Machine Learning in Bioinformatics

This study compares ensemble and non-ensemble machine learning methods for predicting ubiquitination sites, finding ensemble methods generally outperform non-ensemble ones.

MetaLLM predicts metal binding sites in proteins using a transformer-based model trained on protein sequences, achieving over 90% precision for common metal ions.

Hist2Vec is a kernel-based embedding method for biological sequence classification that outperforms existing methods with >76% and >83% accuracy on DNA and protein datasets, respectively.

The paper introduces CATHe, a method using protein language model embeddings to detect remote homologues for CATH superfamilies, achieving high accuracy and validating predictions with AlphaFold structures.

This paper introduces ProteinNPT, a non-parametric transformer for protein property prediction and design, showing improved performance in label-scarce and multi-task settings.

The paper addresses imbalanced multilabel classification for protein subcellular localization using oversampling techniques to improve minority class prediction performance.

This paper introduces nucleotide augmentation (NTA) to improve machine learning models for protein engineering by generating biologically meaningful sequence data, achieving performance gains even with limited training data.

LambdaPP provides fast, accessible protein feature predictions including structure and function, using AI models and embeddings from protein Language Models.

This paper introduces GOProFormer, a transformer-based method for Gene Ontology protein function prediction, showing superior performance over existing methods and proposing a new benchmark dataset.

The paper explores using protein embeddings for nearest neighbor search to detect remote homology, highlighting challenges with multi-domain proteins and proposing solutions.

The paper introduces Pprint2, a method using convolutional neural networks and evolutionary profiles to predict RNA-interacting residues in proteins, achieving an AUC of 0.82.

PenLight improves protein annotation by integrating 3D structure and sequence data using graph neural networks and contrastive learning.

The paper proposes deep learning models for predicting succinylated lysine residues using biochemical and physicochemical properties, achieving improved sensitivity and specificity through threshold tuning.

The paper trains a deep learning model to predict enzyme thermal adaptation using a dataset of enzymes labeled with optimal growth temperatures.

The paper introduces a software package that computes over 50,000 features for protein/peptide function prediction, including residue-level annotations.

The paper introduces ProteinGLUE, a benchmark suite for evaluating self-supervised protein models, showing pre-training improves downstream tasks but smaller models perform better than larger ones.

LMetalSite predicts metal ion-binding sites from protein sequences using a pretrained language model and multi-task learning, achieving higher accuracy than structure-based methods.

The paper introduces DistilProtBert, a distilled version of ProtBert, which effectively distinguishes real proteins from shuffled sequences with high accuracy.

The paper uses a fine-tuned BERT model to predict specific substrates of transmembrane transport proteins with high accuracy.

TMbed predicts transmembrane regions using protein language model embeddings, achieving high accuracy and low false positives for proteome-wide analysis.

DBpred predicts DNA-interacting residues in proteins using deep learning, achieving improved AUROC scores compared to existing methods.

This paper introduces a deep learning method called STEP to predict virus-host protein-protein interactions using protein sequence embeddings and a Siamese neural network, demonstrating its effectiveness on SARS-CoV-2 and JCV proteins.

The paper presents a computational method to generate synthetic eukaryotic signal peptides using building blocks, validated by Signal-BLAST showing 88% of generated sequences were identified as signal peptides.

This paper introduces DistDom, a deep learning method that predicts protein domain boundaries using 1D sequence features and 2D distance maps, achieving higher accuracy than existing methods on CASP14 data.

KinasePhos 3.0 improves phosphorylation site prediction by expanding kinase classification and building 771 models, achieving average accuracy of 87.2%.

This paper predicts protein sequence profiles using a ProtAlbert transformer without alignment, showing high similarity to HSSP profiles.

The paper uses a deep learning model to represent microbial proteins for alignment-free analysis and visualization, showing it can capture structural and functional features.

The paper introduces mebipred, a machine learning tool for predicting metal-binding proteins from sequence data, achieving 90% accuracy and outperforming existing methods.

PepNN introduces deep attention models for predicting peptide binding sites on proteins using sequence and structural data, achieving high performance on benchmark tests.

LiGIoNs is a pHMM-based method for detecting and classifying ligand-gated ion channels, outperforming existing methods and enabling subfamily classification.