Machine Learning in Bioinformatics

This paper introduces a pretraining task for genome language models that predicts evolutionary rate, showing improved performance over sequence-only pretraining.

The paper introduces an AI-driven pipeline using Protein Language Models to discover hidden CLE peptides in plant proteomes, achieving high classification accuracy and identifying previously unannotated candidates.

COPTER improves protein representation by incorporating context using optimal transport, achieving better performance in drug target and perturbation response predictions.

PocketGNN predicts enzyme kcat by integrating local 3D pocket geometry and global sequence semantics, achieving high accuracy and robustness across splits.

PepInter is a deep learning framework that uses pre-trained language models to predict protein-peptide interactions and bioactivity, outperforming existing methods on benchmark datasets.

caRBP-Pred predicts chromatin-associated RNA-binding proteins using short peptide sequences, achieving an AUC of 0.81 and identifying 52 potential caRBPs in mice.

This paper introduces StructFuncNet, a deep learning framework for predicting protein-protein interaction affinity using structural and evolutionary data, achieving state-of-the-art performance across multiple benchmarks.

B-PPI is a bacterial PPI prediction model that outperforms human-centric models on bacterial data with higher AUPRC and F1 scores.

The paper creates a benchmarking dataset for protein function prediction by curating functional annotations from UniProt, enabling training and evaluation of machine learning models.

PatchDNA introduces a flexible, biologically-informed patching method for DNA sequences that outperforms tokenization in DNA benchmarks with smaller models.

This paper introduces a new 20-letter alphabet, TEA, derived from protein language models, enabling efficient homology searches comparable to structure-based methods without structural data.

CrossPPI predicts protein-protein binding affinity using a cross-fusion deep learning model that integrates sequence and structural features, achieving high correlation and low error metrics on a test dataset.

The paper introduces eccDNAMamba, a model for ultra-long eccDNA sequences, showing improved performance in cancer discrimination and copy-number prediction.

SAMAffinity predicts mutation-induced changes in protein-protein binding affinity using a sequence-aware deep learning model, outperforming existing methods on multiple datasets.

The paper develops ACCESS, a multimodal graph neural network to annotate functional dark matter proteins from extreme environments beyond homology-based methods.

Genolator is a multimodal large language model that combines genomic and structural data with natural language to interpret protein function, outperforming existing models in accuracy.

TooTranslator uses language embedding alignment to predict substrates of transport proteins, achieving 80% accuracy with top-100 predictions for unseen substrates.

The paper introduces Profluent-E1, retrieval-augmented protein language models that improve performance by conditioning on homologous sequences, achieving state-of-the-art results in fitness prediction and contact-map prediction.

SignalGen predicts optimal signal peptides for protein expression by integrating mature protein sequences, organism, and localization using a Latent Residual Transformer.

This paper introduces ESMSCOP, a contrast-aware pre-training framework that integrates evolutionary and structural information for improved protein function prediction.

The paper improves prediction of functional protein interface sequences by integrating structural and sequence features using a graph attention model, achieving better generalization across diverse variants.

MotifAE discovers functional sequence patterns from protein language models through an unsupervised framework, improving motif identification and enabling stability prediction.

The paper identifies 15 candidate cytoplasmic retention peptides in human proteins, with 10 having unrelated sequences, and shows that retention sequences are located at specific sites across protein isoforms.

MkAtt-SDN2GO improves protein function prediction by integrating multi-kernel convolution and attention mechanisms, achieving better performance than SDN2GO on Gene Ontology terms.

ProAffinity++ predicts protein-protein binding affinities using sequence and structure representations, outperforming existing methods in accuracy and handling complex cases like antigen-antibody interactions.

This study benchmarks fourteen protein-language models (pLMs) to determine which best captures biological properties, finding that mid-scale models often outperform larger ones for general tasks.

ProteinSight predicts carotenoid-binding sites using a 3D U-Net model, showing high sensitivity and specificity on a curated test set and identifying potential sites on Human Serum Albumin.

DomDiff improves domain boundary identification and classification using a diffusion model with ESM2 embeddings, outperforming existing methods and excelling in rare families.

StoPred predicts protein complex stoichiometry using protein language models and graph attention networks, achieving higher accuracy for both homomeric and heteromeric complexes than existing methods.

AstraPTM2 predicts 39 PTM types on full-length proteins using a transformer model with structural and sequence features, achieving high AUROC and macro-F1 scores.