Machine Learning in Bioinformatics

ProtBLIP2-SST generates open-ended protein function captions by integrating sequence, structure, and text using a two-stage framework.

This paper evaluates pLM-guided data augmentation for protein sequence prediction, showing that it improves performance across tasks with variation levels between 2-30%.

MAERM predicts enzyme-reaction matching relationships using a mixed-attention model, outperforming baselines and enabling enzyme design.

This paper shows that general-purpose language models can detect protein homology without biological pre-training, achieving ROC-AUC 0.79 on a controlled benchmark.

PINPOINT predicts protease-inhibitory activities of fungal SSPs lacking annotated inhibitor domains using protein language models and structural modeling.

The paper develops a new S-acylation prediction tool, SAPPTree, by integrating sequence and structural features from experimentally confirmed sites to improve prediction accuracy.

ProLoc localizes protein functional regions using text guidance, outperforming existing methods on a new benchmark.

The paper evaluates pLM-based biological relation prediction models to determine the sources of their performance, finding that multiple information sources contribute and that aggregate metrics alone cannot establish mechanistic insights.

This paper introduces EnzPlacer, a contrastive learning method that predicts enzyme function from sequence by placing it in known function space, even when the precise function is unknown.

This study explores using reduced amino acid alphabets with sub-word tokenization in protein language models, showing improved efficiency with minimal performance loss.

This study develops machine learning models to predict asparagine deamidation in proteins, showing improved accuracy over existing methods and reducing development costs.

AbTune improves protein language model performance for antibody-related tasks through depth-selective fine-tuning, outperforming baselines in structure prediction and mutation effect prediction.

PRIME is a deep learning framework that predicts metal-binding sites in proteins by integrating evolutionary and structural signals, achieving high accuracy across various metal ions.

The paper evaluates KAN variants and protein embeddings for enzyme classification, showing improved performance with ESM-2 embeddings.

Prot2Prop predicts six protein developability properties using a multitask framework with a frozen ProstT5 encoder and task-specific adapters, achieving strong performance on both classification and regression tasks.

Choragraph uses deep learning to analyze spatial proteomics data, reconstructing missing values and identifying subcellular localization, including dual-residency in Arabidopsis.

The paper introduces LAMBDA, a benchmark to evaluate genomic language models' ability to detect prophage sequences across four complexity levels.

This study develops seqLens, a genomic language model family, and evaluates its performance on phenotypic predictions using diverse pretraining datasets and tokenization strategies.

The paper presents a convolutional neural network for gene prediction using human genome sequences, achieving state-of-the-art results on genes associated with genetic disorders.

GUANinE v1.1 evaluates supervised and unsupervised genomic models, showing each excels on specific tasks and identifying tradeoffs in model performance and resource use.

This paper introduces IonNTXpred, a PLM-based method for predicting ion channel modulators, achieving high accuracy and supporting drug discovery and safety evaluation.

The paper introduces a unified smoothing framework for protein domain bigram models, distinguishing between biologically incompatible and undersampled domain pairs.

EssTFNet predicts human essential genes using a deep learning framework combining adaptive time-frequency analysis and DNA language models, achieving high accuracy and interpretability.

HandShake predicts partner-specific protein-protein binding sites using only sequence data, outperforming existing methods and revealing redundancy inflation in benchmark metrics.

The paper shows that biological meaning in protein embeddings depends on resolution, with different levels of organisation preserving distinct biological relationships.

SPPIDER-seq predicts protein-protein interaction sites by considering partner-specific context, outperforming AlphaFold3 on disordered interfaces and revealing partner-specific patterns in TP53 interactions.

ProtAff predicts protein-protein binding affinity using sequence data, outperforming existing methods on an EGFR benchmark but underperforming structural methods in cross-target discrimination.

This paper introduces a CNN classifier for therapeutic peptide classification using augmented negative sampling, achieving significantly lower false positive rates and competitive performance on external benchmarks.

BALM-PPI predicts protein-protein binding affinity from sequence using a fine-tuned protein language model, achieving accuracy comparable to structure-based methods and providing residue-level explanations.

This work analyzes CAFA5 submissions, showing improved accuracy in protein function prediction and highlights the role of protein language models and 3D structures in performance gains.