Felix Becker

• This cumulative dissertation presents software which enhances state-of-the-art accuracy for two profound bioinformatics problems. The tools combine established mathematical frameworks for biological sequence Analysis and modern deep learning techniques. The cornerstone of this work is a novel hidden Markov model layer, which implements parallelized algorithms for supervised and unsupervised training and inference. It can be combined with other layers and efficiently runs on a GPU. This thesis contains three research articles in which deep learning is combined with hidden Markov models - a synergy that is driven by pattern detection in large amounts of biological sequence data, but also the incorporation of strong inductive biases. We pioneered learnMSA, a tool that constructs multiple sequence alignments for families of protein sequences. It is the first aligner for large numbers of sequences to integrate a deep learning model, surpassing the state-of-the-art accuracy of amino acid-based aligners. It offers runtime and accuracy advantages over long-established software when scaling up the number of input proteins and when aligning distantly related sequences. Article I describes the original release of learnMSA and introduces its unique methodology, while Article II features the addition of parameter-rich, pre-trained protein language models, which incorporate rich prior knowledge about proteins. Furthermore, we developed Tiberius, a deep learning model for ab initio gene prediction in eukaryotic species. The tool is fast and, compared to other ab initio predictors, unprecedentedly accurate, as shown in Article III. Tiberius, receiving only a genome as input, matches the accuracy of Pipelines that additionally require complex extrinsic evidence.