Machine Learning in Bioinformatics

The paper introduces a semi-supervised learning approach for regulatory sequence prediction using pseudo-labeling, showing strong performance improvements over supervised learning and comparable results to DNABERT2.

The paper argues that AI platforms for predicting protein function risk overprediction and highlights the need for multiple evidence types, using an analysis of YciO protein function predictions in E. coli.

This paper introduces SSEmb, a model that integrates protein sequence and structure to improve variant effect predictions and downstream tasks like binding site prediction.

This paper shows that incorporating protein structure information into protein language models improves their performance on structure-related tasks without requiring structure access during inference.

This paper fine-tunes pre-trained genomic models on genotype-phenotype datasets to improve DNA sequence classification, achieving high accuracy on the HERV dataset.

The paper introduces BioCLIP, a contrastive learning framework for pre-training Protein Structure Models using Protein Language Models, achieving performance comparable to specialized methods on various tasks.

The paper introduces a framework using protein language models to learn sequence-structure-function representations and develops a retrieval system for large-scale protein annotation and function prediction.

This study develops a deep learning model to identify disease-resistance proteins in rice and related species, showing superior performance compared to traditional machine learning methods.

This study applies LoRA and QLoRA to predict protein binding sites using ESM-2 without MSA or structural data, showing improved performance and reduced overfitting.

AlzGenPred uses network features with CatBoost to classify Alzheimer's Disease-associated genes from high-throughput sequencing data with high accuracy.

This paper introduces SPOT, a machine learning model that predicts substrates for transport proteins with high accuracy, outperforming existing methods.

The paper introduces OglyPred-PLM, a model using ProtT5-XL-UniRef50 embeddings to predict human O-linked glycosylation sites, outperforming existing methods on a benchmark dataset.

GUANinE v1.0 provides benchmark datasets for evaluating genomic AI sequence-to-function models across functional genomics and evolutionary biology tasks.

Pairing each residue with its Foldseek 3D-structure token and pre-training an ESM-2-sized model on about 40 million AlphaFold2 structures beat ESM-2 and structure-based baselines on variant-effect and function tasks.

NanoBERTa-ASP predicts nanobody-antigen binding sites using a BERT-based model trained specifically on nanobody sequences, achieving high precision and AUC.

DeepGO-SE improves protein function prediction for distantly related proteins by using a pretrained large language model and a neuro-symbolic model that leverages GO axioms.

This paper introduces cdsBERT, a protein language model trained on codon sequences, which outperforms amino acid-based models in enzyme prediction and reveals codon usage patterns in embeddings.

The paper introduces a computational predictor, iLEC-DNA, for identifying leccDNA using sequence-derived features, outperforming existing methods across multiple species.

This study introduces Prot-SCL, a self-supervised machine learning method for predicting protein subcellular localization from primary sequence, achieving strong performance on both validation and testing sets.

The paper introduces three deep learning models for predicting SUMOylation sites and reports that SUMOnet-3 outperforms existing methods with 75.8% AUPR and 87% AUC.

The paper uses MitoPredictor to predict and analyze the mitochondrial proteome of the ctenophore Mnemiopsis leidyi.

The paper introduces Domain-PFP, a method for protein function prediction using function-aware domain embeddings, which outperforms existing methods and large-scale protein language models in GO prediction tasks.

This paper introduces a small-scale protein language model that achieves comparable performance to larger models with significantly reduced computational requirements.

This paper investigates using a fractal-based image representation and deep learning for predicting protein binding sites but concludes the approach is ineffective.

This study evaluates in silico tools for predicting molecular functions of orphan proteins, finding most methods are descriptive rather than predictive.

DNAGPT pre-trains a modified generative pre-trained transformer (GPT) architecture on over 200 billion mammalian base pairs to jointly process sequence and numerical data across diverse DNA analysis tasks.

SPAT annotates proteins based on their likelihood to be at the cell surface using existing annotations, showing good performance on leukemia and public datasets.

This paper introduces CoupleNet, a model that integrates protein sequence and structure to improve representation learning, achieving better performance in fold classification and function prediction.

A 100-billion-parameter protein language model trained with masked and autoregressive objectives together, topping prior results on 15 of 18 benchmarks and powering a fast structure predictor.

T4SEpp predicts bacterial type IV secreted effectors using protein language models and integrates multiple modules, achieving high sensitivity and specificity.