Biology subjects

Song, X.

Publications and source records attributed to Song, X..

At least 109 records · Page 6Linked to original sources

bioRxiv2024

Voice patches in the marmoset auditory cortex revealed by wide-field calcium imaging

Species-specific vocalizations are behaviorally critical sounds. Similar to faces, species-specific vocalizations are important for the survival and social interactions of both humans and vocal animals. Face patches have been found in the brains of both human and non-human primates. In humans, a voice patch system has been identified on the lateral superior temporal gurus (STG) that is selective to human voices over other sounds. In non-human primates, while vocalization-selective regions were found on the rostral portion of the temporal lobe outside of the auditory cortex in both macaques and marmosets using functional magnetic resonance imaging (fMRI), it is yet clear whether vocalization-selective regions are present in the auditory cortex. Using wide-field calcium imaging, a technique with both high temporal and high spatial resolution, we discovered two voice patches in the marmoset auditory cortex that preferentially respond to marmoset vocalizations over other sounds and carry call types and identity information. One patch is located on the posterior primary auditory cortex (A1), and the other one is located on the anterior non-core region of the auditory cortex. These voice patches are functionally connected and hierarchically organized as shown by latency and selectivity analyses. Our findings reveal the existence of voice patches in the auditory cortex of marmosets and support the notion that similar cortical architectures are adapted for recognizing communication signals for both vocalizations and faces in different primate species.

neuroscience↗

bioRxiv2024

Intratumoral Microbiome of Adenoid Cystic Carcinomas and Comparison with other Head and Neck Cancers

BackgroundAdenoid cystic carcinoma (ACC) is a rare, slow growing yet aggressive head and neck malignancy. Despite its clinical significance, our understanding of the cellular evolution and microenvironment in ACC remains limited. MethodsWe investigated the intratumoral microbiome of 50 ACC tumors and 33 adjacent normal tissues using 16S rRNA gene sequencing. This allowed us to characterize the bacterial communities within ACC and explore potential associations between the bacterial community structure, patients clinical characteristics, and tumor molecular features obtained through RNA sequencing. ResultsBacterial composition in ACC displayed significant differences compared to adjacent normal salivary tissue and exhibited diverse levels of species richness. We identified two main microbial subtypes within ACC: oral-like and gut-like. Oral-like microbiomes, characterized by higher diversity and abundance of genera like Neisseria, Leptotrichia, Actinomyces, Streptococcus, Rothia, and Veillonella (commonly found in healthy oral cavities), were associated with the less aggressive ACC-II molecular subtype and improved patient outcomes. Notably, we identified the same oral genera in oral cancer and in head and neck squamous cell carcinomas. In both cancers, they were part of shared oral communities associated with more diverse microbiome, less aggressive tumor phenotype, and better survival. Conversely, gut-like microbiomes in ACC, featuring low diversity and colonization by gut mucus layer-degrading species like Bacteroides, Akkermansia, Blautia, Bifidobacterium, and Enterococcus, were associated with poorer outcomes. Elevated levels of Bacteroides thetaiotaomicron were independently associated with significantly worse survival, regardless of other clinical and molecular factors. Furthermore, this association positively correlated with tumor cell biosynthesis of glycan-based cell membrane components. ConclusionsOur study uncovers specific intratumoral oral genera as potential pan-cancer biomarkers for favorable microbiomes in ACC and other head and neck cancers. These findings highlight the pivotal role of the intratumoral microbiome in influencing ACC prognosis and disease biology.

microbiology↗

bioRxiv2023

Genome organization in double-stranded DNA viruses observed by cryoET

Double-stranded DNA (dsDNA) viruses package their genetic material into protein cages with diameters usually a few hundred times smaller than the length of their genome. Compressing the relatively stiff and highly negatively charged dsDNA into a small volume is energetically costly and mechanistically enigmatic. Multiple models of dsDNA packaging have been proposed based on various experimental evidence and simulation methods, but direct observation of any viral genome organization is lacking. Here, using cryoET and an improved data processing scheme that utilizes information from the encaging protein shell, we present 3D views of dsDNA genome inside individual viral particles at resolution that densities of neighboring DNA duplexes are readily separable. These cryoET observations reveal a "rod-and-coil" fold of the dsDNA that is conserved among herpes simplex virus type 1 (HSV-1) with a spherical capsid, bacteriophage T4 with a prolate capsid, and bacteriophage T7 with a proteinaceous core inside the capsid. Finally, inspired by the genome arrangement in partially packaged T4 particles, we propose a mechanism for the genome packaging process in dsDNA viruses.

molecular biology↗

bioRxiv2023

QuadST: A Powerful and Robust Approach for Identifying Cell-Cell Interaction-Changed Genes on Spatially Resolved Transcriptomics

Spatially resolved transcriptomics (SRT) have enabled profiling spatial organization of cells and their transcriptome in situ. Various analytical methods have been developed to uncover cell-cell interaction processes using SRT data. To improve upon existing efforts, we developed a novel statistical framework called QuadST for the robust and powerful identification of interaction-changed genes (ICGs) for cell-type-pair specific interactions on a single-cell SRT dataset. QuadST is motivated by the idea that in the presence of cell-cell interaction, gene expression level can vary with cell-cell distance between cell type pairs, which can be particularly pronounced within and in the vicinity of cell-cell interaction distance. Specifically, QuadST infers ICGs in a specific cell type pairs interaction based on a quantile regression model, which allows us to assess the strength of distance-expression association across entire distance quantiles conditioned on gene expression level. To identify ICGs, QuadST performs a hypothesis testing with an empirically estimated FDR, whose upper bound is determined by the ratio of cumulative associations at symmetrically smaller and larger distance quantiles simultaneously across all genes. Simulation studies illustrate that QuadST provides consistent FDR control and better power performance than other compared methods. Its application on SRT datasets profiled from mouse brains demonstrates that QuadST can identify ICGs presumed to play a role in specific cell type pair interactions (e.g., synaptic pathway genes among excitatory neuron cell interactions). These results suggest that QuadST can be a useful tool to discover genes and regulatory processes involved in specific cell type pair interactions.

bioinformatics↗

bioRxiv2023

Spatially resolved molecular and cellular atlas of the mouse brain

A comprehensive atlas of genes, cell types, and their spatial distribution across a whole mammalian brain is fundamental for understanding function of the brain. Here, using snRNA-seq and Stereo-seq techniques, we generated a mouse brain atlas with spatial information for 308 cell clusters with single-cell resolution involving over 6 million cells as well as for 29,655 genes. We have identified new astrocyte clusters, and demonstrated that distinct cell clusters exhibit preference for cortical subregions. In addition, we identified 155 genes exhibiting regional specificity in the brainstem, and 513 long non-coding RNA exhibited regional specificity in the adult brain. Parcellation of brain regions based on spatial transcriptomic information showed large overlap with that by traditional method. Furthermore, we have uncovered 411 transcription factor regulons with spatiotemporal specificity during development. Thus, our study has discovered genes and regulon with spatiotemporal specificity, and provided a high-resolution spatial transcriptomic atlas of the mouse brain.

neuroscience↗

bioRxiv2023

Bi-parental graph strategy to represent and analyze hybrid plant genomes

Hybrid plants are universally existed in wild and often exhibit greater performance of complex traits compared with their parents and other selfing plants. This phenomenon, known as heterosis, has been extensively applied in plant breeding for decades. However, the process of decoding hybrid plant genomes has seriously lagged due to the challenges in their genome assembling and the lack of proper methods to further represent and analyze them. Here we report the assembly and analysis of two hybrids: an intraspecific hybrid between two maize inbred lines and an interspecific hybrid between maize and its wild relative teosinte, based on the combination of PacBio High Fidelity (HiFi) sequencing and chromatin conformation capture sequencing data. The haplotypic assemblies are well-phased at chromosomal scale, successfully resolving the complex loci with extensive parental structural variations (SVs). By integrating into a bi-parental genome graph, the haplotypic assemblies can facilitate downstream short-reads based SV calling and allele-specific gene expression analysis, demonstrating outstanding advantages over one single linear genome. Our work provides an entire workflow which hopefully can promote the deciphering of the large numbers of hybrid plant genomes, especially those whose parents are unknown or unavailable and help to understand genome evolution and heterosis.

bioinformatics↗

bioRxiv2023

Towards Generalizability and Robustness in Biological Object Detection in Electron Microscopy Images

Machine learning approaches have the potential for meaningful impact in the biomedical field. However, there are often challenges unique to biomedical data that prohibits the adoption of these innovations. For example, limited data, data volatility, and data shifts all compromise model robustness and generalizability. Without proper tuning and data management, deploying machine learning models in the presence of unaccounted for corruptions leads to reduced or misleading performance. This study explores techniques to enhance model generalizability through iterative adjustments. Specifically, we investigate a detection tasks using electron microscopy images and compare models trained with different normalization and augmentation techniques. We found that models trained with Group Normalization or texture data augmentation outperform other normalization techniques and classical data augmentation, enabling them to learn more generalized features. These improvements persist even when models are trained and tested on disjoint datasets acquired through diverse data acquisition protocols. Results hold true for transformerand convolution-based detection architectures. The experiments show an impressive 29% boost in average precision, indicating significant enhancements in the models generalizibality. This underscores the models capacity to effectively adapt to diverse datasets and demonstrates their increased resilience in real-world applications.

bioengineering↗

bioRxiv2023

A novel antagonist of the CCL5/CCR5 axis suppresses the tumor growth and metastasis of triple-negative breast cancer by CCR5-YAP1 regulation.

Triple-negative breast cancer (TNBC) is the most aggressive subtype of breast cancer (BC) with a high mortality rate, and few effective therapeutic strategies are available. CCL5/CCR5 is an appealing immunotherapeutic target for TNBC. However, its signaling mechanism is poorly understood and its direct antagonists have not been reported. Here, we developed a high-throughput screening (HTS) assay for discovering its antagonists. Verteporfin was identified as a more selective and potent antagonist than the known CCR5 antagonist maraviroc. Without photodynamic therapy, verteporfin demonstrated significant inhibition on TNBC tumor growth through immune regulation, remarkable suppression of lung metastasis by cell-intrinsic mechanism, and a significant extension of overall survival in vivo. Mechanistically, CCR5 was found to be essential for expression of the key hippo effector YAP1. It promoted YAP1 transcription via HIF-1 and exerted further control over the migration of CD8+ T, NK, and MDSC immune cells through chemokines CXCL16 and CXCL8 which were identified from RNA-seq. Moreover, the CCR5-YAP1 axis played a vital role in promoting metastasis by modulating {beta}-catenin and core epithelial-mesenchymal transition transcription factors ZEB1 and ZEB2. It is noteworthy that the regulatory relationship between CCR5 and YAP1 was observed across various BC subtypes, TNBC patients, and showed potential relevance in fifteen additional cancer types. Overall, this study introduced an easy-to-use HTS assay that streamlines the discovery of CCL5/CCR5 axis antagonists. Verteporfin was identified as a specific molecular probe of this axis with great potentials as a therapeutic agent for treating sixteen malignant diseases characterized by heightened CCR5 and YAP1 levels.

cancer biology↗

bioRxiv2023

Chemical Tools for the Gid4 Subunit of the Human E3 Ligase C-terminal to LisH (CTLH) Degradation Complex

We have developed a novel chemical handle (PFI-E3H1) and a chemical probe (PFI-7) as ligands for the Gid4 subunit of the human E3 ligase CTLH degradation complex. Through an efficient initial hit-ID campaign, structure-based drug design (SBDD) and leveraging the sizeable Pfizer compound library, we identified a 500 nM ligand for this E3 ligase through file screening alone. Further exploration identified a vector that is tolerant to addition of a linker for future chimeric molecule design. The chemotype was subsequently optimized to sub-100 nM Gid4 binding affinity for a chemical probe. These novel tools, alongside the suitable negative control also identified, should enable the interrogation of this complex human E3 ligase macromolecular assembly.

biophysics↗

bioRxiv2023

Efficient semi-supervised semantic segmentation of electron microscopy cancer images with sparse annotations

Electron microscopy (EM) enables imaging at nanometer resolution and can shed light on how cancer evolves to develop resistance to therapy. Acquiring these images has become a routine task; however, analyzing them is now the bottleneck, as manual structure identification is very time-consuming and can take up to several months for a single sample. Deep learning approaches offer a suitable solution to speed up the analysis. In this work, we present a study of several state-of-the-art deep learning models for the task of segmenting nuclei and nucleoli in volumes from tumor biopsies. We compared previous results obtained with the ResUNet architecture to the more recent UNet++, FracTALResNet, SenFormer, and CEECNet models. In addition, we explored the utilization of unlabeled images through semi-supervised learning with Cross Pseudo Supervision. We have trained and evaluated all of the models on sparse manual labels from three fully annotated in-house datasets that we have made available on demand, demonstrating improvements in terms of 3D Dice score. From the analysis of these results, we drew conclusions on the relative gains of using more complex models, semi-supervised learning as well as next steps for the mitigation of the manual segmentation bottleneck.

bioengineering↗

bioRxiv2023

A novel Bayesian model for assessing intratumor heterogeneity of tumor infiltrating leukocytes with multi-region gene expression sequencing

Intratumor heterogeneity (ITH) of tumor-infiltrated leukocytes (TILs) is an important phenomenon of cancer biology with potentially profound clinical impacts. Multiregion gene expression sequencing data provide a promising opportunity that allows for explorations of TILs and their intratumor heterogeneity for each subject. Although several existing methods are available to infer the proportions of TILs, considerable methodological gaps exist for evaluating intratumor heterogeneity of TILs with multi-region gene expression data. Here, we develop ICeITH, immune cell estimation reveals intratumor heterogeneity, a Bayesian hierarchical model that borrows cell type profiles as prior knowledge to decompose mixed bulk data while accounting for the within-subject correlations among tumor samples. ICeITH quantifies intratumor heterogeneity by the variability of targeted cellular compositions. Through extensive simulation studies, we demonstrate that ICeITH is more accurate in measuring relative cellular abundance and evaluating intratumor heterogeneity compared with existing methods. We also assess the ability of ICeITH to stratify patients by their intratumor heterogeneity score and associate the estimations with the survival outcomes. Finally, we apply ICeITH to two multi-region gene expression datasets from lung cancer studies to classify patients into different risk groups according to the ITH estimations of targeted TILs that shape either pro- or anti-tumor processes. In conclusion, ICeITH is a useful tool to evaluate intratumor heterogeneity of TILs from multi-region gene expression data.

bioinformatics↗

bioRxiv2023

Numerical simulation and analysis of droplet formation within an amphiphilic particle

An instrument-free particle-templated droplet formation can be achieved upon simple mixing of amphiphilic particles with aqueous and oil phases in a well plate by using a common lab pipette. Here, a two-dimensional, two-phase flow model was established using a finite element method to mimic the droplet formation within a concentric amphiphilic particle, which consisted of an outer hydrophobic layer and an inner hydrophilic layer. Immiscible water and oil phases selectively interacted with the hydrophilic and hydrophobic layers of the particle, respectively, to form an isolated aqueous compartment within a cavity. Three extreme models were also simulated, including completely hydrophilic, completely hydrophobic, and oppositely amphiphilic particle, which indicated that a right order of the particle layers was necessary to capture the droplet inside the cavity. Moreover, we performed a systematic study of particle-templated droplet formation by varying the individual layer thicknesses of particle, particle height, interfacial tension between water and oil, contact angle of interface with different surfaces, velocity of incoming oil media, and distance between neighboring particles. The volume fraction of water droplet trapped within the target cavity region was calculated to characterize the droplet formation. Our work will help to optimize the particle fabrication process, predict the experiment droplet formation, and explain the physical mechanism underlying compartmentalization phenomena.

bioengineering↗

bioRxiv2023

Selective refueling of CAR T cells using ADA1 and 1 CD26 boosts antitumor immunity

CAR T cell therapy has revolutionized the treatment of hematologic malignancies, but its efficacy in solid tumors remains limited due to the immunosuppressive nature of the tumor microenvironment and the inability of T cells to persist and traffic to the tumor site. While current strategies focus on enhancing CAR T cell activity through costimulatory molecules and cytokines, a critical yet often overlooked factor is the competition for nutrients between tumor cells and T cells in the nutrient-deprived tumor microenvironment. To address this challenge, we employed a selective metabolic refueling (MR) strategy by providing T cells with inosine as an alternative fuel source for growth and functionality. In this study, we engineered CAR T cells to co-express a membrane-bound CD26 and a cytoplasmic adenosine deaminase 1 (ADA1) fused to an anti-CD3 scFv. ADA1 irreversibly converts both intracellular and extracellular adenosine to inosine, overcoming adenosine-mediated immunosuppression and providing T cells with inosine for growth. The inclusion of an anti-CD3 scFv fusion partner and overexpressing CD26 boosts ADA1 capture in a membrane proximal manner, providing inosine for T cells and minimizing feeding the tumor cells. We demonstrate that ADA1 is conditionally secreted only in stress conditions and that it activates CAR T cells through trans-signaling in a tumor-specific manner. In addition, we show that, compared to unmodified CAR T cells, CD26-overexpressing CAR T cells have better migration capacity and are less susceptible to TGF-{beta}1 suppression. Finally, we found that, in mice models of human hepatocellular carcinoma (GPC3-MR-CAR) and human non-small cell lung cancer (HER2-MR-CAR), metabolically refueled CAR T cells (MR-CAR) are more efficient in reducing tumor growth than unmodified CAR T cells. Thus, selective refueling CAR T cells using ADA1 and CD26 holds promise for improving the efficacy of CAR T cell therapy of solid tumors.

immunology↗

bioRxiv2023

Imprinting but not cytonuclear interactions affects parent-of-origin effect on seed size in Arabidopsis hybrids

The parent-of-origin effect on seed size can result from imprinting or a combinational effect between cytoplasmic and nuclear genomes, but their relative contributions remain unknown. To discern these confounding effects, we generated cytoplasmic-nuclear substitution (CNS) lines using recurrent backcrossing in the Arabidopsis thaliana ecotypes Col-0 and C24. These CNS lines differ only in the nuclear genome (imprinting) or in the cytoplasm. The CNS reciprocal hybrids with the same cytoplasm display a [~]20% seed size difference as observed in the conventional hybrids. However, seed size is similar between the reciprocal cybrids with fixed imprinting. Transcriptome analyses in the endosperm of CNS hybrids using laser-capture microdissection have identified 104 maternally expressed genes (MEGs) and 90 paternally-expressed genes (PEGs). These imprinted genes are involved in pectin catabolism and cell wall modification in the endosperm. HDG9, an epiallele and one of 11 cross-specific imprinted genes, controls seed size. In the embryo, a handful of imprinted genes is found in the CNS hybrids but only one is expressed higher in the embryo than endosperm. AT4G13495 encodes a long-noncoding RNA (lncRNA), but no obvious seed phenotype is observed in the lncRNA knockout lines. NRPD1, encoding the largest subunit of RNA Pol IV, is involved in the biogenesis of small interfering RNAs. Seed size and embryo is larger in the cross using nrpd1 as the maternal parent than in the reciprocal cross. In spite of limited ecotypes tested, these results suggest potential roles of imprinting and NRPD1-mediated small RNA pathway in seed size variation in hybrids.

genetics↗

bioRxiv2023

Bacterial Sphingolipids Exacerbate Colitis by Inhibiting ILC3-derived IL-22 Production.

Commensal bacteria of the Bacteroidetes phylum are the primary producers of sphingolipids in the gut lumen. These lipids serve dual roles as bacterial virulence factors and regulators of the host mucosal immune system, including regulatory T cells and invariant natural killer T cells (iNKT). Sphingolipid composition is significantly altered in fecal samples of patients with inflammatory bowel disease (IBD). However, the specific mechanisms by which bacterial sphingolipids modulate mucosal homeostasis and regulate intestinal inflammation remain unclear. In this study, we investigated the impact of bacterial sphingolipids on intestinal inflammation by mono-colonizing mice with Bacteroides fragilis strains that either express or lack sphingolipids during DSS-induced colitis. We discovered that B. fragilis sphingolipids exacerbate intestinal inflammation. Mice mono-colonized with B. fragilis lacking sphingolipids exhibited less severe DSS-induced colitis. This amelioration of colitis was associated with increased production of interleukin-22 (IL-22) by innate lymphoid cell type 3 (ILC3). Consistent with the inhibitory effect of sphingolipids on IL-22 production, mice colonized with B. fragilis lacking sphingolipids showed enhanced epithelial STAT3 activity, intestinal cell proliferation, and antimicrobial peptide production following DSS treatment compared to those colonized with B. fragilis producing sphingolipids. Additionally, colitis severity in mice colonized with B. fragilis lacking sphingolipids was exacerbated upon IL-22 blockade. Furthermore, our study reveals that bacterial sphingolipids restrict epithelial IL-18 production following DSS treatment and interfere with IL-22 production by a subset of ILC3 cells expressing both the interleukin-18 receptor (IL-18R) and major histocompatibility complex class II (MHC II). These findings indicate that B. fragilis-derived sphingolipids exacerbate mucosal inflammation by impeding epithelial IL-18 expression, resulting in compromised production of IL-22 by ILC3 cells. HighlightsO_LIB. fragilis-derived sphingolipids exacerbate DSS-induced colitis in mono-colonized C57BL/6 mice. C_LIO_LIB. fragilis-derived sphingolipids constrain ILC3-derived IL-22, leading to reduced colonic epithelial cell proliferation and compromised barrier function. C_LIO_LIB. fragilis-derived sphingolipids restrict epithelial NLRC4 inflammasome activation and IL-18 secretion. C_LIO_LIB. fragilis-derived sphingolipids modulate IL-22 production by IL18R+ MHC II+ ILC3s. C_LI

immunology↗

bioRxiv2023

Anti-PD-1-iRGD Peptide Conjugate Boosts Antitumor Efficacy via Engagement Augmentation and Penetration Enhancement of T cells

Despite the important breakthroughs of immune-checkpoint inhibitors (ICIs) in recent years, the overall objective response rate (ORR) remains limited in various cancers. Here, we synthesized programmed cell death protein-1 (PD-1) antibody iRGD conjugate (PD-1-(iRGD)2) through glycoengineering and bio-orthogonal reaction. PD-1-(iRGD)2 exhibited extra iRGD receptor dependent affinity to several cancer cell lines rather than normal cell lines. Via dual targeting, PD-1-(iRGD)2 engageed tumor cells and T cells thus mediating T cell activation and facilitating tumor elimination. Besides, the attachment of iRGD impressively improved the penetrability of both PD-1 antibody and PD-1+ T cells. In multiple syngeneic mouse models, PD-1-(iRGD)2 effectively reduced tumor growth with satisfactory biosafety. Moreover, results of flow cytometry and single-cell RNA-seq revealed that PD-1-(iRGD)2 remodeled the tumor microenvironment (TME) and expanded a unique population of "better effector" CD8+ tumor infiltrating T cells (TILs) expressing stem and memory associated genes including Tcf7, Il7r, Lef1 and Bach2. Conclusively, PD-1-(iRGD)2 could be a novel and promising therapeutic approach for cancer immunotherapy. Statement of significanceDesigned against the clinical dilemma of unsatisfied response rate after contemporary cancer immunotherapy, PD-1-(iRGD)2 engages T cells and tumor cells, promotes T cell infiltration and expands a unique population of "better effectors" with enhanced therapeutic potential for the treatment of cancer.

immunology↗

bioRxiv2023

ASB3 expression aggravates inflammatory bowel disease by targeting TRAF6 protein stability and affecting the intestinal microbiota

E3 ubiquitin ligase (E3) plays a vital role in regulating inflammatory responses by mediating ubiquitination. Previous studies have shown that ankyrin repeat and SOCS box-containing protein 3 (ASB3) is involved in immunomodulatory functions associated with cancer. However, the impact of ASB3 on the dynamic interplay of microbiota and inflammatory responses in inflammatory bowel disease (IBD) is unclear. Here we systematically identify the E3 ligase ASB3 as a facilitative regulator in the development and progression of IBD. ASB3-/- mice are resistant to DSS-induced colitis. I{kappa}B phosphorylation levels and production of proinflammatory factors IL-1{beta}, IL-6, and TNF- were reduced in colonic tissues of ASB3-/- mice compared to WT mice. This colitis-resistant phenotype was suppressed after coprophagic microbial transfer and reversed after combined antibiotic clearance of the microbiota. Mechanistically, ASB3 specifically catalyzes K48-linked polyubiquitination of TRAF6 in IECs. In contrast, in ASB3-deficient organoids, the integrity of the TRAF6 protein is shielded, consequently decelerating the onset of intestinal inflammation. Taken together, Our findings demonstrates that ASB3 is associated with dysregulation of the colitis microbiota and promotes proinflammatory factors production by disrupting the TRAF6 stability. Strategies to limit the protein level of ASB3 in intestinal epithelial cells may help in the treatment of colitis.

immunology↗

bioRxiv2023

DrugGPT: A GPT-based Strategy for Designing Potential Ligands Targeting Specific Proteins

DrugGPT presents a ligand design strategy based on the autoregressive model, GPT, focusing on chemical space exploration and the discovery of ligands for specific proteins. Deep learning language models have shown significant potential in various domains including protein design and biomedical text analysis, providing strong support for the proposition of DrugGPT. In this study, we employ the DrugGPT model to learn a substantial amount of protein-ligand binding data, aiming to discover novel molecules that can bind with specific proteins. This strategy not only significantly improves the efficiency of ligand design but also offers a swift and effective avenue for the drug development process, bringing new possibilities to the pharmaceutical domain. In our research, we particularly optimized and trained the GPT-2 model to better adapt to the requirements of drug design. Given the characteristics of proteins and ligands, we redesigned the tokenizer using the BPE algorithm, abandoned the original tokenizer, and trained the GPT-2 model from scratch. This improvement enables DrugGPT to more accurately capture and understand the structural information and chemical rules of drug molecules. It also enhances its comprehension of binding information between proteins and ligands, thereby generating potentially active drug candidate molecules. Theoretically, DrugGPT has significant advantages. During the model training process, DrugGPT aims to maximize the conditional probability and employs the back-propagation algorithm for training, making the training process more stable and avoiding the Mode Collapse problem that may occur in Generative Adversarial Networks in drug design. Furthermore, the design philosophy of DrugGPT endows it with strong generalization capabilities, giving it the potential to adapt to different tasks. In conclusion, DrugGPT provides a forward-thinking and practical new approach to ligand design. By optimizing the tokenizer and retraining the GPT-2 model, the ligand design process becomes more direct and efficient. This not only reflects the theoretical advantages of DrugGPT but also reveals its potential applications in the drug development process, thereby opening new perspectives and possibilities in the pharmaceutical field.

bioinformatics↗