Machine Learning in Bioinformatics

This study evaluates seven Genomic Foundation Models across 52 genomic tasks, finding that random baselines are strong and tokenizer choices significantly impact performance.

OPUS-GO provides sequence-level and residue-level annotations for proteins and RNA sequences using a biological language model and modified MIL strategy, improving accuracy and interpretability.

EnzymeCAGE is a geometric foundation model trained on enzyme-reaction pairs to predict enzyme function, de-orphan reactions, identify catalytic sites, and reconstruct biosynthetic pathways.

The paper develops a deep learning model called PepMSND to predict peptide blood stability using multi-level feature engineering and a comprehensive database of 635 samples, achieving high accuracy and AUC.

GeSite identifies nucleic acid-binding residues using a domain-adaptive protein language model and explainable geometric deep learning, showing superior performance on benchmark datasets.

This paper introduces L2G, a method that repurposes pre-trained language models for genomics tasks, achieving better performance than specialized models on multiple benchmarks.

The paper introduces BetaDescribe, a model that generates rich textual descriptions of proteins, enhancing understanding of their functions and properties.

This paper introduces AIDO.DNA, a DNA language model that improves representation of DNA sequences across multiple tasks by scaling model size and training data.

This paper proposes a dynamic model selection approach for genomic language models, showing improved performance through multi-model fusion and identifying limitations related to motif reliance.

This paper predicts PCR amplification outcomes using a deep learning framework combining BERT tokenization and a CNN-BiLSTM model, achieving 84.8% accuracy and improved sensitivity, specificity, and AUC.

This study uses protein language models to rank protein-peptide binding affinities without labeled data or structural information, showing agreement with experimental IC50 values across five datasets.

Fast-Part partitions biological sequence data efficiently while maintaining homology separation and stratification ratios.

The paper introduces ISM, a sequence-only model that improves upon ESM2 by incorporating structural information, achieving better performance on mutation stability and structure prediction tasks.

PLM-interact improves PPI prediction by jointly encoding protein pairs, achieving 16-28% AUPR gains across multiple species.

This paper introduces a hybrid machine learning model for protein classification using amino acids, achieving 95% accuracy and outperforming existing methods by 15%.

DeepEST predicts bacterial protein functions using a multimodal deep learning framework combining gene expression, localization, and structure data, outperforming existing methods on a 25-species benchmark.

This paper introduces LC-PLM, a protein language model based on BiMamba-S architecture, showing improved performance on protein tasks compared to Transformer-based models.

This paper introduces two tools, HAM and HAC, to preprocess protein sequences by masking homologous regions and resolving annotation conflicts, improving machine learning model integrity.

SpatialPPI 2.0 predicts protein-protein interactions using graph neural networks and inter-residue analysis, outperforming existing methods and working with AlphaFold-predicted structures.

This paper introduces a dual-language foundation model that integrates protein and coding sequences, showing improved performance in discriminative tasks and generating coding sequences with higher protein yield.

This paper develops a CNN-based framework to predict plastic-degrading enzymes from environmental sequences, achieving high accuracy and F1 score.

BBATProt predicts protein and peptide functions using explainable deep learning, outperforming existing models in hydrolytic catalysis, peptide activity, and post-translational modification site prediction.

This paper adapts DNA and protein bio-language models for mRNA-focused tasks through various finetuning techniques, showing that medium-scale models perform best for cross-modal knowledge transfer.

This study introduces a sequence-based deep learning method for predicting DNA-binding residues, outperforming existing models and demonstrating interpretability through contrastive learning and graph neural networks.

The paper introduces ProtAC, a framework for cleaning large protein datasets using multimodal learning, achieving improved function prediction and a cleaned UniRef50 dataset.

ProteinAligner improves protein function and property prediction by integrating sequence, structure, and literature text modalities in a multi-modal pretraining framework.

This paper introduces ProPEC, a perplexity-based ensemble model for zero-shot prediction of viral protein fitness, showing improved performance over individual models.

The paper develops pLM-DBPs, a tool for predicting DNA-binding proteins in plants using protein language model embeddings, achieving 83.8% MCC on an independent test set.

BioLLMNet improves RNA-interaction prediction by combining multi-modal language model features using a gating mechanism, achieving state-of-the-art results in RNA-protein, RNA-small molecule, and RNA-RNA interactions.

This paper introduces PROEDIT, a method to remove virus-related knowledge from pre-trained protein models to improve safety in mutation effect predictions without affecting performance on non-virus proteins.