Machine Learning in Bioinformatics

This study identifies information leakage in enzyme-substrate prediction models and datasets, showing performance drops when leakage is removed.

This paper evaluates Evo 2's in-context learning capabilities on genomic sequences, finding strong performance on short sequences but degradation with length, and no benefit from model scaling.

GraPPI develops a self-supervised graph encoder for protein-protein interaction modeling, achieving high performance in binding interface prediction and affinity prediction.

This paper introduces DEFT, a method that combines sequence- and structure-based approaches to improve EC number prediction accuracy and enable high-throughput enzyme repertoire annotation.

SurfGraphPro predicts protein binding sites by integrating protein language models with geometric deep learning on coarse protein surfaces, achieving faster performance than existing methods.

NeoToxPred improves toxicity classification by using a protein language model with carefully curated negative controls, outperforming existing methods on both internal and external benchmarks.

StackHPpred identifies peptide hormones from amino acid sequences using a stacking-based ensemble framework, achieving high accuracy on independent datasets.

Pep-PU-GAN improves peptide function prediction using PU learning and GANs with graph neural networks, achieving high F1 and AUROC scores on neuropeptide classification.

This paper fine-tunes a protein language model on nanobody thermostability data and uses sparse autoencoders to uncover biophysical rules governing stability, identifying both known and novel stabilizing features.

This paper evaluates a self-architecting transformer for protein representation, showing it achieves better performance than existing models with a smaller dataset.

The paper introduces Atlas, a microbiome dataset, and Waypoint, a series of transformer-based models, showing improvements in downstream tasks through pretraining.

ProteinSage uses structural constraints during pretraining to improve protein language modeling with less data, achieving competitive performance on structure-aware tasks.

BioGraphX-RNA predicts RNA subcellular localization using physicochemical graph encoding, showing competitive performance and interpretability across mRNA, miRNA, and lncRNA.

The paper develops a high-quality dataset for protein sub-cellular localisation prediction, showing performance improvements and highlighting data leakage issues in existing models.

FuncSeek improves enzyme function prediction by using contrastive learning with multiple PLMs, outperforming existing methods on benchmark datasets.

DeepTMHMM2 predicts transmembrane protein topology and subcellular location, including re-entrant regions and interfacial helices, and identifies localization across 17 membranes.

TaHL-PTM improves PTM prediction by addressing tokenization issues in decoder-only PLMs, achieving better performance on residue-level tasks.

ProtEnrich enhances protein sequence embeddings with biophysical properties without disrupting their semantic structure, improving performance on structure-related tasks.

The paper introduces InterPET, a benchmark for predicting PETase activity using sequence embeddings and graph architectures, showing that model performance and biological validity are interdependent.

The paper introduces a method using sparse autoencoder features from InterPLM to predict neuropeptide precursors from secreted proteins across species.

This paper introduces a benchmark dataset for enzyme function prediction and shows that current models struggle to distinguish non-functional enzyme variants from functional ones.

PG-LLM benchmarks language models for protein variant ranking, showing they outperform many sequence-based predictors but fall short of specialized models.

The paper evaluates cryptic binding site detection methods by separating coverage from ranking, showing that ranking improvements can significantly enhance recovery.

ETAP-CLF uses ESM3 embeddings and transformer attention to classify proteins into binary categories, achieving high AUROC scores across different biological processes.

ICFinder uses protein language models to identify ion channels and predict permeation residues, outperforming existing methods and providing a public resource.

ESM-ECForest uses ESM-2 embeddings and Random Forest classifiers to predict enzyme functions with high accuracy across EC levels.

This study introduces CrossVOI, a framework for VOC-OR interaction prediction using protein language models and cross-attention, showing improved performance and interpretability of attention mechanisms in human olfactory receptors.

The paper examines the role of singleton protein sequences in training protein language models, finding they contain valuable functional diversity and should be included in training.

This paper develops interpretable biophysical neural networks to predict transcriptional activation domains from protein sequences, revealing how residue properties influence both activation strength and protein abundance.

CLASPP is a unified deep learning model that predicts multiple PTM types from protein sequences, addressing data imbalance and improving prediction performance across organisms.