Machine Learning in Bioinformatics

This work identifies optimal features and develops a machine learning tool for predicting bacteriocin sequences with high accuracy.

The paper introduces bindAdjust, a method that improves binding predictions by incorporating structural information from AlphaFold2, enhancing recall for small molecule binding and expanding predictions to larger binding sites.

The paper introduces a 3-D structure-based method to predict SUMOylation sites with 81% accuracy and a Matthews Correlation Coefficient of 0.4.

The paper tests preprocessing methods and classifiers for protein function prediction using InterProScan data, finding that preprocessing and stacking improve performance, outperforming CAFA3 competitors.

This paper introduces NeCLAS, a machine learning pipeline that predicts nanoscale interactions across domains, outperforming existing models and showing cross-domain validity.

DBP2Vec predicts DNA-binding proteins using a pre-trained protein language model, outperforming existing methods on benchmark datasets.

The paper introduces ATGO, a deep-learning method that improves Gene Ontology prediction for proteins using self-attention transformers and triplet networks, achieving higher accuracy than existing methods.

The paper introduces a web server with three algorithms for predicting sub-peroxisomal and sub-mitochondrial protein localisation and PTS1 signals, aiming to accelerate research in peroxisomal biology.

The paper predicts enzyme substrate selectivity using space-filling curve-based structural representations of proteins, achieving high accuracy with gradient-boosted classifiers.

The paper introduces Markonv, a convolutional layer that models inter-positional correlations using Markov transition matrices, showing effectiveness in analyzing omics data and decoding electrical signals.

CELL-E generates 2D probability density maps of protein localization using text-to-image transformers, avoiding discrete class annotations.

This paper introduces ESP, a machine learning model predicting enzyme-substrate pairs with over 90% accuracy, outperforming existing models for specific enzyme families.

Combining an ESM-1b encoder with a conditional random field decoder, DeepTMHMM classifies alpha-helical and beta-barrel transmembrane proteins and their topology, outperforming most compared tools, including on signal-peptide cleavage sites.

PLAST uses protein language models and discrete cosine transforms to efficiently detect homologs with high accuracy and speed.

LazyPair predicts protein-protein interactions and their types using random forests trained on residue properties, outperforming existing methods in accuracy and scalability.

This paper introduces an alignment-free protein classification framework using sequence embeddings to infer evolutionary relationships and functional motifs across diverse protein superfamilies.

This study evaluates Transformer models for genomic sequence classification, comparing pre-trained and from-scratch approaches on benchmark datasets.

The paper introduces a method to predict protein classification using specific peptides, achieving high precision and recall on enzyme sequences and applying it to GPCR and ZF families.

This paper develops machine learning models to identify viral spike proteins and classify respiratory virus families from sequence data with high accuracy.

TooT-SC predicts eleven substrate classes of transmembrane transport proteins using a combination of amino acid composition and evolutionary information, achieving high performance.

ProtFinder predicts multiple subcellular locations of proteins using protein interaction networks and biological priors, achieving high accuracy on held-out data.

fastMSA accelerates multiple sequence alignment by using a dual encoder based on a protein language model, achieving a 34-fold speed-up while maintaining homolog recall and compatibility with downstream tasks like AlphaFold.

SBP-SITA is a sequence-based tool for predicting S-itaconation sites, achieving an AUC of 0.73 using an ensemble classifier.

The paper introduces an unsupervised method using Transformer models, UMAP, and spectral clustering to organize protein sequences into biologically meaningful clusters.

The paper introduces Multi-Branch-CNN, a parallel CNN method for classifying ion channel interacting peptides, showing improved performance on novel sequences compared to traditional ML algorithms and other CNN variants.

The paper develops machine learning models using sequence-derived features to predict protein abundance across multiple cell lines and identifies key features for this task.

USPNet is a signal peptide predictor using a deep protein language model that outperforms existing methods, especially in metagenomic data without additional group information.

This study introduces LM-GVP, a deep learning framework combining a protein Language Model and Graph Neural Network to predict protein properties from sequence and structure, outperforming existing methods on fluorescence, protease stability, and Gene Ontology function prediction.

This study introduces AttCNNPPISP, a model combining attention mechanisms and CNNs to predict protein-protein interaction sites, showing improved performance over existing methods.

A single pre-trained protein language model, ESM-1v, scores mutations zero-shot about as well as family-specific DeepSequence across 41 deep mutational scans, with no experimental labels or retraining.