In recent years, inspired by the success of large language models (LLMs) for DNA and proteins, several LLMs for RNA have also been developed. These models take massive RNA datasets as inputs and learn, in a self-supervised way, how to represent each RNA base with a semantically rich numerical vector. This is done under the hypothesis that obtaining high-quality RNA representations can enhance data-costly downstream tasks, such as the fundamental RNA secondary structure prediction problem. However, existing RNA-LLM have not been evaluated for this task in a unified experimental setup. Since they are pretrained models, assessment of their generalization capabilities on new structures is a crucial aspect. Nonetheless, this has been just partially addressed in literature. In this work we present a comprehensive experimental and comparative analysis of pretrained RNA-LLM that have been recently proposed. We evaluate the use of these representations for the secondary structure prediction task with a common deep learning architecture. The RNA-LLM were assessed with increasing generalization difficulty on benchmark datasets. Results showed that two LLMs clearly outperform the other models, and revealed significant challenges for generalization in low-homology scenarios. Moreover, in this study we provide curated benchmark datasets of increasing complexity and a unified experimental setup for this scientific endeavor. Source code and curated benchmark datasets with increasing complexity are available in the repository: https://github.com/sinc-lab/rna-llm-folding/.

MOTIVATION: Coding and noncoding RNA molecules participate in many important biological processes. Noncoding RNAs fold into well-defined secondary structures to exert their functions. However, the computational prediction of the secondary structure from a raw RNA sequence is a long-standing unsolved problem, which after decades of almost unchanged performance has now re-emerged due to deep learning. Traditional RNA secondary structure prediction algorithms have been mostly based on thermodynamic models and dynamic programming for free energy minimization. More recently deep learning methods have shown competitive performance compared with the classical ones, but there is still a wide margin for improvement. RESULTS: In this work we present sincFold, an end-to-end deep learning approach, that predicts the nucleotides contact matrix using only the RNA sequence as input. The model is based on 1D and 2D residual neural networks that can learn short- and long-range interaction patterns. We show that structures can be accurately predicted with minimal physical assumptions. Extensive experiments were conducted on several benchmark datasets, considering sequence homology and cross-family validation. sincFold was compared with classical methods and recent deep learning models, showing that it can outperform the state-of-the-art methods.

MOTIVATION: In contrast to messenger RNAs, the function of the wide range of existing long noncoding RNAs (lncRNAs) largely depends on their structure, which determines interactions with partner molecules. Thus, the determination or prediction of the secondary structure of lncRNAs is critical to uncover their function. Classical approaches for predicting RNA secondary structure have been based on dynamic programming and thermodynamic calculations. In the last 4 years, a growing number of machine learning (ML)-based models, including deep learning (DL), have achieved breakthrough performance in structure prediction of biomolecules such as proteins and have outperformed classical methods in short transcripts folding. Nevertheless, the accurate prediction for lncRNA still remains far from being effectively solved. Notably, the myriad of new proposals has not been systematically and experimentally evaluated. RESULTS: In this work, we compare the performance of the classical methods as well as the most recently proposed approaches for secondary structure prediction of RNA sequences using a unified and consistent experimental setup. We use the publicly available structural profiles for 3023 yeast RNA sequences, and a novel benchmark of well-characterized lncRNA structures from different species. Moreover, we propose a novel metric to assess the predictive performance of methods, exclusively based on the chemical probing data commonly used for profiling RNA structures, avoiding any potential bias incorporated by computational predictions when using dot-bracket references. Our results provide a comprehensive comparative assessment of existing methodologies, and a novel and public benchmark resource to aid in the development and comparison of future approaches. AVAILABILITY: Full source code and benchmark datasets are available at: https://github.com/sinc-lab/lncRNA-folding. CONTACT: [email protected].

MOTIVATION: In precision medicine, next-generation sequencing and novel preclinical reports have led to an increasingly large amount of results, published in the scientific literature. However, identifying novel treatments or predicting a drug response in, for example, cancer patients, from the huge amount of papers available remains a laborious and challenging work. This task can be considered a text mining problem that requires reading a lot of academic documents for identifying a small set of papers describing specific relations between key terms. Due to the infeasibility of the manual curation of these relations, computational methods that can automatically identify them from the available literature are urgently needed. RESULTS: We present DL4papers, a new method based on deep learning that is capable of analyzing and interpreting papers in order to automatically extract relevant relations between specific keywords. DL4papers receives as input a query with the desired keywords, and it returns a ranked list of papers that contain meaningful associations between the keywords. The comparison against related methods showed that our proposal outperformed them in a cancer corpus. The reliability of the DL4papers output list was also measured, revealing that 100% of the first two documents retrieved for a particular search have relevant relations, in average. This shows that our model can guarantee that in the top-2 papers of the ranked list, the relation can be effectively found. Furthermore, the model is capable of highlighting, within each document, the specific fragments that have the associations of the input keywords. This can be very useful in order to pay attention only to the highlighted text, instead of reading the full paper. We believe that our proposal could be used as an accurate tool for rapidly identifying relationships between genes and their mutations, drug responses and treatments in the context of a certain disease. This new approach can certainly be a very useful and valuable resource for the advancement of the precision medicine field. AVAILABILITY AND IMPLEMENTATION: A web-demo is available at: http://sinc.unl.edu.ar/web-demo/dl4papers/. Full source code and data are available at: https://sourceforge.net/projects/sourcesinc/files/dl4papers/. CONTACT: [email protected]. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

MOTIVATION: The importance of microRNAs (miRNAs) is widely recognized in the community nowadays because these short segments of RNA can play several roles in almost all biological processes. The computational prediction of novel miRNAs involves training a classifier for identifying sequences having the highest chance of being precursors of miRNAs (pre-miRNAs). The big issue with this task is that well-known pre-miRNAs are usually few in comparison with the hundreds of thousands of candidate sequences in a genome, which results in high class imbalance. This imbalance has a strong influence on most standard classifiers, and if not properly addressed in the model and the experiments, not only performance reported can be completely unrealistic but also the classifier will not be able to work properly for pre-miRNA prediction. Besides, another important issue is that for most of the machine learning (ML) approaches already used (supervised methods), it is necessary to have both positive and negative examples. The selection of positive examples is straightforward (well-known pre-miRNAs). However, it is difficult to build a representative set of negative examples because they should be sequences with hairpin structure that do not contain a pre-miRNA. RESULTS: This review provides a comprehensive study and comparative assessment of methods from these two ML approaches for dealing with the prediction of novel pre-miRNAs: supervised and unsupervised training. We present and analyze the ML proposals that have appeared during the past 10 years in literature. They have been compared in several prediction tasks involving two model genomes and increasing imbalance levels. This work provides a review of existing ML approaches for pre-miRNA prediction and fair comparisons of the classifiers with same features and data sets, instead of just a revision of published software tools. The results and the discussion can help the community to select the most adequate bioinformatics approach according to the prediction task at hand. The comparative results obtained suggest that from low to mid-imbalance levels between classes, supervised methods can be the best. However, at very high imbalance levels, closer to real case scenarios, models including unsupervised and deep learning can provide better performance.

Feature selection is an extremely important matter in pattern recognition, particularly when a large set of features is available without knowledge about the discriminative information provided by each element. The key issue is to define a criterion in order to rank the features, discarding those features that are less relevant, redundant, or noisy. This depends on the particular task, the classifier and the properties of the data. A frequent approach consists on the use of genetic algorithms guided by the classification accuracy. However they are often not able to provide a solution with both a considerable reduction of dimensionality and high accuracy rate. Here we propose a modified version of a genetic algorithm, introducing a novel local improvement approach based on evolution, which is able to obtain better dimensionality-accuracy trade-off. Experimental results on different well known datasets show the advantages of our proposal.