Yufeng Shen

This study shows that larger protein language models can overestimate sequence likelihood, leading to poor fitness prediction, and recommends moderate model sizes for better performance.

The paper develops MDmis, a method using molecular dynamics simulations of intrinsically disordered regions to predict variant pathogenicity, finding Long IDRs show stronger pathogenic effects than Short IDRs.

The paper validates two genes, Cnot1 and Cdc42bpb, as causes of congenital diaphragmatic hernia in mice using patient-specific variants and shows distinct hernia phenotypes.

This paper introduces two protein language models trained on dynamic biophysical data to better predict mutation effects and protein behavior.

MotifAE discovers functional sequence patterns from protein language models through an unsupervised framework, improving motif identification and enabling stability prediction.

PreMode predicts the mode-of-action of missense variants using graph neural networks, achieving state-of-the-art performance in mode-of-action classification.

This study uses AI to show that genome sequencing can predict adverse outcomes after congenital heart defect surgery by identifying damaging genotypes in specific genes.

The paper introduces MisFit, a probabilistic graphical model to estimate selection coefficients of missense variants using population sequence data, showing improved performance in prioritizing disease-associated variants.

The paper examines how different amino acid substitutions at the same KIF1A residue lead to distinct clinical and molecular outcomes in KIF1A-associated neurological disorders.

G2VTCR predicts TCR-antigen binding using graph embeddings of TCR sequences and antigens, outperforming existing methods in classification and clustering.

This study shows that MYRF is essential for mesothelial cell function in lung development and maturation, with its mutation causing congenital diaphragmatic hernia.

The paper identifies rare FOXF1 missense variants in PAH patients, shows dysregulated FOXF1 expression in PAH lungs, and identifies downstream targets linked to PAH pathogenesis.

DeltaSplice predicts splicing-altering mutations and alternative splicing from genomic sequences, outperforming existing methods and identifying disease-related variants in neurodevelopmental disorders.

The paper reports clinical phenotypes and longitudinal adaptive function in KIF1A-associated neurological disorder and correlates them with computational predictions of genotype severity.

The paper shows that disruptions in endosomal trafficking of the Vangl-Celsr polarity complex cause congenital trachea-esophageal malformations in Xenopus, similar to human EA/TEF.

SABRE predicts TCR-antigen recognition using a self-attention model trained on curated databases and experimental data, achieving an AUROC of 0.726.

The paper identifies missense variants in PLS3 causing X-linked congenital diaphragmatic hernia and body wall defects through genetic analysis and functional studies.

The paper identifies rare heterozygous predicted loss of function variants in BSN associated with higher BMI in multiple cohorts.

SHINE uses protein language models to predict pathogenicity of inframe insertion and deletion variants, outperforming existing methods in two test datasets.

This paper introduces gMVP, a graph attention neural network for predicting the functional effect of missense variants, showing improved performance over existing methods in identifying damaging variants and separating de novo variants in neurodevelopmental disorders.

CNV-espresso improves precision of rare CNV detection in exome sequencing by using transfer learning, while maintaining recall, especially for small exon-spanning CNVs.

VBASS integrates single-cell expression and de novo variant data to enhance power in discovering disease risk genes, showing better performance than existing methods on simulated data and identifying more candidate risk genes.

The paper identifies novel risk genes associated with EA/TEF through whole genome sequencing and functional validation using CRISPR-Cas9 in Xenopus.

The paper identifies LONP1 and ALYREF as new candidate risk genes for congenital diaphragmatic hernia through de novo and ultra-rare variant analyses in proband-parent trios and population controls.

The paper introduces POLARIS, a deep-learning method for predicting localized RBP binding affinity to transcripts, achieving high performance across cell lines and enabling interpretation of noncoding variants.

This study links KIF1A variants to clinical severity in KAND by analyzing transport defects and protein dysfunction.

ThreaderAI improves protein structure prediction by enhancing template-query alignment using deep learning, outperforming existing methods on distant homology proteins.

The paper uses deep whole genome sequencing to analyze genetic variants in CDH patients and their families, revealing heterogeneous and multifactorial genetic etiology.

The paper uses single-cell RNA-seq data to develop a machine learning method called A-risk to prioritize autism risk genes, identifying 71 candidates including new ones like CLCN4, PRKAR1B, and NR2F1.