The tokenization of audio with neural audio codec models is a vital part of modern AI pipelines for the generation or understanding of speech, alone or in a multimodal context. Traditionally such tokenization models have concentrated on low parameter-count architectures using only components with strong inductive biases. In this work we show that by applying a transformer architecture with large parameter count to this problem, and applying a flexible Finite Scalar Quantization (FSQ) based bottleneck, it is possible to reach state-of-the-art speech quality at extremely low bit-rates of $400$ or $700$ bits-per-second. The trained models strongly out-perform existing baselines in both objective and subjective tests.

MOTIVATION: Next Generation Sequencing technologies make it possible to detect rare genetic variants in individual patients. Currently, more than a dozen software and web services have been created to predict the pathogenicity of variants related with changing of amino acid residues. Despite considerable efforts in this area, at the moment there is no ideal method to classify pathogenic and harmless variants, and the assessment of the pathogenicity is often contradictory. In this article, we propose to use peptides structural formulas of proteins as an amino acid residues substitutions description, rather than a single-letter code. This allowed us to investigate the effectiveness of chemoinformatics approach to assess the pathogenicity of variants associated with amino acid substitutions. RESULTS: The structure-activity relationships analysis relying on protein-specific data and atom centric substructural multilevel neighborhoods of atoms (MNA) descriptors of molecular fragments appeared to be suitable for predicting the pathogenic effect of single amino acid variants. MNA-based Naïve Bayes classifier algorithm, ClinVar and humsavar data were used for the creation of structure-activity relationships models for 10 proteins. The performance of the models was compared with 11 different predicting tools: 8 individual (SIFT 4G, Polyphen2 HDIV, MutationAssessor, PROVEAN, FATHMM, MVP, LIST-S2, MutPred) and 3 consensus (M-CAP, MetaSVM, MetaLR). The accuracy of MNA-based method varies for the proteins (AUC: 0.631-0.993; MCC: 0.191-0.891). It was similar for both the results of comparisons with the other individual predictors and third-party protein-specific predictors. For several proteins (BRCA1, BRCA2, COL1A2, and RYR1), the performance of the MNA-based method was outstanding, capable of capturing the pathogenic effect of structural changes in amino acid substitutions. AVAILABILITY AND IMPLEMENTATION: The datasets are available as supplemental data at Bioinformatics online. A python script to convert amino acid and nucleotide sequences from single-letter codes to SD files is available at https://github.com/SmirnygaTotoshka/SequenceToSDF. The authors provide trial licenses for MultiPASS software to interested readers upon request.