Biology subjects

Pollastri, G.

Publications and source records attributed to Pollastri, G..

7 recordsLinked to original sources

bioRxiv2026

Developing SCL2205 : A Protein Sequence-based Spatial Modelling Dataset for the Protein Language Model Frontier

Deep learning (DL) has advanced computational genome annotation tasks such as protein sub-cellular localisation (SCL) prediction. Nonetheless, its potential remains underutilised, primarily because of the limited availability of high-quality reference data and suboptimal input preparation strategies. In this study, we develop and analyse a high-quality dataset derived from the latest release of the universal protein knowledgebase (UniProtKB), designed to address existing challenges and support robust DL-based SCL modelling. The dataset was constructed through extensive quality preprocessing to ensure reliability, manual label mapping to enhance the quantity and diversity of the training data, and stringent partitioning to minimise data leakage. We validated the dataset using independent test sets, achieving up to 10.8% performance improvement, measured by the area under the precision-recall curve (PR-AUC), compared to the state-of-the-art (SoTA). Furthermore, we highlighted potential performance metric inflation in existing SoTA predictors by demonstrating, for the first time, at least 4.8% training-to-testing data leakage (pre-sequence representation) when using only 10% of the training set under homology augmentation (augmentation based on sequence similarity database searches; details in Sub-section 2.1), a commonly used data augmentation strategy in DL-based SCL prediction modelling. SCL2205 will efficiently support the development of robust, trustworthy, and generalisable DL-based SCL predictors, while minimising data leakage and promoting reproducibility. It is openly available under the Creative Commons Zero (CC0 1.0) licence on DRYAD and is conveniently deployed as a package on the Python Package Index - p-scldata.

bioinformatics↗

bioRxiv2025

LINKER-Pred: A Public Web Server for Accurate Prediction of Disordered Flexible Linkers in Proteins

Disordered Flexible Linkers (DFLs) are unstructured regions that play critical roles in inter-domain communication and multivalent protein interactions. Despite their biological significance, the accurate identification of DFLs remains challenging due to limited experimental annotations and sparsity of dedicated prediction tools. Here we introduce LINKER-Pred, a publicly available web server featuring two convolutional neural network-based predictors trained on a novel large-scale dataset of linkers connecting folded domains (DLD dataset) and DisProt linkers. LINKER-Pred2 combines ProtTrans and MSA-Transformer embeddings within an ensemble CNN framework, achieving state-of-the-art performance on CAID2 and CAID3 benchmarks. LINKER-Pred-Lite excludes MSA-based features, improving speed while maintaining competitive predictive accuracy. LINKER-Pred predictors offer robust residue-level DFL predictions directly from sequence, providing a scalable solution for DFL annotation across proteomes. The LINKER-Pred web server and associated resources are freely available at https://linkerpred.chemeslab.org/, offering the research community an accessible tool for studying protein disorder and modularity.

bioinformatics↗

bioRxiv2025

Expanding the Landscape of Disordered Flexible Linkers: A Structural and Computational Framework for DLD dataset assembly

Disordered flexible linkers (DFLs) are functional elements found within intrinsically disordered regions that carry out key functions by connecting domains and/or short linear motifs. Understanding the features of DFLs is limited by the lack of comprehensive datasets and accurate predictive models. In this study, we propose a classification for DFLs that includes linkers joining two domains (DLD), a domain and a motif or two short linear motifs. We developed a workflow that allows the systematic identification of DLD-type linkers from protein structures and created a comprehensive dataset known as the DLD dataset. The DLD dataset includes 1640 independent domain linkers (IDLs) which expands currently available linker datasets and annotates related regions such as dependent-domain linkers, intra-domain loops, and termini. Our data collection process integrates missing residue completion and smoothing of short secondary structure stretches enabling to capture a higher number of longer IDLs. We assessed the features of IDLs using t-SNE analysis and protein language model embedding with a CNN-based classifier as well as PCA analysis. IDLs can be distinguished from other disordered and folded protein regions, and their features highly overlap with DisProt Linkers, considered the gold standard for linker annotation. The DLD dataset offers a valuable resource for researchers seeking to investigate the features of disordered flexible linkers and to improve the accuracy and generalizability of DFL predictive models. The DLD dataset is available via an interactive web server at https://dld.chemeslab.org/ where linkers are annotated with sequence and structural features and can be visualized using a structure viewer.

bioinformatics↗

bioRxiv2024

PUNCH2: explore the strategy for Intrinsically Disordered Protein predictor

Intrinsically disordered proteins (IDPs) lack stable three-dimensional structures, which poses significant challenges for their computational prediction. This study introduces PUNCH2, a novel approach for predicting intrinsically disordered regions (IDRs) in proteins. We address key issues such as the scarcity of comprehensive IDR databases, effective feature extraction, and robust model architecture. By integrating sequences from experimental PDB datasets and fully disordered DisProt sequences for training, PUNCH2 achieves superior prediction accuracy. Various sequence embeddings, including One-Hot, MSA-based, and PLM-based methods, were evaluated, with ProtTrans-based embeddings showing the best performance. The optimal model architecture features multiple convolutional layers, enhancing predictive confidence. PUNCH2 and its lighter variant, PUNCH2-light, demonstrate high efficiency and accuracy in benchmarking against top predictors on the CAID2 dataset, offering promising tools for advancing IDP research and understanding.

bioinformatics↗

bioRxiv2023

Random Forest approach for the identification of relationships between epigenetic marks and its application to robust assignment of chromatin states

Structural changes of chromatin modulate access to DNA for all proteins involved in transcription. These changes are linked to variations in epigenetic marks that allow to classify chromatin in different functional states depending on the pattern of these marks. Importantly, alterations in chromatin states are known to be linked with various diseases. For example, there are abnormalities in epigenetic patterns in different types of cancer. For most of these diseases, there is not enough epigenomic data available to accurately determine chromatin states for the cells affected in each of them, mainly due to high costs of performing this type of experiments but also because of lack of a sufficient amount of sample or degradation thereof. In this work we describe a cascade method based on a random forest algorithm to infer epigenetic marks, and by doing so, to reduce the number of experimentally determined marks required to assign chromatin states. Our approach identified several relationships between patterns of different marks, which strengthens the evidence in favor of a redundant epigenetic code.

genomics↗

bioRxiv2020

Protein Profiles: Biases and Protocols

The use of evolutionary profiles to predict protein secondary structure, as well as other protein structural features, has been standard practice since the 1990s. Using profiles in the input of such predictors, in place or in addition to the sequence itself, leads to significantly more accurate predictors. While profiles can enhance structural signals, their role remains somewhat surprising as proteins do not use profiles when folding in vivo. Furthermore, the same sequence-based redundancy reduction protocols initially derived to train and evaluate sequence-based predictors, have been applied to train and evaluate profile-based predictors. This can lead to unfair comparisons since profile may facilitate the bleeding of information between training and test sets. Here we use the extensively studied problem of secondary structure prediction to better evaluate the role of profiles and show that: (1) high levels of profile similarity between training and test proteins are observed when using standard sequence-based redundancy protocols; (2) the gain in accuracy for profile-based predictors, over sequence-based predictors, strongly relies on these high levels of profile similarity between training and test proteins; and (3) the overall accuracy of a profile-based predictor on a given protein dataset provides a biased measure when trying to estimate the actual accuracy of the predictor, or when comparing it to other predictors. We show, however, that this bias can be avoided by implementing a new protocol (EVALpro) which evaluates the accuracy of profile-based predictors as a function of the profile similarity between training and test proteins. Such a protocol not only allows for a fair comparison of the predictors on equally hard or easy examples, but also completely removes the need for selecting arbitrary similarity cutoffs when selecting test proteins. The EVALpro program is available for download from the SCRATCH suite (http://scratch.proteomics.ics.uci.edu).

bioinformatics↗

bioRxiv2018

Porter 5: state-of-the-art ab initio prediction of protein secondary structure in 3 and 8 classes

MotivationAlthough Secondary Structure Predictors have been developed for more than 60 years, current ab initio methods have still some way to go to reach their theoretical limits. Moreover, the continuous effort towards harnessing ever increasing data sets and more sophisticated, deeper Machine Learning techniques, has not come to an end.\n\nResultsHere we present Porter 5, the last release of one of the best performing ab initio secondary structure predictor. Version 5 achieves 84% accuracy (84% SOV) when tested on 3 classes, and 73% accuracy (82% SOV) on 8 classes, on a large independent set, significantly outperforming all the most recent ab initio predictors we have tested.\n\nAvailabilityThe web and standalone versions of Porter5 are available at http://distilldeep.ucd.ie/porter.\n\nContactname@bio.com\n\nSupplementary informationSupplementary data are available at Bioinformatics online.

bioinformatics↗