Marinka Žitnik

Medea, an AI agent, performs therapeutic reasoning across biological contexts by preserving context, verifying steps, and reconciling evidence, outperforming existing models in multiple domains.

This paper introduces ClinVar-BERT, a model that prioritizes variants for clinical review by analyzing variant summaries and distinguishing functional evidence, achieving high accuracy in predicting loss of function.

The paper uses language models to analyze ClinVar variant summaries, identifying evidence gaps in Variants of Uncertain Significance and reclassifying some as likely benign or pathogenic.

CONCERT, a niche-aware generative model, predicts spatial perturbation responses more accurately than existing methods across multiple tissues and experimental setups.

PDGrapher predicts combinatorial therapeutic targets to reverse disease phenotypes using a causally inspired GNN, outperforming existing methods in cell line and genetic perturbation datasets.

SHEPHERD uses few-shot learning and a knowledge graph to improve diagnosis of rare genetic diseases using simulated data and real-world patient cohorts.

ATOMICA learns transferable representations of intermolecular interactions across diverse molecular types, enabling cross-modal comparisons and experimentally testable hypotheses.

Fleming is an AI agent that identifies lead compounds for Mycobacterium tuberculosis by combining discriminative and generative models, achieving higher discrimination and hit rates than alternative approaches.

This paper shows that standard protein language models do not support evolutionary reasoning, but a new model trained with phylogenetic supervision performs better on tree reconstruction and clustering.

This paper introduces a multimodal foundation model for predicting protein phenotypes using sequence, structure, and natural language, showing competitive performance and novel insights into disease-related proteins.

This paper evaluates how well deep learning models generalize to new molecular sequences using a new framework called SPECTRA across 18 datasets and 19 models.

PocketGen generates high-fidelity protein pockets with superior binding affinity and structural validity, operating ten times faster than physics-based methods.

This paper introduces FoldMark, a watermarking method for protein generative models that ensures biosecurity by enabling user tracing and maintaining structural quality.

This paper introduces a geometric deep learning model for context-aware protein representations, improving therapeutic target nomination in immune-related diseases.

The paper introduces Therapeutics Commons (TDC-2), a resource integrating cell-type-specific biomarkers with machine learning tasks for drug discovery.

This paper introduces PrimeKG, a knowledge graph integrating disease-related data to support precision medicine by connecting molecular, genetic, and clinical information.

This paper introduces a multiscale interactome framework to explain drug mechanisms by integrating drugs, diseases, proteins, and functions, showing drugs often target functionally related proteins rather than disease-associated ones.

MARS identifies and annotates known and novel cell types in single-cell RNA-seq data by transferring latent representations across datasets and generating interpretable names for new cell types.

OnClass classifies single-cell data into Cell Ontology terms and identifies marker genes for all categories, even those absent in training data.

The paper studies interactome resilience across 1,840 species, finding that interactomes become more resilient over evolution, correlating with environmental adaptability.

The paper introduces a multi-task learning framework for biomedical named entity recognition that improves performance by sharing information across entity types.

Network Enhancement improves the signal-to-noise ratio of weighted biological networks, enhancing downstream analysis performance across multiple applications.

This paper introduces a method to prioritize network communities for experimental validation, showing significant improvement in prioritization efficiency.

The paper introduces Decagon, a graph convolutional network that predicts polypharmacy side effects by modeling drug interactions and side effects as a multimodal graph, achieving up to 69% better performance than baselines.

The paper analyzes 519 disease pathways in the human interactome, finding that most do not form single connected components and that existing methods struggle with disconnected pathways.