Tunca Doğan

SELFormerMM integrates multiple molecular modalities to improve property prediction, outperforming single-modality models.

This paper reports on the fourth CAFA challenge, showing improvements in Gene Ontology term prediction but highlighting challenges in new evaluation settings.

CROssBARv2 integrates heterogeneous biomedical data into a knowledge graph with embeddings and LLM-driven tools for predictive modeling and question-answering.

The paper introduces MoCL-GAT, a self-supervised learning framework for molecular representation, achieving state-of-the-art performance on molecular property prediction tasks.

OmniPath integrates molecular knowledge from 168 resources to support multi-omics analysis, offering curated data, predictions, and tools for causal modeling across bulk, single-cell, and spatial data.

ECLIPSE predicts compound-protein interactions using heterogeneous biomedical knowledge graphs and graph neural networks, achieving strong performance on benchmark tasks.

ProtHGT predicts protein functions using a heterogeneous graph transformer that integrates biological knowledge graphs and language models, achieving state-of-the-art performance across GO sub-ontologies.

This study introduces PROBE, a benchmarking framework for evaluating protein representations on function-related tasks, demonstrating its use with ESM2 and three multimodal PLMs.

Domain2GO predicts protein domain-GO term associations using statistical resampling and co-annotation patterns, achieving high performance in molecular function and biological process prediction with low computational cost.

ASCARIS provides a comprehensive representation of single amino acid variations by integrating positional and structural features, showing competitive performance in variant effect prediction.

The paper reannotates ligand binding sites in UniProtKB using ChEBI identifiers and improves search facilities for these annotations.

This study evaluates computational methods for predicting drug-target interactions, finding that random data splits lead to over-optimism, learned embeddings perform well, and PCM models favor compound features.

DRUIDom predicts compound-target interactions using protein domain modularity, identifying 27,032 high-confidence associations and validating them experimentally on LIM kinases.

The paper introduces CROssBAR, a resource that integrates biomedical data and uses deep learning to predict entity relations, displaying results in knowledge graphs for systems-level analysis of diseases like COVID-19.

This study evaluates protein representation learning methods across three tasks: semantic similarity, function prediction, and drug target classification, highlighting their strengths and limitations.

DEEPScreen predicts drug-target interactions using convolutional neural networks with 2-D compound representations, achieving better or comparable performance to existing methods.

The paper reports results from the CAFA3 challenge, showing improved computational protein function prediction and new annotations for over 1000 genes through experimental screens in C. albicans, P. aureginosa, and Drosophila.