Xingyi Cheng

The paper introduces GB.GeneUnet, a transformer-based model pretrained on multi-species genomes to predict gene expression at megabase-scale contexts, achieving state-of-the-art performance.

This paper introduces VCHarness, an AI system that automatically builds perturbation-response models for virtual cells, outperforming manual methods and reducing development time.

AIDO.Builder automates biomedical model development by generating executable pipelines, selecting strategies, and iteratively refining models without manual intervention.

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

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.

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

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.

ProteinChat is a multi-modal large language model that predicts protein functions from amino acid sequences, outperforming general-purpose LLMs and task-specific models.

MSAGPT improves protein structure prediction by generating virtual MSAs through pre-training, enhancing accuracy in low-MSA scenarios.

This paper introduces scFoundation, a large-scale pretrained model for single-cell transcriptomics, achieving state-of-the-art results in multiple downstream tasks.

This paper investigates optimal training of protein language models by analyzing the relationship between model size, training data, and performance, finding that diverse training data and scaling laws improve results.

A 100-billion-parameter protein language model trained with masked and autoregressive objectives together, topping prior results on 15 of 18 benchmarks and powering a fast structure predictor.

xTrimoGene is a scalable transformer model for scRNA-seq data that reduces computational demands while maintaining accuracy, achieving state-of-the-art results in downstream tasks.

xTrimoDock improves rigid protein docking by combining cross-modal representation learning and spectral algorithms, outperforming existing methods in docking accuracy.