Srijit Seal

SHOT-CCR improves classification of genetic perturbations in Cell Painting datasets by reducing batch effects through biologically guided adversarial training.

The paper presents a model validation framework for molecular property prediction models in drug discovery, revealing four failure modes and showing that common evaluation protocols overestimate performance.

A human-supervised multi-agent LLM system covered drug repurposing from knowledge graph building to assay analysis in three case studies; retrospective COVID-19 screen AUC-ROC 0.986 without predefined thresholds.

Cell Painting in primary human hepatocytes exposed to 1,085 compounds detects bioactivity at lower concentrations than cytotoxicity assays and predicts cell-based, but not cell-free, ToxCast endpoints across feature extractors.

The paper shows that cell counting can predict compound bioactivity as well as Cell Painting or gene expression, but Cell Painting adds unique value in protein-target-specific benchmarks.

This paper uses transfer learning to identify potent antibacterials from large chemical libraries, finding 54% of tested compounds active against E. coli with sub-micromolar potency.

This study evaluates k-fold n-step forward cross-validation for improving out-of-distribution bioactivity predictions in drug discovery and recommends its use with specific metrics.

This paper uses conditional neural processes to predict adverse drug reactions from sparse data and finds they are well-calibrated and effective across organ systems.

This paper reviews methodological advancements in Cell Painting over a decade, highlighting improvements in protocols, feature extraction, and machine learning applications for biological insights.

This paper develops PKSmart, a computational model to predict human pharmacokinetic parameters using molecular features and animal PK data, achieving performance comparable to industry models.

This paper develops a machine learning model to predict drug-induced liver injury by integrating in vitro and in vivo data with chemical features, achieving improved detection and human-specific sensitivity.

This study develops public classifiers for predicting drug-induced cardiotoxicity using FDA DICTrank data, identifying key chemical and biological features that distinguish toxic from non-toxic compounds.

The paper develops similarity-based merger models that combine chemical structure and cell morphology predictions to improve assay hit calls across multiple biological assays.

The paper combines chemical structure, gene expression, and cell morphology data to improve mitochondrial toxicity prediction in drug discovery.