Machine Learning in Bioinformatics

This study uses pseudoperplexity to assess whether ProtT5 memorizes its training data, finding detectable but limited memorization.

This paper evaluates two tools for predicting transition metal binding sites in proteins, showing they can identify known sites from apo structures but have limitations in complex cases.

The paper introduces TaxoFormer, a hierarchical transformer model that predicts the full taxonomic lineage of protein sequences using a structured tokenization scheme and achieves accurate lineage prediction on a large dataset.

This paper introduces a method to compress protein language model embeddings using polynomial trajectories and PCA, achieving efficient storage and reconstruction without loss of downstream task performance.

This study develops a framework for predicting protein function and domain annotation at isoform resolution using structural and sequence data, showing strong performance in benchmark tests and capturing domain changes across isoforms.

This paper introduces ESM-MSR, a model fine-tuned for predicting protein stability with multiple mutations, showing improved performance on several benchmarks.

The paper performs a large-scale statistical analysis of sequence-derived biochemical features to distinguish soluble and insoluble proteins, finding that size-related features and negative charge proportion are the most discriminative.

This paper introduces ReCLIP, a transformer-based framework that models protein-protein interactions at the residue level, showing high accuracy in predicting mutation effects and peptide-MHC binding.

The paper evaluates DNA foundation models' ability to predict R-loop-forming sequences, finding they perform well in some settings but struggle with cross-platform and cross-species generalization.

OmniGene-4 is a multi-modal bio-language MoE model that decomposes task performance by router states, achieving strong benchmarks in homology and biological-knowledge tasks while being computationally efficient.

This paper develops AdventML, a model for predicting enzyme optimal temperature using transformer embeddings and resampling, achieving R2 ~0.64 and improved extreme temperature predictions.

The paper introduces PE-TFD, a method using protein embeddings and machine learning to discover novel protein functions, demonstrating its effectiveness in identifying hydrogenases not found by traditional methods.

ProtAttn-QuadNet predicts protein-protein interactions using ProtBERT embeddings, achieving high accuracy on benchmark datasets and outperforming existing methods.

ProSiteHunter improves protein binding site prediction by integrating sequence-based features and a fine-tuned language model, achieving better performance than existing methods.

PEPE efficiently extracts multimodal protein embeddings from language models, enabling scalable and memory-efficient generation of large embedding datasets for downstream biological tasks.

This paper introduces FunBind, a multimodal model for protein function prediction that improves accuracy and enables zero-shot prediction of novel functions.

The paper introduces ProteomeLM-HPI, a language model adapted to predict cross-species host-pathogen interactions using proteome context and fine-tuning.

SSPSPredictor predicts phase-separating proteins using sequence and structure data, showing higher LLPS propensity in intrinsically disordered proteins and pathogenic variants.

This paper introduces ProtEGNN, a model combining protein language models with geometric equivariance to improve protein property and function prediction with fewer parameters than fine-tuned PLMs.

Deep-Palm is a deep learning framework that accurately predicts S-palmitoylation sites by integrating sequence, physicochemical, embedding, and structural features, outperforming existing methods.

OryzaG3 is a rice-specific DNA language model pretrained on a pangenome, achieving competitive performance on a genomic benchmark while improving inference efficiency.

The paper introduces CO_SCPLOWARBONC_SCPLOW, a family of efficient DNA language models that address challenges in genomic sequence modeling by adapting LLM techniques to DNA's unique properties.

CLIPepPI predicts domain-peptide specificity using contrastive learning, achieving competitive performance across benchmarks and enabling proteome-wide analysis.

TREAD improves repeat annotation by using protein embeddings instead of sequence profiles, outperforming HMMER in low-data and high-divergence scenarios.

ProtmRNA transfers knowledge from protein language models to mRNA sequences, achieving competitive performance with less computational resources.

The paper introduces two models, PiCAP and CAPSIF2, to predict protein-carbohydrate interactions at the protein and residue levels, achieving high accuracy and demonstrating substantial interplay across proteomes.

This paper develops a hybrid feature approach combining deep learning and amino acid frequency features to improve protein-protein interaction prediction using a random forest classifier.

The paper investigates genomic language model representations by analyzing sparse auto-encoder features, finding they encode granular sequence syntax and conservation patterns rather than complex regulatory logic.

This study benchmarks long-context genome language models on biosynthetic gene clusters, showing layer selection impacts performance and revealing how models encode biological information.

This paper improves variant effect prediction by modifying protein language models with sparse mechanistic features, achieving significant performance gains across multiple datasets.