Fanglei Xue

Native homo-oligomeric transmembrane β-barrels (TMBs) have been widely explored for molecular sensing, sequencing, and separation applications, but their uniform lumen limits spatial resolution and the localized interactions required for both analyte discrimination and filtration. Monomeric TMBs have been designed using energy-based methods but this has required extensive expert input, limiting scalability and functionality. Here, we present a rapid generative AI method for TMB design that combines diffusion-based backbone generation conditioned on β-barrel structural features with TMB-optimized sequence design. We characterized 48 designs (spanning 10-16 strands) that exhibited measurable conductances pertaining to pore diameters of 0.7-1.5 nm, and determined crystal structures of two designs that have atomic-level agreement with the designed models. We demonstrate the versatility of the generative method by designing nanopores with copper binding sites for selective ion sensing, larger pores that support DNA translocation, and longer pores with increased hydrophobic thickness, that mediate ion transport across three-dimensional networks of monoglyceride bilayers and synthetic polymer-lipid hybrid membranes composed of block copolymers.

Mutations in isocitrate dehydrogenase 1 (IDH1) and 2 (IDH2) are common in multiple types of human cancer and cause accumulation of the oncometabolite D -2-hydroxyglutarate (D2HG) instead of α-ketoglutarate, driving cancers like gliomas and acute myeloid leukaemia by blocking cell differentiation and promoting tumour growth. Here we discovered protein O -2-hydroxyglutarylation by D2HG using chemical proteomics and further revealed distinct chiral preferences for D2HG and L -2-hydroxyglutarate (L2HG) modifications. D2HG modifications are upregulated in IDH-mutant cells or upon D2HG treatment, while L2HG modifications increase under hypoxic conditions or following L2HG treatment. Notably, two kinases MRCKA and SLK are modified by D2HG and L2HG, respectively, and confirmed by synthetic peptide standards. Phosphoproteomics revealed reduced phosphorylation of MRCKA and SLK substrates, suggesting crosstalk between D/L-2HG modification and kinase activity. These findings highlight distinctive roles of D/L-2HG modifications in cancer progression and suggest potential avenues for therapeutic targeting of oncometabolite-induced post-translational modifications.

Targeted protein degradation (TPD) has rapidly emerged as a powerful modality for drugging previously “undruggable” proteins. TPD employs small molecules like PROTACs and molecular glue degraders (MGD) to induce target protein degradation via the formation of a ternary complex with an E3 ligase. However, the rational design of these degraders is severely hindered by the difficulty of obtaining these ternary structures. Here we introduce DeepTernary, a novel end-to-end deep learning approach using an SE(3)-equivariant encoder and a query-based decoder to accurately and rapidly predict these critical structures. Trained on carefully curated TernaryDB, DeepTernary achieves state-of-the-art performance on PROTAC benchmarks without prior exposure to known PROTACs and shows notable prediction capability on the more challenging MGD benchmark with a blind docking protocol. Remarkably, the buried surface areas calculated from predicted structures correlate with experimental degradation potency metrics. Overall, DeepTernary offers a powerful tool for the development of targeted protein degraders.

Facial Expression Recognition (FER) in the wild is an extremely challenging task. Recently, some Vision Transformers (ViT) have been explored for FER, but most of them perform inferiorly compared to Convolutional Neural Networks (CNN). This is mainly because the new proposed modules are difficult to converge well from scratch due to lacking inductive bias and easy to focus on the occlusion and noisy areas. TransFER, a representative transformer-based method for FER, alleviates this with multi-branch attention dropping but brings excessive computations. On the contrary, we present two attentive pooling (AP) modules to pool noisy features directly. The AP modules include Attentive Patch Pooling (APP) and Attentive Token Pooling (ATP). They aim to guide the model to emphasize the most discriminative features while reducing the impacts of less relevant features. The proposed APP is employed to select the most informative patches on CNN features, and ATP discards unimportant tokens in ViT. Being simple to implement and without learnable parameters, the APP and ATP intuitively reduce the computational cost while boosting the performance by ONLY pursuing the most discriminative features. Qualitative results demonstrate the motivations and effectiveness of our attentive poolings. Besides, quantitative results on six in-the-wild datasets outperform other state-of-the-art methods.