Machine Learning in Bioinformatics

The paper introduces two transformer-based models for multi-task protein function annotation, achieving high performance on various functional labels.

Squidly predicts catalytic residues using a biology-informed contrastive learning framework, outperforming existing methods and integrating with BLAST for efficient annotation.

HAVEN predicts viral hosts from protein sequences, achieving high accuracy for both common and rare hosts while outperforming larger models.

GeneChat predicts gene function from DNA sequences using a multi-modal large language model, outperforming GPT-4o on standard metrics.

The paper introduces a deep learning model for annotating LPMO families, achieving high accuracy and outperforming existing tools.

SAKPE predicts enzyme kinetic parameters with improved accuracy by incorporating weighted protein representations, outperforming existing methods in kcat, Km, and Ki predictions.

AntiCP3 predicts anticancer proteins using evolutionary information and hybrid methods, achieving an AUC of 0.91.

This paper introduces KaML-ESM2 and KaML-ESMC models to predict protein pKa values, outperforming existing methods but facing challenges with engineered residues.

MPBind predicts protein binding sites for multiple partners using protein language models and equivariant graph neural networks, outperforming existing methods.

The paper introduces a biology-driven tokenizer for DNA sequences, integrates it into transformer models, and compares its performance with alternative tokenization strategies.

The paper compares MMseqs2 clustering methods with traditional ML algorithms on protein sequences and finds MMseqs2 methods significantly outperform.

The paper introduces two sequence-based models to predict kinase-specific phosphorylation sites, showing improved performance and zero-shot capability.

This paper introduces a unified model combining a protein language model with an autoregressive transformer to predict protein-ligand binding sites, achieving broad ligand coverage and improved performance for underrepresented ligands.

KcatNet predicts enzyme catalytic efficiency using geometric deep learning, outperforming existing methods and providing insights into enzyme-substrate interactions.

This paper introduces T6CNN, a model that improves T6SE prediction by integrating signal sequences, evolutionary data, and pre-trained features, achieving higher accuracy, sensitivity, and specificity than existing tools.

The paper enhances protein-protein interaction detection using a fine-tuned transformer on AlphaFold 2 Multimer outputs and benchmarks it against machine learning algorithms.

This paper extends Prot2Token to align protein language models for diverse prediction tasks, improving performance across multiple applications through pre-training and integration with ESM-2.

AstraPTM predicts 25 PTMs in a single pass using a transformer-based framework, achieving high accuracy and sensitivity for both common and rare modifications.

ESMStabP predicts protein thermostability with improved accuracy using embeddings from ESM2 and thermophilic classifications, achieving high R2 and PCC values.

This study uses large language models to rediscover the central dogma by aligning DNA and protein sequences, achieving 81% accuracy in a classification task.

The paper trains sparse autoencoders on a protein language model to identify features that represent protein sequences, revealing both generic and family-specific patterns and their potential biological roles.

AllMetal3D predicts metal ion localization, identity, and coordination geometry in proteins, outperforming existing tools for some metals but limited in coordination geometry classification.

NanoBEP predicts nanobody-protein binding affinity using a random forest model trained on antigen-antibody and monomeric protein complexes, achieving high accuracy.

HIT-EC improves EC number prediction with a hierarchical transformer model, showing better performance and interpretability than existing methods.

The paper introduces ProtRAP-LM, a transformer-based model using language model embeddings to predict membrane contact probability and relative accessibility of proteins, achieving faster and more accurate results than MSA-based methods.

CPI-Pred predicts compound-protein interaction parameters using deep learning, outperforming existing models on unseen compounds and enzymes.

This study shows that transformer models struggle to capture DNA's functional complexity, suggesting inherent limitations that scale cannot overcome.

The paper improves functional protein generation by optimizing likelihood in both amino acid sequence and latent space, outperforming existing models in AMP and MDH generation.

Evolla, a protein-language model, improves functional inference and discovers novel proteins through natural language queries and benchmarking.

This study evaluates nine clustering algorithms for unsupervised protein function annotation using pseudo-amino acid composition features, finding some methods effective for functional family segregation but less so for fine-grained differences.