James Zou

SPHERE enables the use of sensitive datasets in AI and open science by creating synthetic twins that preserve statistical properties and privacy.

The Virtual Biotech uses AI agents to integrate multi-omic and clinical data for drug discovery, identifying genomic features linked to trial success and proposing therapeutic strategies.

The paper presents a benchmarking challenge to predict Alzheimer's disease neuropathology from single-nucleus transcriptomics using human and AI models, finding that top methods achieve near-perfect categorical staging predictions.

Eva is a foundation model for tissue imaging data that learns multi-scale spatial representations and performs various tasks like cross-modal inference and patient stratification.

CellVoyager autonomously analyzes scRNA-seq data to generate new hypotheses and findings, outperforming existing models and producing scientifically interesting results.

Eubiota enables autonomous discovery in the gut microbiome by integrating modular agents and reinforcement learning, identifying key pathways and designing therapeutic interventions.

BrainBridge integrates single-cell transcriptomic data with phenotypes to model brain biology, prioritizing genes and cell types linked to Alzheimer's, aging, and sex differences.

SpatialProp predicts the effects of multi-gene perturbations on whole tissue sections using spatially resolved single-cell transcriptomics and graph neural networks.

Squidiff predicts transcriptomic changes across cell types in response to environmental stimuli using a diffusion model, enabling in silico screening for drug discovery and precision medicine.

The paper introduces CytoTRACE 2, a deep learning framework to assess cell potency from scRNA-seq data, showing improved performance across multiple datasets and applications.

The paper develops a pipeline to annotate protein functions using latent features from protein language models, identifying missing annotations and structurally matching proteins.

This paper introduces InterPLM, a framework using sparse autoencoders to extract interpretable features from protein language models, revealing both known and novel biological concepts.

RobPicker improves macromolecule identification in cryo-ET by addressing label noise and class imbalance through a meta-learning framework, outperforming existing methods in detecting small and rare species.

ROSIE generates multiplex immunofluorescence staining from H&E images by imputing protein expression and localization using deep learning.

DANCE 2.0 automates and interprets single-cell preprocessing by recommending optimal pipelines for various tasks, improving performance across six major analyses.

This study uses spatial multi-omics and deep learning to identify immunotherapy response and resistance signatures in hepatocellular carcinoma.

RAPID is an automated tool using large language models to generate medical reporting checklists, achieving high accuracy on CONSORT and CONSORT-AI datasets.

This study evaluates the effectiveness of RL finetuning o1-mini for biomedical reasoning tasks and finds mixed results in gene classification and small molecule prediction.

SyntheMol-RL generates novel and synthesizable antibiotics by optimizing for antibacterial activity and solubility, outperforming prior models and showing in vitro and in vivo efficacy.

The paper introduces CORAL, a deep generative model for spatial multiomics data that deconvolves low-resolution data into single-cell profiles and identifies spatial domains and cell-cell interactions.

spEMO integrates histopathology and spatial multi-omic data using foundation models, improving performance in spatial analysis and medical report generation.

CellMemory interprets out-of-distribution cells using a bottlenecked Transformer, outperforming pre-trained models and enabling accurate analysis of spatially resolved single-cell data.

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.

Role-specialised LLM agents under an LLM principal investigator, with light human steering, assembled an ESM/AlphaFold-Multimer/Rosetta pipeline and designed 92 SARS-CoV-2 nanobodies; two bound recent variants better.

GenePert uses GenePT embeddings from ChatGPT to predict gene expression changes from perturbations, outperforming existing models across multiple datasets.

The paper develops spatial aging clocks to analyze cell proximity effects in mouse brain aging and rejuvenation, revealing pro-aging and pro-rejuvenating influences of specific cell types.

SPRITE improves spatial gene expression predictions by integrating gene networks and spatial relationships, enhancing clustering and classification in spatial transcriptomics data.

ADMET-AI, a Chemprop-RDKit graph neural network trained on 41 TDC datasets, has the best average leaderboard rank and scores one million molecules in 3.1 hours.

ProteinCLIP refines pre-trained protein language model embeddings using contrastive learning between amino acid sequences and functional text, achieving state-of-the-art results in protein tasks.

TISSUE improves spatial transcriptomics analysis by providing uncertainty-calibrated predictions, enhancing downstream tasks like clustering and differential expression analysis.