Suffix-based data structures are widely used for pattern recognition. In one dimension, there exist suffix automata and suffix trees, whereas in two dimensions only suffix trees are known [1]. A two-dimensional suffix at$(i, j)$is the largest square submatrix starting there and may not be unique. We introduce the CDAIG, an automaton that recognizes such two-dimensional suffixes analogously to the Isuffix Tree [1] and is built by adapting [2]. We evaluated the complexity of the CDAIG by aligning 200 randomly selected 1000 bp-long coding chunks from the Escherichia coli genome with homologs from Enterobacteriaceae species identified by BLAST [3]. Each alignment was truncated to 750 bp, all differences (substitutions and gaps) were counted, and a CDAIG was built for each pair. Figure 1 shows that edge and suffix counts increase with the number of differences, capturing the rise in sequence divergence and reflecting the phylogenetic distance between E. coli and the compared species. Datasets are available at [E. coli dataset], and the full implementation at [CDAIG data structure].

High-throughput crystallography experiments at synchrotron facilities generate massive volumes of raw diffraction data, challenging storage and long-term data management. The Hierarchical Data Format version 5 (HDF5) is widely used to organize these datasets, but its native compressors often provide limited compression ratios. In this study, we systematically evaluate a broad range of lossless compression algorithms on rotational macromolecular crystallography (MX) and serial synchrotron crystallography (SSX) datasets, including HDF5-native coders. Our results show that preprocessing strategies significantly influence compression performance, with some codecs benefiting from bit- or Byteshuffling, while others perform optimally without it. JPEG XL and the algorithm described in [1] achieve the highest weighted average compression ratios across all datasets, outperforming the best HDF5-native methods. These findings provide a benchmark for selecting optimal lossless compression strategies for large-scale crystallographic datasets.

We present AIDetx, a framework for detecting AI-generated text using finite-context models (FCMs) by classifying human and AI-generated text. Unlike deep learning classifiers, which are resource-intensive and less interpretable, AIDetx builds separate compression models for human and AI-generated text (shown in Fig 1), classifying new inputs by determining which model compresses the text more effectively. Evaluations on the HC3 and AI-human-text datasets showed strong performance, achieving F1 scores of 97% and 99%, with notable computational efficiency, no need for GPUs, while offering enhanced interpretability compared to deep learning models.

We explore the flexibility of a generalized version of finite-context models (FCMs) to identify parameter configurations that are well-suited to specific data classes. This optimization is achieved through the use of genetic algorithms, resulting in highly efficient compressors, even without prior knowledge of the structure of the data. To illustrate the idea, we used three datasets, DS1-3, containing the parameters of publicly available large language models, as well as a dataset, DS4, containing data from a high energy physics experiment, namely from the ALICE detector at CERN.

Protein sequence compression has been a topic of research for more than two decades. Despite the obvious advantages due to storage space and transmission time reduction, better compression algorithms also help answering the question: How much information is on a certain amount of data? In this paper, we provide a step forward regarding answering this question for protein sequence data, using combinations of copy models and finite-context models, assisted by a multilayer perceptron. Compared to the state-of-the-art protein sequence compressor, AC2, the proposed approach attains better compression using one tenth of the time and less memory.

We provide a compression-based method that explores metagenomic classification and recursive filtering by similarity for sorting the order of the reads and increase the overall compression. Our main objective is to decrease the size of FASTQ files from metagenomics, without loss of information, while discarding only the information associated to the order of the reads.

Authorship attribution is a classical classification problem. We use it here to illustrate the performance of a compression-based measure that relies on the notion of relative compression. Besides comparing with recent approaches that use multiple discriminant analysis and support vector machines, we compare it with the Normalized Conditional Compression Distance (a direct approximation of the Normalized Information Distance) and the popular Normalized Compression Distance. The Normalized Relative Compression (NRC) attained 100% correct classification in the data set used, showing consistency between the compression ratio and the classification performance, a characteristic not always present in other compression-based measures.

The number of genomic sequences is growing substantially. Besides discarding part of the data, the only efficient possibility for coping with this trend is data compression. We present an efficient compressor for genomic sequences, allowing both reference-free and referential compression. This compressor uses a mixture of context models of several orders, according to two model classes: reference and target. A new type of context model, which is capable of tolerating substitution errors, is introduced. For ensuring flexibility regarding hardware specifications, the compressor uses cache-hashes in high order models. The results show additional compression gains over several specific top tools in different levels of redundancy. The implementation is available at http://bioinformatics.ua.pt/software/geco/.

Finite-context modeling is a powerful tool for compressing and hence for representing DNA sequences. We describe an algorithm to detect genomic regularities, within a blind discovery strategy. The algorithm uses information profiles built using suitable combinations of finite-context models. We used the genome of the fission yeast Schizosaccharomyces pombe strain 972 h-for illustration, unveiling locations of low information content, which are usually associated with DNA regions of potential biological interest.

We describe a compression-based distance for genomic sequences. Instead of using the usual conjoint information content, as in the classical Normalized Compression Distance (NCD), it uses the conditional information content. To compute this Normalized Conditional Compression Distance (NCCD), we need a normal conditional compressor, that we built using a mixture of static and dynamic finite-context models. Using this approach, we measured chromosomal distances between Hominidae primates and also between Muroidea (rat and mouse), observing several insights of evolution that so far have not been reported in the literature.