Machine Learning in Bioinformatics

ProteinCLIP refines pre-trained protein language model embeddings using contrastive learning between amino acid sequences and functional text, achieving state-of-the-art results in protein tasks.

The paper predicts transcriptional activation domain function using graph neural networks, achieving high accuracy and outperforming existing methods.

GP-HTNLoc is a multi-label subcellular localization model for ncRNAs that addresses class imbalance and outperforms existing methods.

Peptriever uses a bi-encoder transformer model to efficiently rank proteins interacting with peptides, achieving high AUC-ROC and AUC-PR scores.

ChatNT is a conversational agent that understands biological sequences and achieves state-of-the-art results on genomics tasks while being accessible to non-coders.

The paper uses sequence composition features to classify paralogous proteins and identifies key amino acids driving classification.

This paper introduces TooT-PLM-ionCT, a framework using six protein language models to classify ion channels and transporters from membrane proteins, achieving strong performance in distinguishing these classes.

GPSite predicts protein binding sites across multiple ligands using a multi-task network, outperforming existing methods and enabling genome-scale annotations.

The paper develops a toxin classifier, BioLMTox, trained on a comprehensive dataset to improve in-silico toxin classification accuracy and recall.

The paper constructs a database of N-phosphorylation sites and develops a predictor achieving high accuracy for human N-phosphorylation types.

This paper introduces a protein language model that predicts mutation impacts and lineage classification in viral genomes, offering faster and more accurate functional predictions than traditional methods.

PROTGOAT improves protein function prediction using an ensemble of pre-trained protein language models, achieving high performance in CAFA 5 benchmark.

ViRNN is a deep learning model that predicts viral hosts from genomic sequences, achieving 90.55% accuracy and outperforming existing methods.

This study evaluates how different tokenization methods affect transformer models' performance on various biological sequence tasks, showing improved accuracy and efficiency with optimized tokenizers.

TraitProtNet uses deep learning to predict traits in fungi from genome data, highlighting protein importance over expression levels through interpretability methods.

The paper introduces a new amino acid similarity scoring method, E-score, based on contextual embeddings from deep learning, showing improved alignment quality compared to BLOSUM matrices.

DeepSS2GO predicts protein functions using secondary structure features, outperforming existing methods and achieving five times faster prediction.

The paper introduces PIC, a deep learning model for predicting human protein essentiality, outperforming existing methods and quantifying essentiality across humans, cell lines, and mice orthologues.

This paper introduces a method to validate protein-protein interfaces using learned representations, achieving 0.91 accuracy on a benchmark dataset.

TemStaPro predicts protein thermostability using sequence embeddings from protein language models, showing good agreement with experimental data.

The paper demonstrates that positional embeddings from protein language models can improve sensitivity in homology detection without sacrificing search speed.

The paper uses protein language models to distinguish human and viral proteins, identifying immune escape mechanisms through error analysis.

CatPred predicts in vitro enzyme kinetic parameters using deep learning, achieving competitive performance with uncertainty quantification on diverse datasets.

This study introduces xCAPT5, a hybrid model combining CAPT5 and XGBoost for PPI prediction, outperforming existing methods on multiple tasks.

The paper introduces GATLncLoc+C&S, a model for predicting LncRNA subcellular localization using a corrective graph attention network, achieving high accuracy on benchmark datasets.

DeePNAP predicts protein-nucleic acid binding affinity and mutation effects from sequence alone, achieving high accuracy using a large heterogeneous dataset.

This paper develops two deep learning models for classifying CRISPR-Cas proteins, achieving high accuracy in distinguishing CAS from non-CAS and sub-classifying CAS9 and CAS12.

The paper develops a computational method to predict protein-protein interactions using dipeptide compositions as key features, showing strong correlation across organisms.

The paper introduces GenomicLLM, a large language model trained on genomic and natural language data, capable of performing various analysis tasks with comparable performance to existing tools.

This paper extends EvoAug, an evolution-inspired data augmentation method for genomic deep learning, to TensorFlow, demonstrating comparable performance to the original PyTorch-based version.