Machine Learning in Bioinformatics

This paper benchmarks deep learning-based embeddings for predicting peroxisomal protein-protein interactions using sequence data and integrates AlphaFold2-Multimer predictions for improved accuracy.

This paper fine-tunes protein embeddings to improve interpretable function prediction, showing that fine-tuned models outperform pre-trained ones in both known and novel function classification.

This paper evaluates how different performance metrics and generalization definitions affect the assessment of protein regression models.

This paper demonstrates that large language models can predict protein-protein interactions with high accuracy using synthetic data and fine-tuned models.

This paper introduces genomicBERT, a lightweight genome analysis model that improves efficiency and performance through specialized DNA vocabulary and tokenization, validated across diverse case studies.

ULDNA improves DNA-binding site prediction using a LSTM-attention network with pre-trained language models, achieving higher accuracy than existing methods on 1287 proteins.

AnnoPRO improves protein function annotation by using image-like representations and multimodal deep learning, showing better performance for rare protein families.

The paper develops a machine learning model to classify bacterial nanowire proteins using feature engineering, achieving high accuracy with XGBoost.

The paper compares sequence-based and structure-based methods for predicting protein function, finding that structure-based approaches can offer better accuracy when using high-quality protein structures.

DRPbind predicts DNA, RNA, and protein-binding residues in disordered protein sequences using SVM and features from evolutionary, physicochemical, and structural data.

This study introduces two models, PromoR and PromoS, for promoter recognition and classification in E. coli, achieving accuracies of 0.887 and 0.781 respectively.

This paper evaluates prompt tuning as a cost-efficient alternative to fine-tuning protein language models for sequence generation, finding it more efficient in resource usage but less effective than fine-tuning in sequence quality.

MsPBRsP predicts protein binding residues using a language model without MSA tools, achieving higher accuracy and faster performance than existing methods.

This paper introduces a method to improve protein representation learning by retrieving and integrating similar sequences without alignment, achieving faster and more accurate predictions than traditional methods.

HMMSTRTM predicts transmembrane protein topology using a Hidden Markov Model, achieving high accuracy in residue prediction.

The paper evaluates pre-trained language models for antibody tasks, introduces an evaluation benchmark, and investigates their performance and benefits in real-world applications.

The paper introduces KeAP, a knowledge-enhanced auto-encoder for protein representation learning that outperforms existing methods on multiple downstream tasks.

This paper develops a fast method for predicting multi-label subcellular localization of mRNA at genome scale, achieving high AUROC scores using hybrid techniques.

Protein-specific tuning of masking, depth and positional embeddings, rather than more parameters, yields an encoder-decoder model that tops structure and function benchmarks at a fraction of the compute.

PROSTATA predicts protein stability changes from point mutations using a transformer-based model, outperforming existing neural network approaches.

WUREN is a multimodal model that fuses sequence, graph, and structural features to predict protein interaction interfaces and antigen epitopes, achieving state-of-the-art performance.

ProtLM.SCL is a protein language model that improves subcellular localization prediction for Gram-negative bacteria proteins, outperforming PSORTb in accuracy and precision.

The paper shows that codon-based language models outperform amino acid sequence models in protein engineering tasks like species recognition and melting point estimation.

The paper analyzes how a pre-trained BERT model processes nucleotide sequences, identifies k-mer embeddings as key, and proposes an accelerated pre-training method.

The paper improves protein subcellular localization prediction by combining sequence-based language models and structure-based graph neural networks, achieving better performance with lower computational cost.

This study evaluates PTM predictors using protein language models, revealing a training bias from proteomics data that overestimates prediction accuracy and suggests a solution through negative sample filtering.

PRESGENE is a web server that predicts essential genes in non-model organisms using machine learning, addressing dataset imbalance and limited labeled data challenges.

The paper uses machine learning to predict quantitative phenotypes like optimal growth temperature, pH, and oxygen tolerance from protein and nucleic acid sequences.

The paper analyzes the geometric properties of hidden representations in protein language models and identifies three distinct phases that reveal a general strategy for reconstructing missing data.

This paper introduces a neural machine translation model to generate natural language descriptions of protein functions, enabling the discovery of novel functions beyond existing label spaces.