Eric P. Xing

EMBER, a gene-centric foundation model, improves cross-context perturbation-response prediction and genetic-interaction ranking using contextualized gene embeddings.

PertAdapt improves genetic perturbation prediction by adapting single-cell foundation models with a perturbation adapter and adaptive loss, outperforming non-pretrained baselines across multiple datasets.

AIDO.Tissue is a spatial transcriptomics foundation model that uses cell-guided pretraining to improve spatial cell type classification, niche prediction, and density estimation.

AIDO.ModelGenerator enables rapid, reproducible multimodal foundation model development by integrating datasets, pre-trained models, and experiment recipes to improve biological predictions.

This paper introduces an uncertainty-aware discrete diffusion model for protein design, improving sequence generation from 3D structures by incorporating position-specific uncertainty and pretrained models.

This paper benchmarks foundation model representations for predicting cellular perturbation responses, identifying effective representation learning strategies.

The paper introduces contextualized network inference to estimate sample-specific transcriptional networks for 7000 tumors, improving accuracy and enabling novel prognostic subtyping.

The paper introduces MASPR, a machine learning method using protein language models to predict adenylation domain substrate specificities, enabling the prediction of hybrid NRP-PK natural products.

GET is a foundation model that accurately predicts gene expression and regulatory interactions across 213 human cell types using chromatin accessibility and sequence data.

This paper introduces a VQ-VAE-based protein structure tokenizer that balances reconstruction and retrieval, improving alignment with sequence models for better structure prediction.

This paper introduces AIDO.RAGPLM and AIDO.RAGFold, models that improve protein structure prediction by integrating retrieved MSA with language models, achieving faster performance and better accuracy with limited MSA data.

This paper introduces AIDO.DNA, a DNA language model that improves representation of DNA sequences across multiple tasks by scaling model size and training data.

AIDO.Protein is a mixture-of-experts model for protein understanding and design, achieving state-of-the-art performance with improved efficiency.

This paper introduces AIDO.RNA, a large-scale foundation model for RNA function and structure prediction, achieving state-of-the-art performance on multiple tasks.

The paper introduces AIDO.Cell, a series of dense Transformer models pre-trained on human single-cell RNA-seq data, achieving state-of-the-art results in clustering, classification, and perturbation modeling.

ProteinAligner improves protein function and property prediction by integrating sequence, structure, and literature text modalities in a multi-modal pretraining framework.

AttentionPert predicts gene expression under multiplexed genetic perturbations and generalizes to unseen conditions, outperforming existing methods in accuracy and handling diverse perturbations.

This paper introduces a method to jointly analyze two independent GWAS datasets with different phenotypes, addressing confounding factors and showing its utility in studying Alzheimer's disease and substance use disorder.

This paper develops a domain adversarial neural network to reduce scRNA-seq data dimensionality while minimizing batch effects, improving cell type separation across datasets.

The paper introduces drPCA, a supervised dimensionality reduction method that improves differentiation in case/control studies by using differential vectors, and shows it outperforms PCA and PLS-DA in cancer gene expression datasets.

This paper introduces a system using deep reinforcement learning to automatically curate a genetic association database by simulating human research behaviors.

This paper introduces a method to remove confounding factors in deep neural networks to improve prediction accuracy in healthcare applications.

This paper develops a truncated-rank sparse linear mixed model for variable selection in heterogeneous datasets, demonstrating its effectiveness on genomic data from plants, mice, and humans.

This paper introduces a Squared-LMM model to correct confounding factors in genomic association mapping, showing improved accuracy in predicting trait-locus associations using synthetic and real data.

This paper introduces a personalized regression method to analyze pan-cancer gene expression data, revealing sample-specific patterns that population-level methods miss.