Lei Xie

PK-MUSE improves pharmacokinetics prediction by combining mechanistic models with neural corrections, achieving better performance across unseen chemical structures.

GEM-GPT generates personalized therapeutic molecules to shift cell-type-specific disease states toward healthy phenotypes, outperforming baselines in chemical perturbation and CRISPR experiments.

This paper evaluates five AI frameworks on real scientific tasks, finding they excel in hypothesis generation and reporting but struggle with depth and reliability in real-world research.

ProMoNet improves drug discovery by predicting protein-ligand binding affinity and kinetics using sequence-based pre-training without 3D structural data.

DrugPTM-Bench provides a large-scale phosphoproteomics dataset for modeling drug-induced PTM regulation, highlighting challenges in imbalanced classification and enabling multi-task predictive tasks.

This study identifies terazosin as a drug that reverses AD-related molecular signatures and preserves noradrenergic circuits in a sex-dependent manner in a rat model of Alzheimer's disease.

This paper evaluates AI frameworks' ability to autonomously conduct scientific research, finding they cannot complete full research cycles but may assist with specific subtasks.

The paper introduces eMOSAIC, a method for uncertainty quantification in multi-target ligand binding affinity predictions, showing it outperforms existing methods in out-of-distribution settings.

PRISM-ML integrates multi-omics and EHR data to identify drug repurposing opportunities in Alzheimer's disease by analyzing molecular heterogeneity and predicting therapeutic targets.

The paper investigates BT-11's effects on AD-like pathology in rats, finding improvements in cognition and reduction in Aβ plaques and neuronal loss.

This study shows RG2833, an HDAC1/3 inhibitor, improves spatial memory in female Alzheimer's model rats but not males, with no effect on amyloid beta or microgliosis.

This paper evaluates the combined treatment of diazoxide and dibenzoylmethane in Alzheimer's transgenic rats, showing improved spatial memory, reduced pathology, and upregulated protective genes.

The paper develops PrePROTAC, an interpretable machine learning model to predict PROTAC-induced protein degradation targets, identifying over 600 novel understudied proteins and proposing PROTAC compounds for Alzheimer's disease targets.

The paper introduces MMAPLE, a semi-supervised meta-learning framework that improves prediction of out-of-distribution molecular interactions by leveraging unlabeled data, achieving significant improvements over baselines.

PortalCG predicts chemical-protein interactions for dark proteins using a deep learning framework that integrates binding site information and meta-learning, outperforming existing methods in out-of-distribution scenarios.

The paper investigates the PGD2 pathway in a transgenic rat model of Alzheimer's disease and shows that timapiprant, a DP2 antagonist, mitigates AD pathology.

The paper compares spatial learning in male and female TgF344-AD rats, finding females perform better despite higher amyloid plaques and microgliosis.

The paper introduces a deep learning framework, CIGER, for predicting gene expression rankings induced by chemicals and applies it to pancreatic cancer drug repurposing, showing improved performance over existing methods.

DeepREAL predicts genome-scale receptor activity of ligand binding for novel chemicals using deep learning, addressing data scarcity and distribution shifts.

This paper introduces MultiDCP, a deep learning model for multiplex chemical phenomics to screen compounds for novel cells and patients, showing improved performance in drug repurposing for Alzheimer's disease.

The paper introduces CODE-AE, a context-aware deconfounding autoencoder, to predict patient-specific drug responses from in vitro screens and discovers personalized therapies and biomarkers.

TranSynergy improves synergistic drug combination prediction by integrating biological knowledge and enhancing interpretability, revealing novel pathways and predicting new combinations for ovarian cancer.

This paper introduces DeepCE, a deep learning framework for chemical-induced gene expression prediction and compound screening, showing improved performance over existing methods and demonstrating its application to COVID-19 drug repurposing.

This paper introduces DISAE, a deep learning framework that predicts ligands for unannotated proteins and applies it to assign ligands to human orphan GPCRs.

The paper introduces PLANS, a self-training method that improves deep learning models for drug discovery by using unlabeled and partially labeled data, achieving significant performance gains in predicting Cytochrome P450 binding activity.

This paper introduces DISAE, a protein pre-training method for deorphanization, which improves generalizability and outperforms existing methods in predicting small molecule binding to orphan proteins.

This study identifies R47H-TREM2 mutation-induced proinflammatory microglial states in human AD and tauopathy mice, showing Akt pathway involvement and potential therapeutic targets.

This paper constructs a disease-centric microbiome network and identifies potential drug targets to modulate the microbiome ecosystem, showing diabetes drugs may be repurposed for microbiome modulation.

The paper develops a Bayesian method to improve gene expression signal detection in LINCS L1000 data, showing better reliability and informativeness than existing approaches.

The paper develops a multi-target screening platform to discover dual-action drugs for cancer therapy, identifying levosimendan as a potential candidate that reduces heart failure risk and inhibits cancer growth.