MOTIVATION: For the alignment of large numbers of protein sequences, tools are predominant that decide to align two residues using only simple prior knowledge, e.g. amino acid substitution matrices, and using only part of the available data. The accuracy of state-of-the-art programs declines with decreasing sequence identity and when increasingly large numbers of sequences are aligned. Recently, transformer-based deep-learning models started to harness the vast amount of protein sequence data, resulting in powerful pretrained language models with the main purpose of generating high-dimensional numerical representations, embeddings, for individual sites that agglomerate evolutionary, structural, and biophysical information. RESULTS: We extend the traditional profile hidden Markov model so that it takes as inputs unaligned protein sequences and the corresponding embeddings. We fit the model with gradient descent using our existing differentiable hidden Markov layer. All sequences and their embeddings are jointly aligned to a model of the protein family. We report that our upgraded HMM-based aligner, learnMSA2, combined with the ProtT5-XL protein language model aligns on average almost 6% points more columns correctly than the best amino acid-based competitor and scales well with sequence number. The relative advantage of learnMSA2 over other programs tends to be greater when the sequence identity is lower and when the number of sequences is larger. Our results strengthen the evidence on the rich information contained in protein language models' embeddings and their potential downstream impact on the field of bioinformatics. Availability and implementation: https://github.com/Gaius-Augustus/learnMSA, PyPI and Bioconda, evaluation: https://github.com/felbecker/snakeMSA.

MOTIVATION: For more than 25 years, learning-based eukaryotic gene predictors were driven by hidden Markov models (HMMs), which were directly inputted a DNA sequence. Recently, Holst et al. demonstrated with their program Helixer that the accuracy of ab initio eukaryotic gene prediction can be improved by combining deep learning layers with a separate HMM postprocessor. RESULTS: We present Tiberius, a novel deep learning-based ab initio gene predictor that end-to-end integrates convolutional and long short-term memory layers with a differentiable HMM layer. Tiberius uses a custom gene prediction loss and was trained for prediction in mammalian genomes and evaluated on human and two other genomes. It significantly outperforms existing ab initio methods, achieving F1 scores of 62% at gene level for the human genome, compared to 21% for the next best ab initio method. In de novo mode, Tiberius predicts the exon-intron structure of two out of three human genes without error. Remarkably, even Tiberius's ab initio accuracy matches that of BRAKER3, which uses RNA-seq data and a protein database. Tiberius's highly parallelized model is the fastest state-of-the-art gene prediction method, processing the human genome in under 2 hours. AVAILABILITY AND IMPLEMENTATION: https://github.com/Gaius-Augustus/Tiberius.

BACKGROUND: The Earth Biogenome Project has rapidly increased the number of available eukaryotic genomes, but most released genomes continue to lack annotation of protein-coding genes. In addition, no transcriptome data is available for some genomes. RESULTS: Various gene annotation tools have been developed but each has its limitations. Here, we introduce GALBA, a fully automated pipeline that utilizes miniprot, a rapid protein-to-genome aligner, in combination with AUGUSTUS to predict genes with high accuracy. Accuracy results indicate that GALBA is particularly strong in the annotation of large vertebrate genomes. We also present use cases in insects, vertebrates, and a land plant. GALBA is fully open source and available as a docker image for easy execution with Singularity in high-performance computing environments. CONCLUSIONS: Our pipeline addresses the critical need for accurate gene annotation in newly sequenced genomes, and we believe that GALBA will greatly facilitate genome annotation for diverse organisms.

MOTIVATION: The comparison of genomes using models of molecular evolution is a powerful approach for finding, or toward understanding, functional elements. In particular, comparative genomics is a fundamental building brick in annotating ever larger sets of alignable genomes completely, accurately and consistently. RESULTS: We here present our new program ClaMSA that classifies multiple sequence alignments using a phylogenetic model. It uses a novel continuous-time Markov chain machine learning layer, named CTMC, whose parameters are learned end-to-end and together with (recurrent) neural networks for a learning task. We trained ClaMSA discriminatively to classify aligned codon sequences that are candidates of coding regions into coding or non-coding and obtained four times fewer false positives for this task on vertebrate and fly alignments than existing methods at the same true positive rate. ClaMSA and the CTMC layer are general tools that could be used for other machine learning tasks on tree-related sequence data. AVAILABILITY AND IMPLEMENTATION: Freely from https://github.com/Gaius-Augustus/clamsa. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

BACKGROUND: An important initial phase of arguably most homology search and alignment methods such as required for genome alignments is seed finding. The seed finding step is crucial to curb the runtime as potential alignments are restricted to and anchored at the sequence position pairs that constitute the seed. To identify seeds, it is good practice to use sets of spaced seed patterns, a method that locally compares two sequences and requires exact matches at certain positions only. RESULTS: We introduce a new method for filtering alignment seeds that we call geometric hashing. Geometric hashing achieves a high specificity by combining non-local information from different seeds using a simple hash function that only requires a constant and small amount of additional time per spaced seed. Geometric hashing was tested on the task of finding homologous positions in the coding regions of human and mouse genome sequences. Thereby, the number of false positives was decreased about million-fold over sets of spaced seeds while maintaining a very high sensitivity. CONCLUSIONS: An additional geometric hashing filtering phase could improve the run-time, accuracy or both of programs for various homology-search-and-align tasks.

BACKGROUND: BRAKER is a suite of automatic pipelines, BRAKER1 and BRAKER2, for the accurate annotation of protein-coding genes in eukaryotic genomes. Each pipeline trains statistical models of protein-coding genes based on provided evidence and, then predicts protein-coding genes in genomic sequences using both the extrinsic evidence and statistical models. For training and prediction, BRAKER1 and BRAKER2 incorporate complementary extrinsic evidence: BRAKER1 uses only RNA-seq data while BRAKER2 uses only a database of cross-species proteins. The BRAKER suite has so far not been able to reliably exceed the accuracy of BRAKER1 and BRAKER2 when incorporating both types of evidence simultaneously. Currently, for a novel genome project where both RNA-seq and protein data are available, the best option is to run both pipelines independently, and to pick one, likely better output. Therefore, one or another type of the extrinsic evidence would remain unexploited. RESULTS: We present TSEBRA, a software that selects gene predictions (transcripts) from the sets generated by BRAKER1 and BRAKER2. TSEBRA uses a set of rules to compare scores of overlapping transcripts based on their support by RNA-seq and homologous protein evidence. We show in computational experiments on genomes of 11 species that TSEBRA achieves higher accuracy than either BRAKER1 or BRAKER2 running alone and that TSEBRA compares favorably with the combiner tool EVidenceModeler. CONCLUSION: TSEBRA is an easy-to-use and fast software tool. It can be used in concert with the BRAKER pipeline to generate a gene prediction set supported by both RNA-seq and homologous protein evidence.

BACKGROUND: Vast amounts of next generation sequencing RNA data has been deposited in archives, accompanying very diverse original studies. The data is readily available also for other purposes such as genome annotation or transcriptome assembly. However, selecting a subset of available experiments, sequencing runs and reads for this purpose is a nontrivial task and complicated by the inhomogeneity of the data. RESULTS: This article presents the software VARUS that selects, downloads and aligns reads from NCBI's Sequence Read Archive, given only the species' binomial name and genome. VARUS automatically chooses runs from among all archived runs to randomly select subsets of reads. The objective of its online algorithm is to cover a large number of transcripts adequately when network bandwidth and computing resources are limited. For most tested species VARUS achieved both a higher sensitivity and specificity with a lower number of downloaded reads than when runs were manually selected. At the example of twelve eukaryotic genomes, we show that RNA-Seq that was sampled with VARUS is well-suited for fully-automatic genome annotation with BRAKER. CONCLUSIONS: With VARUS, genome annotation can be automatized to the extent that not even the selection and quality control of RNA-Seq has to be done manually. This introduces the possibility to have fully automatized genome annotation loops over potentially many species without incurring a loss of accuracy over a manually supervised annotation process.

MOTIVATION: Gene finding in eukaryotic genomes is notoriously difficult to automate. The task is to design a work flow with a minimal set of tools that would reach state-of-the-art performance across a wide range of species. GeneMark-ET is a gene prediction tool that incorporates RNA-Seq data into unsupervised training and subsequently generates ab initio gene predictions. AUGUSTUS is a gene finder that usually requires supervised training and uses information from RNA-Seq reads in the prediction step. Complementary strengths of GeneMark-ET and AUGUSTUS provided motivation for designing a new combined tool for automatic gene prediction. RESULTS: We present BRAKER1, a pipeline for unsupervised RNA-Seq-based genome annotation that combines the advantages of GeneMark-ET and AUGUSTUS. As input, BRAKER1 requires a genome assembly file and a file in bam-format with spliced alignments of RNA-Seq reads to the genome. First, GeneMark-ET performs iterative training and generates initial gene structures. Second, AUGUSTUS uses predicted genes for training and then integrates RNA-Seq read information into final gene predictions. In our experiments, we observed that BRAKER1 was more accurate than MAKER2 when it is using RNA-Seq as sole source for training and prediction. BRAKER1 does not require pre-trained parameters or a separate expert-prepared training step. AVAILABILITY AND IMPLEMENTATION: BRAKER1 is available for download at http://bioinf.uni-greifswald.de/bioinf/braker/ and http://exon.gatech.edu/GeneMark/ CONTACT: [email protected] or [email protected] SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

MOTIVATION: As the tree of life is populated with sequenced genomes ever more densely, the new challenge is the accurate and consistent annotation of entire clades of genomes. We address this problem with a new approach to comparative gene finding that takes a multiple genome alignment of closely related species and simultaneously predicts the location and structure of protein-coding genes in all input genomes, thereby exploiting negative selection and sequence conservation. The model prefers potential gene structures in the different genomes that are in agreement with each other, or-if not-where the exon gains and losses are plausible given the species tree. We formulate the multi-species gene finding problem as a binary labeling problem on a graph. The resulting optimization problem is NP hard, but can be efficiently approximated using a subgradient-based dual decomposition approach. RESULTS: The proposed method was tested on whole-genome alignments of 12 vertebrate and 12 Drosophila species. The accuracy was evaluated for human, mouse and Drosophila melanogaster and compared to competing methods. Results suggest that our method is well-suited for annotation of (a large number of) genomes of closely related species within a clade, in particular, when RNA-Seq data are available for many of the genomes. The transfer of existing annotations from one genome to another via the genome alignment is more accurate than previous approaches that are based on protein-spliced alignments, when the genomes are at close to medium distances. AVAILABILITY AND IMPLEMENTATION: The method is implemented in C ++ as part of Augustus and available open source at http://bioinf.uni-greifswald.de/augustus/ CONTACT: [email protected] or [email protected] information: Supplementary data are available at Bioinformatics online.

BACKGROUND: The identification of protein-protein interaction sites is a computationally challenging task and important for understanding the biology of protein complexes. There is a rich literature in this field. A broad class of approaches assign to each candidate residue a real-valued score that measures how likely it is that the residue belongs to the interface. The prediction is obtained by thresholding this score.Some probabilistic models classify the residues on the basis of the posterior probabilities. In this paper, we introduce pairwise conditional random fields (pCRFs) in which edges are not restricted to the backbone as in the case of linear-chain CRFs utilized by Li et al. (2007). In fact, any 3D-neighborhood relation can be modeled. On grounds of a generalized Viterbi inference algorithm and a piecewise training process for pCRFs, we demonstrate how to utilize pCRFs to enhance a given residue-wise score-based protein-protein interface predictor on the surface of the protein under study. The features of the pCRF are solely based on the interface predictions scores of the predictor the performance of which shall be improved. RESULTS: We performed three sets of experiments with synthetic scores assigned to the surface residues of proteins taken from the data set PlaneDimers compiled by Zellner et al. (2011), from the list published by Keskin et al. (2004) and from the very recent data set due to Cukuroglu et al. (2014). That way we demonstrated that our pCRF-based enhancer is effective given the interface residue score distribution and the non-interface residue score are unimodal.Moreover, the pCRF-based enhancer is also successfully applicable, if the distributions are only unimodal over a certain sub-domain. The improvement is then restricted to that domain. Thus we were able to improve the prediction of the PresCont server devised by Zellner et al. (2011) on PlaneDimers. CONCLUSIONS: Our results strongly suggest that pCRFs form a methodological framework to improve residue-wise score-based protein-protein interface predictors given the scores are appropriately distributed. A prototypical implementation of our method is accessible at http://ppicrf.informatik.uni-goettingen.de/index.html.

MOTIVATION: As improved DNA sequencing techniques have increased enormously the speed of producing new eukaryotic genome assemblies, the further development of automated gene prediction methods continues to be essential. While the classification of proteins into families is a task heavily relying on correct gene predictions, it can at the same time provide a source of additional information for the prediction, complementary to those presently used. RESULTS: We extended the gene prediction software AUGUSTUS by a method that employs block profiles generated from multiple sequence alignments as a protein signature to improve the accuracy of the prediction. Equipped with profiles modelling human dynein heavy chain (DHC) proteins and other families, AUGUSTUS was run on the genomic sequences known to contain members of these families. Compared with AUGUSTUS' ab initio version, the rate of genes predicted with high accuracy showed a dramatic increase. AVAILABILITY: The AUGUSTUS project web page is located at http://augustus.gobics.de, with the executable program as well as the source code available for download.

BACKGROUND: Methods of determining whether or not any particular HIV-1 sequence stems - completely or in part - from some unknown HIV-1 subtype are important for the design of vaccines and molecular detection systems, as well as for epidemiological monitoring. Nevertheless, a single algorithm only, the Branching Index (BI), has been developed for this task so far. Moving along the genome of a query sequence in a sliding window, the BI computes a ratio quantifying how closely the query sequence clusters with a subtype clade. In its current version, however, the BI does not provide predicted boundaries of unknown fragments. RESULTS: We have developed Unknown Subtype Finder (USF), an algorithm based on a probabilistic model, which automatically determines which parts of an input sequence originate from a subtype yet unknown. The underlying model is based on a simple profile hidden Markov model (pHMM) for each known subtype and an additional pHMM for an unknown subtype. The emission probabilities of the latter are estimated using the emission frequencies of the known subtypes by means of a (position-wise) probabilistic model for the emergence of new subtypes. We have applied USF to SIV and HIV-1 sequences formerly classified as having emerged from an unknown subtype. Moreover, we have evaluated its performance on artificial HIV-1 recombinants and non-recombinant HIV-1 sequences. The results have been compared with the corresponding results of the BI. CONCLUSIONS: Our results demonstrate that USF is suitable for detecting segments in HIV-1 sequences stemming from yet unknown subtypes. Comparing USF with the BI shows that our algorithm performs as good as the BI or better.

MOTIVATION: Existing coalescent models and phylogenetic tools based on them are not designed for studying the genealogy of sequences like those of HIV, since in HIV recombinants with multiple cross-over points between the parental strains frequently arise. Hence, ambiguous cases in the classification of HIV sequences into subtypes and circulating recombinant forms (CRFs) have been treated with ad hoc methods in lack of tools based on a comprehensive coalescent model accounting for complex recombination patterns. RESULTS: We developed the program ARGUS that scores classifications of sequences into subtypes and recombinant forms. It reconstructs ancestral recombination graphs (ARGs) that reflect the genealogy of the input sequences given a classification hypothesis. An ARG with maximal probability is approximated using a Markov chain Monte Carlo approach. ARGUS was able to distinguish the correct classification with a low error rate from plausible alternative classifications in simulation studies with realistic parameters. We applied our algorithm to decide between two recently debated alternatives in the classification of CRF02 of HIV-1 and find that CRF02 is indeed a recombinant of Subtypes A and G. AVAILABILITY: ARGUS is implemented in C++ and the source code is available at http://gobics.de/software.

MOTIVATION: Computational annotation of protein coding genes in genomic DNA is a widely used and essential tool for analyzing newly sequenced genomes. However, current methods suffer from inaccuracy and do poorly with certain types of genes. Including additional sources of evidence of the existence and structure of genes can improve the quality of gene predictions. For many eukaryotic genomes, expressed sequence tags (ESTs) are available as evidence for genes. Related genomes that have been sequenced, annotated, and aligned to the target genome provide evidence of existence and structure of genes. RESULTS: We incorporate several different evidence sources into the gene finder AUGUSTUS. The sources of evidence are gene and transcript annotations from related species syntenically mapped to the target genome using TransMap, evolutionary conservation of DNA, mRNA and ESTs of the target species, and retroposed genes. The predictions include alternative splice variants where evidence supports it. Using only ESTs we were able to correctly predict at least one splice form exactly correct in 57% of human genes. Also using evidence from other species and human mRNAs, this number rises to 77%. Syntenic mapping is well-suited to annotate genomes closely related to genomes that are already annotated or for which extensive transcript evidence is available. Native cDNA evidence is most helpful when the alignments are used as compound information rather than independent positionwise information. AVAILABILITY: AUGUSTUS is open source and available at http://augustus.gobics.de. The gene predictions for human can be browsed and downloaded at the UCSC Genome Browser (http://genome.ucsc.edu).

BACKGROUND: For many types of analyses, data about gene structure and locations of non-coding regions of genes are required. Although a vast amount of genomic sequence data is available, precise annotation of genes is lacking behind. Finding the corresponding gene of a given protein sequence by means of conventional tools is error prone, and cannot be completed without manual inspection, which is time consuming and requires considerable experience. RESULTS: Scipio is a tool based on the alignment program BLAT to determine the precise gene structure given a protein sequence and a genome sequence. It identifies intron-exon borders and splice sites and is able to cope with sequencing errors and genes spanning several contigs in genomes that have not yet been assembled to supercontigs or chromosomes. Instead of producing a set of hits with varying confidence, Scipio gives the user a coherent summary of locations on the genome that code for the query protein. The output contains information about discrepancies that may result from sequencing errors. Scipio has also successfully been used to find homologous genes in closely related species. Scipio was tested with 979 protein queries against 16 arthropod genomes (intra species search). For cross-species annotation, Scipio was used to annotate 40 genes from Homo sapiens in the primates Pongo pygmaeus abelii and Callithrix jacchus. The prediction quality of Scipio was tested in a comparative study against that of BLAT and the well established program Exonerate. CONCLUSION: Scipio is able to precisely map a protein query onto a genome. Even in cases when there are many sequencing errors, or when incomplete genome assemblies lead to hits that stretch across multiple target sequences, it very often provides the user with the correct determination of intron-exon borders and splice sites, showing an improved prediction accuracy compared to BLAT and Exonerate. Apart from being able to find genes in the genome that encode the query protein, Scipio can also be used to annotate genes in closely related species.

BACKGROUND: Jumping alignments have recently been proposed as a strategy to search a given multiple sequence alignment A against a database. Instead of comparing a database sequence S to the multiple alignment or profile as a whole, S is compared and aligned to individual sequences from A. Within this alignment, S can jump between different sequences from A, so different parts of S can be aligned to different sequences from the input multiple alignment. This approach is particularly useful for dealing with recombination events. RESULTS: We developed a jumping profile Hidden Markov Model (jpHMM), a probabilistic generalization of the jumping-alignment approach. Given a partition of the aligned input sequence family into known sequence subtypes, our model can jump between states corresponding to these different subtypes, depending on which subtype is locally most similar to a database sequence. Jumps between different subtypes are indicative of intersubtype recombinations. We applied our method to a large set of genome sequences from human immunodeficiency virus (HIV) and hepatitis C virus (HCV) as well as to simulated recombined genome sequences. CONCLUSION: Our results demonstrate that jumps in our jumping profile HMM often correspond to recombination breakpoints; our approach can therefore be used to detect recombinations in genomic sequences. The recombination breakpoints identified by jpHMM were found to be significantly more accurate than breakpoints defined by traditional methods based on comparing single representative sequences.

BACKGROUND: In order to improve gene prediction, extrinsic evidence on the gene structure can be collected from various sources of information such as genome-genome comparisons and EST and protein alignments. However, such evidence is often incomplete and usually uncertain. The extrinsic evidence is usually not sufficient to recover the complete gene structure of all genes completely and the available evidence is often unreliable. Therefore extrinsic evidence is most valuable when it is balanced with sequence-intrinsic evidence. RESULTS: We present a fairly general method for integration of external information. Our method is based on the evaluation of hints to potentially protein-coding regions by means of a Generalized Hidden Markov Model (GHMM) that takes both intrinsic and extrinsic information into account. We used this method to extend the ab initio gene prediction program AUGUSTUS to a versatile tool that we call AUGUSTUS+. In this study, we focus on hints derived from matches to an EST or protein database, but our approach can be used to include arbitrary user-defined hints. Our method is only moderately effected by the length of a database match. Further, it exploits the information that can be derived from the absence of such matches. As a special case, AUGUSTUS+ can predict genes under user-defined constraints, e.g. if the positions of certain exons are known. With hints from EST and protein databases, our new approach was able to predict 89% of the exons in human chromosome 22 correctly. CONCLUSION: Sensitive probabilistic modeling of extrinsic evidence such as sequence database matches can increase gene prediction accuracy. When a match of a sequence interval to an EST or protein sequence is used it should be treated as compound information rather than as information about individual positions.