Yufeng Shen
Research Interests
Make your next move.
Prepare your application
Review drafts before you send them.
Explore related searches
Articles (13)
Congenital diaphragmatic hernia (CDH) is a severe congenital anomaly characterized by impairment of both diaphragm and lung development in utero. CDH presents as a spectrum of forms and severities, with diaphragm defects arising in the dorsal/posterior region typically correlating with more severe pulmonary disease and higher risk of mortality than those appearing in ventral/anterior regions. The genetic etiology underlying CDH is complex, with many genes implicated showing variable expressivity and incomplete penetrance in both human patients and mouse models. Here we present in vivo validation of two genes previously unassociated with CDH: the CDC42-interacting kinase CDC42BPB; and CNOT1, a scaffolding protein of the CCR4-NOT protein complex, critical for mRNA regulation through modifications such as deadenylation. Each gene was found to have a damaging, de novo missense variant in a recent large-scale CDH patient sequencing screen. Loss of Cdc42bpb leads to ventral diaphragmatic hernias, heart septal defects and minor lung epithelial differentiation defects in mouse embryos. Installation of the orthologous patient-specific missense variant through CRISPR/Cas9 editing leads to less severe ventral diaphragm defects. Mouse embryos with either one or two copies of the orthologous Cnot1 variant, c.1867C>T (p.R623W), develop dorsal diaphragmatic hernias with low (<50%) penetrance, and mutants showed alterations in mRNA isoform expression consistent with the molecular role of Cnot1 in RNA splicing. These results underscore the power of in vivo functional modeling to validate genes and patient-specific variants uncovered by patient sequencing, reveal two previously unrecognized genetic causes of CDH, and highlight the heterogeneity of different patient anatomic presentations.
Year:
2026
Protein language models (pLMs) learn sequence patterns at evolutionary scale, but these patterns remain inaccessible within these “black box” models. To discover them, we developed MotifAE, an unsupervised framework based on the sparse autoencoder (SAE) architecture that projects pLM embeddings into an interpretable, sparse latent space. MotifAE introduces an additional smoothness loss to encourage coherent feature activation, which markedly improves the identification of known functional motifs compared to the standard SAE. The sequence patterns captured by MotifAE exhibit rich diversity, align with known functional motifs, and are reflected in the model’s weight space. Beyond short motifs, MotifAE also captures structural domains, with latent feature activation scores correlating with residue importance for different domain functions. By aligning MotifAE features with experimental data, we further identified features associated with domain folding stability. These features enable the prediction of a stability-specific fitness landscape that improves stability prediction and supports the engineering of domains with enhanced stability. Overall, MotifAE provides a general framework for systematic sequence pattern discovery and interpretation, with the potential to advance protein function analysis, mutation effect interpretation, and rational protein engineering.
Year:
2025
Predictive models for missense variant pathogenicity offer little functional interpretation for intrinsically disordered regions, since they rely on conservation and coevolution across homologous sequences. To understand the extent to which biophysics modulates model performance compared to genomic conservation, we model biophysics of IDRs explicitly for improved interpretation of variant effects. We develop MDmis, a method that uses biophysical features extracted from molecular dynamics (MD) simulations of IDRs to predict pathogenicity. We find that pathogenic variants in Long IDRs manifest differently, with transient order and depleted solvent access, compared to those in Short IDRs. Using MD simulations of sequences with single missense variants, we identify stronger evidence for pathogenic effects in Long IDRs compared to Short IDRs. MDmis, when combined with conservation information, achieves strong predictive accuracy of pathogenicity of variants in Long IDRs. Overall, extracting information from MD simulations can help understand the drivers of predictive performance and elucidate biophysical behaviors affected by pathogenic variants.
Year:
2025
The binding of peptide-MHC complexes by T cell receptors (TCRs) is crucial for T cell antigen recognition in adaptive immunity. High-throughput multiplex assays have generated valuable data and insights about antigen specificity of TCRs. However, identifying which TCRs recognize which antigens remains a significant challenge due to the immense diversity of TCR. Here we describe G2VTCR (Graph2Vec-based Representation and Embedding of TCR and Targets for Enhanced Recognition Analysis), a computational method that uses atomic level graph embedding to predict TCR-antigen recognition. G2VTCR represents antigens and the third complementarity-determining region (CDR3) of TCR sequences using graphs, in which nodes encode atomic identities and edges encode chemical bonds between atoms, and then uses Weisfeiler-Lehman iterations to produce embeddings. The embeddings can be used for supervised classification tasks in TCR-antigen binding prediction and unsupervised clustering of TCRs. We evaluated G2VTCR using publicly available paired TCR-CDR3/antigen data generated by antigen-stimulation experiments. We show that G2VTCR has better performance in both classification and clustering than other embedding methods including pre-trained protein language models. We investigated the impact of Weisfeiler-Lehman iterations and the sample size of TCR CDR3 on classification performance. Our results highlight the utility of atomic level graphical embedding of immune repertoire sequences for antigen specificity prediction.
Year:
2025
Protein language models, and models that incorporate structure or homologous sequences, estimate sequence likelihoods p(sequence) that reflect the protein fitness landscape and are commonly used in mutation effect prediction and protein design. It is widely believed in deep learning field that larger model performs better across tasks. However, for fitness prediction, language model performance declines beyond a certain size, raising concerns about their scalability. Here, we showed that model size, training dataset, and stochastic elements can bias the predicted p(sequence) away from real fitness. Model performance on fitness prediction depends on how well p(sequence) matches evolutionary patterns in homologs, which is best achieved at a moderate p(sequence) level for most proteins. At extreme predicted wild-type sequence likelihoods, models predict uniformly low or high likelihoods for nearly all mutations, failing to reflect the real fitness landscape. Notably, larger models tend to predict proteins with higher p(sequence) , which may exceed the moderate range and thus reduce performance. Our findings clarify the scaling behavior of protein models on fitness prediction and provide practical guidelines for their application and future development.
Year:
2025
KIF1A is a neuron-specific kinesin motor responsible for intracellular transport along axons. Pathogenic KIF1A mutations cause KIF1A-associated neurological disorders (KAND), a spectrum of severe neurodevelopmental and neurodegenerative conditions. While individual KIF1A mutations have been studied, how different substitutions at the same residue affect motor function and disease progression remains unclear. Here, we systematically examine the molecular and clinical consequences of mutations at three key motor domain residues—R216, R254, and R307—using single-molecule motility assays and genotype-phenotype associations. Our findings reveal that mutations at R216 and R254 produce residue-specific effects, with some substitutions partially retaining function, whereas R307 mutations universally abolish motility, likely due to its critical role in microtubule binding. Mutant motor properties correlate with developmental outcomes, with R307 mutations linked to the most sever impairments. These results demonstrate that even single-residue substitutions can lead to distinct molecular and clinical phenotypes, highlighting the finely tuned mechanochemical properties of KIF1A. By establishing residue-specific genotype-phenotype relationships, this work provides fundamental insights into KAND pathogenesis and informs targeted therapeutic strategies.
Year:
2025
The mesothelium is a squamous monolayer that ensheathes internal organs, lines the body cavity, and the diaphragm. It serves as a protective barrier, coated in glycocalyx, and secretes lubricants to facilitate tissue movement. How the mesothelium forms is poorly understood. Here, we investigate Myrf , a transcription factor gene expressed in the mesothelium, because it carries variants in patients with Congenital Diaphragmatic Hernia (CDH), a disorder that affects the diaphragm, lung, and other organs. In mice, inactivation of Myrf early in organogenesis resulted in CDH and defective mesothelial specification, compromising its function as a signaling center for lung growth. Inactivation of Myrf later led to enhanced mesothelium differentiation into mesenchymal cell types through partial epithelial-to-mesenchymal transition (EMT), resulting in a unique accumulation of smooth muscle encasing the lung. In this role, MYRF functions in parallel with YAP/TAZ. Together, these findings establish MYRF as a critical regulator of mesothelium development, and when mutated, causes CDH.
Year:
2025
Structural dynamics are fundamental to protein functions and mutation effects. Current protein deep learning models are predominantly trained on sequence and/or static structure data, which often fail to capture the dynamic nature of proteins. To address this, we introduce SeqDance and ESMDance, two protein language models trained on dynamic biophysical properties derived from molecular dynamics simulations and normal mode analyses of over 64,000 proteins. Both models can be directly applied to predict dynamic properties of unseen ordered and disordered proteins. SeqDance, trained from scratch, has attentions that capture dynamic interaction and co-movement between residues, and its embeddings encode rich representations of protein dynamics that can be further utilized to predict conformational properties beyond the training tasks via transfer learning. SeqDance predicted dynamic property changes reflect mutation effect on protein folding stability. ESMDance, built upon ESM2 (Evolutionary Scale Model II) outputs, substantially outperforms ESM2 in zero-shot prediction of mutation effects for designed and viral proteins which lack evolutionary information. Together, SeqDance and ESMDance offer a new framework for integrating protein dynamics into language models, enabling more generalizable predictions of protein behavior and mutation effects. Significance Statement The sequence—structure (ensemble)—function relationship is central to biology. Protein dynamics in the structure ensemble play a decisive role in determining function and mutation effects, and are widely used to study thermodynamics, folding pathways, and dynamic interactions of ordered proteins, as well as the conformational variability of intrinsically disordered proteins. However, current state-of-the-art protein deep learning models, such as AlphaFold2,3 and ESM, focus on static structures and sequences, which failed to directly capture protein dynamics. Here, we address this gap by developing protein language models to learn dynamic properties of over 64,000 proteins. We show that the model’s Transformer attentions capture protein dynamic interactions, and our model can be applied to predict conformational properties and mutation effects.
Year:
2024
Alternative splicing plays a crucial role in protein diversity and gene expression regulation in higher eukaryotes, and mutations causing dysregulated splicing underlie a range of genetic diseases. Computational prediction of alternative splicing from genomic sequences not only provides insight into gene-regulatory mechanisms but also helps identify disease-causing mutations and drug targets. However, the current methods for the quantitative prediction of splice site usage still have limited accuracy. Here, we present DeltaSplice, a deep neural network model optimized to learn the impact of mutations on quantitative changes in alternative splicing from the comparative analysis of homologous genes. The model architecture enables DeltaSplice to perform “reference-informed prediction” by incorporating the known splice site usage of a reference gene sequence to improve its prediction on splicing-altering mutations. We benchmarked DeltaSplice and several other state-of-the-art methods on various prediction tasks, including evolutionary sequence divergence on lineage-specific splicing and splicing-altering mutations in human populations and neurodevelopmental disorders, and demonstrated that DeltaSplice outperformed consistently. DeltaSplice predicted ∼15% of splicing quantitative trait loci (sQTLs) in the human brain as causal splicing-altering variants. It also predicted splicing-altering de novo mutations outside the splice sites in a subset of patients affected by autism and other neurodevelopmental disorders (NDDs), including 19 genes with recurrent splicing-altering mutations. Integration of splicing-altering mutations with other types of de novo mutation burdens allowed the prediction of eight novel NDD-risk genes. Our work expanded the capacity of in silico splicing models with potential applications in genetic diagnosis and the development of splicing-based precision medicine.
Year:
2024
Databases with mental and behavioral health surveys suffer from missingness when participants skip the entire survey, affecting the data quality and sample size. We investigated the missing data patterns and evaluate the imputation performance in Simons Powering Autism Research (SPARK), a large-scale autism cohort consists of over 117,000 participants. Four common methods were assessed – Multiple Imputation by Chained Equations (MICE), K-Nearest Neighbors (KNN), MissForest, and Multiple Imputation with Denoising Autoencoders (MIDAS). In a complete subset of 15,196 autism participants, we simulated three types of missingness patterns. We observed that MIDAS and KNN performed the best as the rate of random missingness increased and when blockwise missingness was simulated. The average computational times for MIDAS and KNN were 10 minutes, 35 minutes for MissForest, and 290 minutes for MICE. MIDAS and KNN both provide promising imputation performance in mental and behavioral health survey data that exhibit blockwise missingness patterns.
Year:
2024
Year:
2024
Year:
2024
Year:
2022
Collaborators (12)
Xin Sun
Professor
University of California, San Diego
David J. McCulley
Associate Professor
UC San Diego Health System
Arne Gennerich
Professor
Albert Einstein College of Medicine
Jacqueline Montes
Associate Professor
Columbia University Irving Medical Center
Samuel Rowbotham
Assistant Professor
Massachusetts General Hospital
Chang Shu
Assistant Professor
University of Southern California
Tristan T. Sands
-
Wendy Chung
Kennedy Family Professor
Boston Children's Hospital
Carol Bult
Professor and Knowlton Family Chair
Jackson Laboratory
Lu Rao
Research Assistant Professor
Albert Einstein College of Medicine
Jennifer M. Bain
Columbia University Irving Medical Center
Frank A. Provenzano
Assistant Professor
Columbia University
Make your next move.
Prepare your application
Review drafts before you send them.