MOTIVATION: Amyloid signaling sequences adopt the cross-β fold that is capable of self-replication in the templating process. Propagation of the amyloid fold from the receptor to the effector protein is used for signal transduction in the immune response pathways in animals, fungi, and bacteria. So far, a dozen of families of amyloid signaling motifs (ASMs) have been classified. Unfortunately, due to the wide variety of ASMs it is difficult to identify them in large protein databases available, which limits the possibility of conducting experimental studies. To date, various deep learning (DL) models have been applied across a range of protein-related tasks, including domain family classification and the prediction of protein structure and protein-protein interactions. RESULTS: In this study, we develop tailor-made bidirectional LSTM and BERT-based architectures to model ASM, and compare their performance against a state-of-the-art machine learning grammatical model. Our research is focused on developing a discriminative model of generalized ASMs, capable of detecting ASMs in large datasets. The DL-based models are trained on a diverse set of motif families and a global negative set, and used to identify ASMs from remotely related families. We analyze how both models represent the data and demonstrate that the DL-based approaches effectively detect ASMs, including novel motifs, even at the genome scale. AVAILABILITY AND IMPLEMENTATION: The models are provided as a Python package, asmscan-bilstm, and a Docker image at https://github.com/chrispysz/asmscan-proteinbert-run. The source code can be accessed at https://github.com/jakub-galazka/asmscan-bilstm and https://github.com/chrispysz/asmscan-proteinbert. Data and results are at https://github.com/wdyrka-pwr/ASMscan.

NLR proteins are intracellular receptors constituting a conserved component of the innate immune system of cellular organisms. In fungi, NLRs are characterized by high diversity of architectures and presence of amyloid signaling. Here, we explore the diverse world of effector and signaling domains of fungal NLRs using state-of-the-art bioinformatic methods including MMseqs2 for fast clustering, probabilistic context-free grammars for sequence analysis, and AlphaFold2 deep neural networks for structure prediction. In addition to substantially improving the overall annotation, especially in basidiomycetes, the study identifies novel domains and reveals the structural similarity of MLKL-related HeLo- and Goodbye-like domains forming the most abundant superfamily of fungal NLR effectors. Moreover, compared to previous studies, we found several times more amyloid motif instances, including novel families, and validated aggregating and prion-forming properties of the most abundant of them in vitro and in vivo. Also, through an extensive in silico search, the NLR-associated amyloid signaling was identified in basidiomycetes. The emerging picture highlights similarities and differences in the NLR architectures and amyloid signaling in ascomycetes, basidiomycetes and other branches of life.

BACKGROUND: Amyloid signaling motifs are a class of protein motifs which share basic structural and functional features despite the lack of clear sequence homology. They are hard to detect in large sequence databases either with the alignment-based profile methods (due to short length and diversity) or with generic amyloid- and prion-finding tools (due to insufficient discriminative power). We propose to address the challenge with a machine learning grammatical model capable of generalizing over diverse collections of unaligned yet related motifs. RESULTS: First, we introduce and test improvements to our probabilistic context-free grammar framework for protein sequences that allow for inferring more sophisticated models achieving high sensitivity at low false positive rates. Then, we infer universal grammars for a collection of recently identified bacterial amyloid signaling motifs and demonstrate that the method is capable of generalizing by successfully searching for related motifs in fungi. The results are compared to available alternative methods. Finally, we conduct spectroscopy and staining analyses of selected peptides to verify their structural and functional relationship. CONCLUSIONS: While the profile HMMs remain the method of choice for modeling homologous sets of sequences, PCFGs seem more suitable for building meta-family descriptors and extrapolating beyond the seed sample.

BACKGROUND: The field of protein sequence analysis is dominated by tools rooted in substitution matrices and alignments. A complementary approach is provided by methods of quantitative characterization. A major advantage of the approach is that quantitative properties defines a multidimensional solution space, where sequences can be related to each other and differences can be meaningfully interpreted. RESULTS: Quantiprot is a software package in Python, which provides a simple and consistent interface to multiple methods for quantitative characterization of protein sequences. The package can be used to calculate dozens of characteristics directly from sequences or using physico-chemical properties of amino acids. Besides basic measures, Quantiprot performs quantitative analysis of recurrence and determinism in the sequence, calculates distribution of n-grams and computes the Zipf's law coefficient. CONCLUSIONS: We propose three main fields of application of the Quantiprot package. First, quantitative characteristics can be used in alignment-free similarity searches, and in clustering of large and/or divergent sequence sets. Second, a feature space defined by quantitative properties can be used in comparative studies of protein families and organisms. Third, the feature space can be used for evaluating generative models, where large number of sequences generated by the model can be compared to actually observed sequences.

BACKGROUND: In the last decade, there have been many applications of formal language theory in bioinformatics such as RNA structure prediction and detection of patterns in DNA. However, in the field of proteomics, the size of the protein alphabet and the complexity of relationship between amino acids have mainly limited the application of formal language theory to the production of grammars whose expressive power is not higher than stochastic regular grammars. However, these grammars, like other state of the art methods, cannot cover any higher-order dependencies such as nested and crossing relationships that are common in proteins. In order to overcome some of these limitations, we propose a Stochastic Context Free Grammar based framework for the analysis of protein sequences where grammars are induced using a genetic algorithm. RESULTS: This framework was implemented in a system aiming at the production of binding site descriptors. These descriptors not only allow detection of protein regions that are involved in these sites, but also provide insight in their structure. Grammars were induced using quantitative properties of amino acids to deal with the size of the protein alphabet. Moreover, we imposed some structural constraints on grammars to reduce the extent of the rule search space. Finally, grammars based on different properties were combined to convey as much information as possible. Evaluation was performed on sites of various sizes and complexity described either by PROSITE patterns, domain profiles or a set of patterns. Results show the produced binding site descriptors are human-readable and, hence, highlight biologically meaningful features. Moreover, they achieve good accuracy in both annotation and detection. In addition, findings suggest that, unlike current state-of-the-art methods, our system may be particularly suited to deal with patterns shared by non-homologous proteins. CONCLUSION: A new Stochastic Context Free Grammar based framework has been introduced allowing the production of binding site descriptors for analysis of protein sequences. Experiments have shown that not only is this new approach valid, but produces human-readable descriptors for binding sites which have been beyond the capability of current machine learning techniques.