Biology subjects

Marsalkova, E.

Publications and source records attributed to Marsalkova, E..

5 recordsLinked to original sources

bioRxiv2026

Advancing Knotted Protein Design with ESM3: Guided Generation and Topological Insights

Multimodal protein language models have transformed protein design, yet their capacity to capture complex topological features remains poorly understood. We use knotted proteins, rare structures in which the backbone forms a nontrivial topological knot, as a test case to probe this capacity using ESM3, a generative protein language model. Topology-aware guided decoding strongly enriches ESM3 outputs for knotted topologies, producing structures classified as knotted at an 89% success rate (95% CI: 81- 94%), compared to ~0.5% for unguided diffusion-based approaches. A confidence analysis shows that freshly generated artificial knots have lower ESM3 pLDDT and pTM than real knotted proteins evaluated under the same pipeline, motivating a cautious interpretation of generated examples as model samples pending independent validation. In contrast, the robustness analyses on real knotted proteins are high-confidence: on average 84% of the protein sequence must be altered before the knot breaks, and the loss follows a sharp threshold rather than gradual degradation. Strikingly, structural drift accumulates well before topological disruption, suggesting that topology is more robust than specific three-dimensional arrangement. These findings position knotted proteins as a useful probe of how generative protein models represent rare, global structural features.

bioengineering↗

bioRxiv2026

miRBind2 enables sequence-only prediction of miRNA binding and transcript repression

MotivationMicroRNAs (miRNAs) regulate gene expression by guiding Argonaute proteins to partially complementary sites on target RNAs. While classical prediction methods rely on engineered features such as seed match categories, evolutionary conservation, and site context, recent advances in deep learning offer the potential to learn targeting rules directly from sequence. We developed a sequence-based deep learning model that improves miRNA target site prediction, and further validated the learned target site representations by extending the model to gene-level functional repression prediction. ResultsWe introduce miRBind2, a deep learning method for miRNA target site prediction that incorporates a novel pairwise nucleotide representation capturing all possible miRNA-target nucleotide interactions, with a CNN-based architecture. miRBind2 outperforms previous SotA models across four independent datasets from the debiased miRBench benchmark, while using 92% fewer parameters. We show that the convolutional features and weights learned by miRBind2 can be transferred to transcript-level prediction by extending the miRBind2 architecture and fine-tuning it on miRNA perturbation experiments. This miRBind2-3UTR model predicts gene repression from sequence alone. On a dataset of 50,549 miRNA-gene pairs, miRBind2-3UTR significantly outperforms TargetScan. These results show that deep models pretrained on target site data can capture regulatory signals and predict functional repression without requiring conventional engineered biological features. AvailabilityModels and source code are freely available via GitHub (https://github.com/BioGeMT/miRBind_2.0). A publicly available web-tool for novel predictions and visualization is available at : (https://huggingface.co/spaces/dimostzim/BioGeMT-miRBind2) Contactpanagiotis.alexiou@um.edu.mt

bioinformatics↗

bioRxiv2026

AlphaFind v2: Similarity Search in AlphaFold DB and TED Domains across Structural Contexts

The availability of large-scale protein structure collections enables structure-based analysis of their function and evolution beyond what is possible from sequence alone. However, applying three-dimensional structure comparison at scale remains computationally demanding and limits practical exploration of large experimental and predicted collections. This creates a need for fast, structure-based search methods that retain biological relevance while enabling large-scale exploration. In this paper, we present AlphaFind v2, an application for finding structurally similar proteins in the AlphaFold Database (https://alphafold.ebi.ac.uk/) of predicted structures. AlphaFind v2 uses fast pre-filtering via state-of-the-art protein embeddings that preserve structural information, followed by refinement with US-align. The application presents multiple complementary search modes, including (i) search over full protein chains, (ii) search aware of the AlphaFold pLDDT metric, restricting similarity computation to the most stable and structurally relevant regions, (iii) search over protein domains from the TED database (https://ted.cathdb.info/), and (iv) a multidomain search mode, combining multiple chain-level domain matches within a single score and alignment. The application accepts protein identifiers and returns similar proteins with metrics, rich metadata, and interactive superpositions. AlphaFind v2 additionally allows searching within an organism or CATH label and matches the proteins with experimental structures. AlphaFind v2 is accessible at https://alphafind.ics.muni.cz/. Graphical Abstract O_FIG O_LINKSMALLFIG WIDTH=200 HEIGHT=79 SRC="FIGDIR/small/710735v1_ufig1.gif" ALT="Figure 1"> View larger version (22K): org.highwire.dtl.DTLVardef@5e7ce7org.highwire.dtl.DTLVardef@15a3458org.highwire.dtl.DTLVardef@12299ccorg.highwire.dtl.DTLVardef@9f21af_HPS_FORMAT_FIGEXP M_FIG C_FIG

bioinformatics↗

bioRxiv2025

Balancing Speed and Precision in Protein Folding: A Comparison of AlphaFold2, ESMFold, and OmegaFold

We compared the performance of three widely used protein structure prediction tools--AlphaFold2, ESMFold, and OmegaFold--using a dataset of over 1,300 newly created records from the PDB database. These structures, resolved between July 2022 and July 2024, ensure unbiased evaluation, as they were unavailable during the training of these tools. Using metrics such as root mean square deviation (RMSD), template modeling score (TM-score), and predicted local distance difference test (pLDDT), we found that AlphaFold2 consistently achieves the highest accuracy but depends on high-quality sequence alignments. In contrast, ESMFold and OmegaFold provide faster predictions and excel in challenging cases, such as rapidly evolving or designed proteins with limited sequence homology. Comparing ESMFold and OmegaFold, ESMFold achieves higher confidence scores (pLDDT) and structural similarity (TM-score). OmegaFold is competitive in specific contexts, such as de novo-designed proteins or sequences with limited evolutionary information. Additionally, we demonstrate that machine learning models trained on protein language model embeddings and pLDDT confidence scores can predict potential structure prediction failures, helping to identify challenging cases early in the pipeline.

bioinformatics↗

bioRxiv2024

miRBench: A Comprehensive microRNA Binding Site Prediction Training and Benchmarking Dataset

MotivationMicroRNAs (miRNAs) are crucial regulators of gene expression, but the precise mechanisms governing their binding to target sites remain unclear. A major contributing factor to this is the lack of unbiased experimental datasets for training accurate prediction models. While recent experimental advances have provided numerous miRNA-target interactions, these are solely positive interactions. Generating negative examples in silico is challenging and prone to introducing biases, such as the miRNA frequency class bias identified in this work. Biases within datasets can compromise model generalization, leading models to learn dataset-specific artifacts rather than true biological patterns. ResultsWe introduce a novel methodology for negative sample generation that effectively mitigates the miRNA frequency class bias. Using this methodology, we curate several new, extensive datasets and benchmark several state-of-the-art methods on them. We find that a simple convolutional neural network model, retrained on some of these datasets, is able to outperform state-of-the-art methods. This highlights the potential for leveraging unbiased datasets to achieve improved performance in miRNA binding site prediction. To facilitate further research and lower the barrier to entry for machine learning researchers, we provide an easily accessible Python package, miRBench, for dataset retrieval, sequence encoding, and the execution of state-of-the-art models. AvailabilityThe miRBench Python Package is accessible at https://github.com/katarinagresova/miRBench/releases/tag/v1.0.0

bioinformatics↗