Bo Wang

The paper shows that conserved motif vocabularies constrain genome access in cell type evolution across species.

A reasoning LLM combining ESM3 embeddings with GO-GPT term predictions produces protein function annotations that experts rated tie-or-better than UniProt in 79% of cases (SFT model).

Orthrus is a Mamba-based RNA foundation model pre-trained with biological augmentations and contrastive learning, achieving better mRNA property prediction than existing models with less fine-tuning data.

This paper introduces a pretraining task for genome language models that predicts evolutionary rate, showing improved performance over sequence-only pretraining.

This paper shows deep learning models outperform baselines when using well-calibrated metrics for genetic perturbation prediction.

BioML-bench evaluates AI agents on end-to-end biomedical ML tasks across four domains, finding agents underperform compared to humans and diverse ML strategies correlate with higher scores.

The paper introduces OncoGAN, a generative AI tool that creates synthetic cancer genomes to address the lack of gold-standard datasets for benchmarking cancer genome analysis tools.

mRNABench evaluates self-supervised mRNA foundation models using 13 datasets and 81 tasks to assess their ability to predict mRNA properties.

This paper introduces a framework to interpret attention mechanisms in genomic transformer models by correlating attention scores with biological annotations, showing how biological features are learned through pre-training and fine-tuning.

This paper uses a large language model to perform zero-shot cell type annotations on scRNA-seq data, showing improved performance over existing models and providing interpretable biological rationales.

This paper introduces SpatialDIVA, a method to disentangle spatial transcriptomics and histopathology imaging data, showing it outperforms baselines in variation separation and batch effect removal.

LUMI-lab autonomously designs ionizable lipids for mRNA delivery, identifying brominated tails and achieving 20.3% gene editing efficacy in mice.

Further pretraining a single-cell foundation model on 30 million spatial profiles, with a protocol-aware mixture-of-experts decoder and a neighbourhood objective, supports multi-slide integration, deconvolution and imputation.

MorphoDiff predicts high-resolution cell morphology changes under different interventions using diffusion models and perturbation embeddings.

This paper introduces DeepMet, a chemical language model that predicts and discovers unknown metabolites using mass spectrometry data.

MethylGPT is a transformer-based model trained on human DNA methylation data that accurately predicts methylation values and improves age and disease prediction.

CpGPT is a foundation model for DNA methylation analysis that outperforms specialized models in aging-related tasks and can impute methylation profiles across diverse conditions.

A generative transformer pretrained on over 33 million cells can be fine-tuned for cell-type annotation, batch and multi-omic integration, perturbation prediction and gene network inference, claiming superior performance.

DeepVelo estimates cell-specific transcriptional kinetics in scRNA-seq data using GCNs, enabling analysis of complex differentiation processes with time-varying rates.

This paper reproduces and improves the reusability of the TCRP method for predicting drug response across preclinical and clinical datasets.

This study investigates the molecular effects of androgen in the human breast using single-cell transcriptome, chromatin, and spatial profiling of transgender men undergoing androgen therapy, revealing changes in gene expression, immune populations, vasculature, and metabolism.

scFormer uses transformers to learn cell and gene embeddings for single-cell data, achieving state-of-the-art results on multiple tasks.

This study evaluates how dataset imbalance affects single-cell RNA sequencing integration and introduces metrics to assess integration stability in imbalanced settings.

BIONIC integrates biological networks using deep learning to improve accuracy and scalability, identifying novel gene interactions in yeast.

This paper introduces a graph neural network model for predicting molecule carcinogenicity, showing improved performance over existing methods.

scDeepHash provides a scalable method for automated cell-type annotation and cell retrieval in scRNA-seq datasets using deep hashing, showing competitive performance against existing methods.

OCAT integrates large-scale scRNA-seq datasets efficiently and achieves state-of-the-art performance in cell type clustering and downstream analyses.

simATAC generates synthetic scATAC-seq datasets with known cell labels to benchmark analytical pipelines.

Network Enhancement improves the signal-to-noise ratio of weighted biological networks, enhancing downstream analysis performance across multiple applications.

SIMLR is a tool for single-cell RNA-seq analysis that improves clustering and visualization through multi-kernel learning.