Biology subjects

Stegmayer, G.

Publications and source records attributed to Stegmayer, G..

11 recordsLinked to original sources

bioRxiv2026

On the robustness of scRNA-seq foundation models for plant perturbation response prediction under cross-experiment shift

Foundation models for single-cell transcriptomics promise to learn generalizable representations of cellular states. However, recent evidence suggests they often fail to outperform simple machine learning baselines. Furthermore, their ability to generalize across unseen experimental conditions remains poorly understood, particularly in plants, where rigorous evaluation beyond cell type annotation and batch integration is lacking. To address this, we introduce an Arabidopsis thaliana foundation model, scAraFM, and benchmark it across several perturbation conditions under three increasingly challenging protocols: random splits from a single experiment, replicate-based splits, and cross-experiment transfer learning. We found that random splits overestimate performance by up to 30 points relative to cross-experiment evaluations. Across representation strategies, preserving gene identity consistently outperforms the standard pooled embeddings. Moreover, simple baselines using raw reads remain competitive in single-experiment settings, challenging current claims of universal advantage of foundation models. In contrast, under cross-experiment transfer, pretrained representations show added value, particularly with few labelled samples, suggesting that the benefits of foundation models emerge precisely in the regimes that matter for practical deployment. Overall, our results demonstrate that conclusions about foundation models depend critically on the evaluation design, and that preserving per-gene structure aids generalization in downstream tasks, supporting robust predictions across unseen experimental contexts.

bioinformatics↗

bioRxiv2026

Homology-aware cross-validation strategies for generalization assessment in RNA structure prediction

RNA secondary structure prediction is a fundamental challenge in bioinformatics, essential for understanding the functional roles of non-coding RNAs. Recently, deep learning models have transformed the field with impressive results, leading to critical discussions regarding the validity of current cross-validation strategies. On the one hand, traditional random partitioning yields overoptimistic results due to data leakage from uncontrolled homology. On the other hand, removing from the training set all sequences that exhibit even the slightest resemblance to the testing sequences penalizes learning-based methods by requiring generalization to completely out-of-distribution sequences. While it is very simple to remove sequences and retrain a machine learned model, it is very difficult to remove the experimental data used for parameter tuning and the sequences used for the development of classical thermodynamic methods. Thus, these methods often benefit from an implicit knowledge leakage. In this work we critically review existing cross-validation strategies for RNA secondary structure prediction: random splitting, clustering-based splitting, and leaving one RNA family out for testing. We analyze the advantages and limitations of each strategy, also expanding them towards the future directions to ensure fair comparisons across the full range of sequence similarities, with the same rigor for both classical and learning-based methods. Data and source code are available at https://github.com/sinc-lab/xvalRNAfolding

bioinformatics↗

bioRxiv2026

emb2dis: a novel protein disorder prediction tool based on ResNets, dilated convolutions & protein language models

Intrinsically disordered proteins (IDPs) play an important role in a wide range of biological functions and are linked to several diseases. Due to technical difficulties and the high cost of experimental determination of disorder in proteins, combined with the exponential increase of unannotated protein sequences, the development of computational methods for disorder prediction became an active area of research in the last few decades. In this work, we present emb2dis, a deep learning model that uses protein language models (pLMs) to predict disorder from sequence. The emb2dis tool is a pre-trained model that receives as input a protein sequence, calculates its pLM embedding and passes it to a deep learning model. In contrast to existing approaches, emb2dis integrates informative sequence representations with a novel architecture that combines residual networks (ResNets) and dilated convolutions. This design effectively enlarges the receptive field of the convolution operation, enabling the model to better capture an extended context of each amino acid. At the output, emb2dis assigns a disorder propensity score to each residue in the sequence. The model was evaluated on datasets from the latest CAID3 blind benchmark for disorder prediction, where it achieved first place in the Disorder-PDB category, exhibiting strong performance with high AUC and Fmax scores. Additionally, it ranked among the top ten methods on the Disorder-NOX dataset. We provide a freely available web-demo for emb2dis and a source code repository for local installation. Weblink for the toolhttps://sinc.unl.edu.ar/web-demo/emb2dis/ The importance of the emb2dis tool is that it provides a new deep learning approach and significant improvements in the prediction of protein disorder, with a simple web interface and graphical output detailing per-residue disorder.

bioinformatics↗

bioRxiv2025

RNAcomp2D: a visual tool to compare multiple predictions of RNA secondary structure

Ribonucleic acids (RNAs) are involved in many important biological processes. In particular, non-coding RNAs are crucial regulators of cellular processes, playing a significant role in gene expression. RNA secondary structure is key to infer their specific function and for understanding how they interact with other molecules. Many computational models have been developed in the last decade to predict the secondary structure, achieving increasingly higher success rates. However, each new method has its own input-output interface, programming language, computational requirements and, sometimes, a dedicated server to run the model or just a source code in a repository. Thus, nowadays it is very hard to obtain predictions from multiple methods and compare them at once. A unified interface is urgently needed, which allows accessing several methods at the same time, visualizing and comparing predictions among them, and also with a reference structure when available. We introduce here RNAcomp2D, a web-based tool that allows users to enter an RNA sequence, or select one from RNAcentral, and obtains predictions of RNA secondary structures using several state-of-the-art methods. Both classical thermodynamic methods and the latest deep learning models are packaged in containers and accessible in an unified website. All the predictions, and the reference structure if available, are shown at the same time in a single graphical interface. Moreover, as new models continue to be developed, this tool is designed to be scalable, allowing the addition of more prediction methods in the future. Web server is available at https://sinc.unl.edu.ar/web-demo/rnacomp2d/. Data and source code are available at https://github.com/sinc-lab/RNAcomp2D

bioinformatics↗

bioRxiv2025

GNN2Pfam: Integrating protein sequence and structure with graph neural networks for Pfam domain annotation

The challenge of establishing the relationship between protein sequences and their function cannot yet be considered completely solved. State-of-the-art annotation of Pfam domains is based on hidden Markov models (HMMs) built from hand-crafted sequence alignments. However, while this approach has been highly successful during the last decades since its proposal, there is yet a very large number of proteins that remain unannotated because there is no possible alignment to already known and functionally characterized sequences, or HMM fails to discriminate between similar domains. Adding structural information using deep and graph neural networks (GNNs) presents an opportunity to build upon existing models in those more challenging cases. GNNs excel at capturing complex relationships in data and can learn a model that shares information across all existing families, thus being able to generalize Pfam domain predictions to novel sequences. In this work we propose GNN2Pfam, an end-to-end GNN-based method for Pfam family domain annotation. Our strategy allows one single model to be trained for all species and families. This novel proposal uses the protein 3D structure together with a sequence representation obtained from a large pre-trained model. The GNN2Pfam model is based on a graph derived from amino acid interactions in the 3D structure, learning both sequential and structural features from this representation. Experiments show that the proposed GNN-based model can clearly outperform the HMM state-of-the-art predictive performance in Pfam domains annotations. These results suggest that GNN models can be the key component of future protein annotation tools. Data and source code are available at https://github.com/efenoy/GNN2Pfam.

bioinformatics↗

bioRxiv2025

ET-Pfam: Ensemble transfer learning for protein family prediction

MotivationDue to the rapid growth of sequence generation, which has surpassed the expert curators ability to manually review and annotate them, the computational annotation of proteins remains a significant challenge in bioinformatics nowadays. The Pfam database contains a large collection of proteins that are nowadays annotated with domain families through multiple sequence alignments and profile Hidden Markov models (pHMMs). However, such computational annotation methods have some limitations such as problems for handling large datasets and the fact that multiple sequence alignments are computationally challenging to compute with high accuracy due to the increase in complexity as the number of sequences and lengths grow. Additionally, each HMM is independently obtained for each family missing the opportunity of learning patterns across families, that is from a complete view of all the dataset. As an alternative, some deep learning (DL) models have been recently proposed, nevertheless with simple representations of the inputs and moderate improvements in performance. ResultsIn this work we present ET-Pfam, a novel approach based on transfer learning and ensembles of multiple DL classifiers to predict functional families in the Pfam database. Several base DL models are first trained using learned representations from a protein large language model, with different hyperparameters to increase diversity. Then, the base models are integrated using classical ensemble strategies and novel voting approaches by learning weights for each model and for each Pfam family. Results demonstrate that the proposed ET-Pfam method can consistently diminish classification error rates compared to individual DL models, boosting prediction performance. Among the novel ensemble strategies presented here, the learned weights by family voting achieved the best performance, with the lowest error rate (7.00%), significantly surpassing the best individual base model error (12.91%) and three competitors of the state- of-the-art on the same Pfam dataset. AvailabilityData and source code are available at https://github.com/sinc-lab/ET-Pfam.

bioinformatics↗

bioRxiv2024

Identification of genes involved in Kranz anatomy evolution of non-model grasses using unsupervised machine learning

Kranz syndrome is a set of leaf anatomical and functional characteristics of species using C4 photosynthesis. The current model for the evolution of C4 photosynthesis from a C3 ancestor proposes a series of gradual anatomical changes followed by a biochemical adaptation of the C4 cycle enzymatic machinery. In this work, leaf anatomical traits from closely related C3, C4 and intermediate species (Proto-Kranz, PK) were analyzed together with gene expression data to discover potential drivers for the establishment of Kranz anatomy using unsupervised machine learning. Species-specific Self-Organizing Maps (SOM) were developed to group features (genes and phenotypic traits) into clusters (neurons) according to their expression along the leaf developmental gradient. The analysis with SOM allowed us to identify candidate genes as enablers of key anatomical traits differentiation related to the area of mesophyll (M) and bundle sheath (BS) cells, vein density, and the interface between M and BS cells. At the same time, we identified a small subset of genes that displaced together with the change in the area of the BS cell along evolution suggesting a salient role in the origin of Kranz anatomy in grasses. HighlightKranz syndrome is a set of leaf anatomical and functional characteristics of species using C4 photosynthesis. We discovered, with a novel machine learning methodology, some key genes for Kranz differentiation.

plant biology↗

bioRxiv2023

sincFold: end-to-end learning of short- and long-range interactions for RNA folding

MotivationCoding and non-coding RNA molecules participate in many important biological processes. Non-coding 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 thanks 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 still leaving a wide margin for improvement. ResultsIn 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 against classical methods and recent deep learning models, showing that it can outperform state-of-the-art methods. AvailabilityThe source code is available at https://github.com/sinc-lab/sincFold (v0.16) and the web access is provided at https://sinc.unl.edu.ar/web-demo/sincFold Contactlbugnon@sinc.unl.edu.ar

bioinformatics↗

bioRxiv2022

gGN: learning to represent graph nodes as low-rank Gaussian distributions

Unsupervised learning of node representations from knowledge graphs is critical for numerous downstream tasks, ranging from large-scale graph analysis to measuring semantic similarity between nodes. This study presents gGN as a novel representation that defines graph nodes as Gaussian distributions. Unlike existing representations that approximate such distributions using diagonal covariance matrices, our proposal approximates them using low-rank perturbations. We demonstrate that this low-rank approximation is more expressive and better suited to represent complex asymmetric relations between nodes. In addition, we provide a computationally affordable algorithm for learning the low-rank representations in an unsupervised fashion. This learning algorithm uses a novel loss function based on the reverse Kullback-Leibler divergence and two ranking metrics whose joint minimization results in node representations that preserve not only node depths but also local and global asymmetric relationships between nodes. We assessed the representation power of the low-rank approximation with an in-depth systematic empirical study. The results show that our proposal was significantly better than the diagonal approximation for preserving graph structures. Moreover, gGN also outperformed 17 methods on the downstream task of measuring semantic similarity between graph nodes.

bioinformatics↗

bioRxiv2020

High precision in microRNA prediction: a novel genome-wide approach based on convolutional deep residual networks

MotivationMicroRNAs (miRNAs) are small non-coding RNAs that have a key role in the regulation of gene expression. The importance of miRNAs is widely acknowledged by the community nowadays, and the precise prediction of novel candidates with computational methods is still very needed. This could be done by searching homologous with sequence alignment tools, but this will be restricted only to sequences very similar to the known miRNA precursors (pre-miRNAs). Further-more, other important properties of pre-miRNAs, such as the secondary structure, are not taken into account by these methods. Many machine learning approaches were proposed in the last years to fill this gap, but these methods were tested in very controlled conditions, which are not fulfilled, for example, when predicting in newly sequenced genomes, where no miRNAs are known. If these methods are used under real conditions, the precision achieved is far from the one published. ResultsThis work provides a novel approach for dealing with the computational prediction of pre-miRNAs: a convolutional deep residual neural network. The proposed model has been tested on several complete genomes of animals and plants, achieving a precision up to 5 times higher than other approaches at the same recall rates. Also, a novel validation methodology is used to ensure that the performance reported can be achieved when using the method on new unknown species. AvailabilityTo provide fast an easy access to mirDNN, a web demo is available here. It can process fasta files with multiple sequences to calculate the prediction scores, and can generate the nucleotide importance plots. The full source code of this project is available here and here. Contactcyones@sinc.unl.edu.ar

bioinformatics↗

bioRxiv2020

HextractoR: an R package for automatic extraction of hairpins from genome-wide data

Extracting stem-loop sequences (hairpins) from genome-wide data is very important nowadays for some data mining tasks in bioinformatics. The genome preprocessing is very important because it has a strong influence on the later steps and the final results. For example, for novel miRNA prediction, all well-known hairpins must be properly located. Although there are some scripts that can be adapted and put together to achieve this task, they are outdated, none of them guarantees finding correspondence to well-known structures in the genome under analysis, and they do not take advantage of the latest advances in secondary structure prediction. We present here an R package for automatic extraction of hairpins from genome-wide data (HextractorR). HextractoR makes an exhaustive and smart analysis of the genome in order to obtain a very good set of short sequences for further processing. Moreover, genomes can be processed in parallel and with low memory requirements. Results obtained showed that HextractoR has effectively outperformed other methods. HextractoR it is freely available at CRAN and Sourceforge.

bioinformatics↗