. Despite their effectiveness, machine learning (and in particular deep learning) approaches are not always accessible or easy to implement for those with limited computational expertise. Here we present PARROT, a general framework for training and applying deep learning-based predictors on large protein datasets. Using an internal recurrent neural network architecture, PARROT is capable of tackling both classification and regression tasks while only requiring raw protein sequences as input. We showcase the potential uses of PARROT on three diverse machine learning tasks: predicting phosphorylation sites, predicting transcriptional activation function of peptides generated by high-throughput reporter assays, and predicting the fibrillization propensity of amyloid beta with data generated by deep mutational scanning. Through these examples, we demonstrate that PARROT is easy to use, performs comparably to state-of-the-art computational tools, and is applicable for a wide array of biological problems.

iAtbP-Hyb-EnC: Prediction of antitubercular peptides via heterogeneous feature representation and genetic algorithm based ensemble learning model

Tuberculosis (TB) is a worldwide illness caused by the bacteria Mycobacterium tuberculosis. Owing to the high prevalence of multidrug-resistant tuberculosis, numerous traditional strategies for developing novel alternative therapies have been presented. The effectiveness and dependability of these procedures are not always consistent. Peptide-based therapy has recently been regarded as a preferable alternative due to its excellent selectivity in targeting specific cells without affecting the normal cells. However, due to the rapid growth of the peptide samples, predicting TB accurately has become a challenging task. To effectively identify antitubercular peptides, an intelligent and reliable prediction model is indispensable. An ensemble learning approach was used in this study to improve expected results by compensating for the shortcomings of individual classification algorithms. Initially, three distinct representation approaches were used to formulate the training samples: k-space amino acid composition, composite physiochemical properties, and one-hot encoding. The feature vectors of the applied feature extraction methods are then combined to generate a heterogeneous vector. Finally, utilizing individual and heterogeneous vectors, five distinct nature classification models were used to evaluate prediction rates. In addition, a genetic algorithm-based ensemble model was used to improve the suggested model's prediction and training capabilities. Using Training and independent datasets, the proposed ensemble model achieved an accuracy of 94.47% and 92.68%, respectively. It was observed that our proposed "iAtbP-Hyb-EnC" model outperformed and reported ~10% highest training accuracy than existing predictors. The "iAtbP-Hyb-EnC" model is suggested to be a reliable tool for scientists and might play a valuable role in academic research and drug discovery. The source code and all datasets are publicly available at https://github.com/Farman335/iAtbP-Hyb-EnC.

In-silico prediction of in-vitro protein liquid-liquid phase separation experiments outcomes with multi-head neural attention

MOTIVATION

Proteins able to undergo Liquid-Liquid Phase Separation (LLPS) in-vivo and in-vitro are drawing a lot of interest, due to their functional relevance for cell life. Nevertheless, the proteome-scale experimental screening of these proteins seems unfeasible, because besides being expensive and time consuming, LLPS is heavily influenced by multiple environmental conditions such as concentration, pH and temperature, thus requiring a combinatorial number of experiments for each protein.

RESULTS

To overcome this problem, we propose an Neural Network model able to predict the LLPS behavior of proteins given specified experimental conditions, effectively predicting the outcome of in-vitro experiments. Our model can be used to rapidly screen proteins and experimental conditions searching for LLPS, thus reducing the search space that needs to be covered experimentally. We experimentally validate Droppler's prediction on the the TAR DNA-binding protein in different experimental conditions, showing the consistency of its predictions.

SUPPLEMENTARY INFORMATION

Supplementary data are available at Bioinformatics online.

Applications of Machine and Deep Learning in Adaptive Immunity

Adaptive immunity is mediated by lymphocyte B and T cells, which respectively express a vast and diverse repertoire of B cell and T cell receptors and, in conjunction with peptide antigen presentation through major histocompatibility complexes (MHCs), can recognize and respond to pathogens and diseased cells. In recent years, advances in deep sequencing have led to a massive increase in the amount of adaptive immune receptor repertoire data; additionally, proteomics techniques have led to a wealth of data on peptide-MHC presentation. These large-scale data sets are now making it possible to train machine and deep learning models, which can be used to identify complex and high-dimensional patterns in immune repertoires. This article introduces adaptive immune repertoires and machine and deep learning related to biological sequence data and then summarizes the many applications in this field, which span from predicting the immunological status of a host to the antigen specificity of individual receptors and the engineering of immunotherapeutics.

Incorporating Machine Learning into Established Bioinformatics Frameworks

The exponential growth of biomedical data in recent years has urged the application of numerous machine learning techniques to address emerging problems in biology and clinical research. By enabling the automatic feature extraction, selection, and generation of predictive models, these methods can be used to efficiently study complex biological systems. Machine learning techniques are frequently integrated with bioinformatic methods, as well as curated databases and biological networks, to enhance training and validation, identify the best interpretable features, and enable feature and model investigation. Here, we review recently developed methods that incorporate machine learning within the same framework with techniques from molecular evolution, protein structure analysis, systems biology, and disease genomics. We outline the challenges posed for machine learning, and, in particular, deep learning in biomedicine, and suggest unique opportunities for machine learning techniques integrated with established bioinformatics approaches to overcome some of these challenges.

Current cancer driver variant predictors learn to recognize driver genes instead of functional variants

BACKGROUND

Identifying variants that drive tumor progression (driver variants) and distinguishing these from variants that are a byproduct of the uncontrolled cell growth in cancer (passenger variants) is a crucial step for understanding tumorigenesis and precision oncology. Various bioinformatics methods have attempted to solve this complex task.

RESULTS

In this study, we investigate the assumptions on which these methods are based, showing that the different definitions of driver and passenger variants influence the difficulty of the prediction task. More importantly, we prove that the data sets have a construction bias which prevents the machine learning (ML) methods to actually learn variant-level functional effects, despite their excellent performance. This effect results from the fact that in these data sets, the driver variants map to a few driver genes, while the passenger variants spread across thousands of genes, and thus just learning to recognize driver genes provides almost perfect predictions.

CONCLUSIONS

To mitigate this issue, we propose a novel data set that minimizes this bias by ensuring that all genes covered by the data contain both driver and passenger variants. As a result, we show that the tested predictors experience a significant drop in performance, which should not be considered as poorer modeling, but rather as correcting unwarranted optimism. Finally, we propose a weighting procedure to completely eliminate the gene effects on such predictions, thus precisely evaluating the ability of predictors to model the functional effects of single variants, and we show that indeed this task is still open.

Amino acid encoding for deep learning applications

BACKGROUND

The 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.

RESULTS

By 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.

CONCLUSION

Our 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.

Insight into the protein solubility driving forces with neural attention

Protein solubility is a key aspect for many biotechnological, biomedical and industrial processes, such as the production of active proteins and antibodies. In addition, understanding the molecular determinants of the solubility of proteins may be crucial to shed light on the molecular mechanisms of diseases caused by aggregation processes such as amyloidosis. Here we present SKADE, a novel Neural Network protein solubility predictor and we show how it can provide novel insight into the protein solubility mechanisms, thanks to its neural attention architecture. First, we show that SKADE positively compares with state of the art tools while using just the protein sequence as input. Then, thanks to the neural attention mechanism, we use SKADE to investigate the patterns learned during training and we analyse its decision process. We use this peculiarity to show that, while the attention profiles do not correlate with obvious sequence aspects such as biophysical properties of the aminoacids, they suggest that N- and C-termini are the most relevant regions for solubility prediction and are predictive for complex emergent properties such as aggregation-prone regions involved in beta-amyloidosis and contact density. Moreover, SKADE is able to identify mutations that increase or decrease the overall solubility of the protein, allowing it to be used to perform large scale in-silico mutagenesis of proteins in order to maximize their solubility.

The solubility of proteins is a crucial biophysical aspect when it comes to understanding many human diseases and to improve the industrial processes for protein production. Due to its relevance, computational methods have been devised in order to study and possibly optimize the solubility of proteins. In this work we apply a deep-learning technique, called neural attention to predict protein solubility while “opening” the model itself to interpretability, even though Machine Learning models are usually considered black boxes. Thank to the attention mechanism, we show that i) our model implicitly learns complex patterns related to emergent, protein folding-related, aspects such as to recognize β-amyloidosis regions and that ii) the N-and C-termini are the regions with the highes signal fro solubility prediction. When it comes to enhancing the solubility of proteins, we, for the first time, propose to investigate the synergistic effects of tandem mutations instead of “single” mutations, suggesting that this could minimize the number of required proposed mutations.