Generative design of compounds with desired potency from target protein sequences using a multimodal biochemical language model.

Deep learning models adapted from natural language processing offer new opportunities for the prediction of active compounds via machine translation of sequential molecular data representations. For example, chemical language models are often derived for compound string transformation. Moreover, given the principal versatility of language models for translating different types of textual representations, off-the-beaten-path design tasks might be explored. In this work, we have investigated generative design of active compounds with desired potency from target sequence embeddings, representing a rather provoking prediction task. Therefore, a dual-component conditional language model was designed for learning from multimodal data. It comprised a protein language model component for generating target sequence embeddings and a conditional transformer for predicting new active compounds with desired potency. To this end, the designated "biochemical" language model was trained to learn mappings of combined protein sequence and compound potency value embeddings to corresponding compounds, fine-tuned on individual activity classes not encountered during model derivation, and evaluated on compound test sets that were structurally distinct from training sets. The biochemical language model correctly reproduced known compounds with different potency for all activity classes, providing proof-of-concept for the approach. Furthermore, the conditional model consistently reproduced larger numbers of known compounds as well as more potent compounds than an unconditional model, revealing a substantial effect of potency conditioning. The biochemical language model also generated structurally diverse candidate compounds departing from both fine-tuning and test compounds. Overall, generative compound design based on potency value-conditioned target sequence embeddings yielded promising results, rendering the approach attractive for further exploration and practical applications. SCIENTIFIC CONTRIBUTION: The approach introduced herein combines protein language model and chemical language model components, representing an advanced architecture, and is the first methodology for predicting compounds with desired potency from conditioned protein sequence data.

Meta-learning for transformer-based prediction of potent compounds.

For many machine learning applications in drug discovery, only limited amounts of training data are available. This typically applies to compound design and activity prediction and often restricts machine learning, especially deep learning. For low-data applications, specialized learning strategies can be considered to limit required training data. Among these is meta-learning that attempts to enable learning in low-data regimes by combining outputs of different models and utilizing meta-data from these predictions. However, in drug discovery settings, meta-learning is still in its infancy. In this study, we have explored meta-learning for the prediction of potent compounds via generative design using transformer models. For different activity classes, meta-learning models were derived to predict highly potent compounds from weakly potent templates in the presence of varying amounts of fine-tuning data and compared to other transformers developed for this task. Meta-learning consistently led to statistically significant improvements in model performance, in particular, when fine-tuning data were limited. Moreover, meta-learning models generated target compounds with higher potency and larger potency differences between templates and targets than other transformers, indicating their potential for low-data compound design.

Designing highly potent compounds using a chemical language model.

Compound potency prediction is a major task in medicinal chemistry and drug design. Inspired by the concept of activity cliffs (which encode large differences in potency between similar active compounds), we have devised a new methodology for predicting potent compounds from weakly potent input molecules. Therefore, a chemical language model was implemented consisting of a conditional transformer architecture for compound design guided by observed potency differences. The model was evaluated using a newly generated compound test system enabling a rigorous assessment of its performance. It was shown to predict known potent compounds from different activity classes not encountered during training. Moreover, the model was capable of creating highly potent compounds that were structurally distinct from input molecules. It also produced many novel candidate compounds not included in test sets. Taken together, the findings confirmed the ability of the new methodology to generate structurally diverse highly potent compounds.

DeepCubist: Molecular Generator for Designing Peptidomimetics based on Complex three-dimensional scaffolds.

Mimicking bioactive conformations of peptide segments involved in the formation of protein-protein interfaces with small molecules is thought to represent a promising strategy for the design of protein-protein interaction (PPI) inhibitors. For compound design, the use of three-dimensional (3D) scaffolds rich in sp3-centers makes it possible to precisely mimic bioactive peptide conformations. Herein, we introduce DeepCubist, a molecular generator for designing peptidomimetics based on 3D scaffolds. Firstly, enumerated 3D scaffolds are superposed on a target peptide conformation to identify a preferred template structure for designing peptidomimetics. Secondly, heteroatoms and unsaturated bonds are introduced into the template via a deep generative model to produce candidate compounds. DeepCubist was applied to design peptidomimetics of exemplary peptide turn, helix, and loop structures in pharmaceutical targets engaging in PPIs.

Apatinib combined with camrelizumab in the treatment of recurrent/metastatic nasopharyngeal carcinoma: a prospective multicenter phase II study.

Background: Preclinical studies demonstrated that immune checkpoint inhibitors combined with antiangiogenic drugs have a synergistic anti-tumor effect. This present phase II trial aimed to evaluate the efficacy and safety of apatinib combined with camrelizumab in patients with recurrent/metastatic nasopharyngeal carcinoma (RM-NPC). Methods: Patients with RM-NPC were administered with apatinib at 250 mg orally once every day and with camrelizumab at 200 mg via intravenous infusion every 2 weeks until the disease progressed or toxicity became unacceptable. The objective response rate (ORR) was the primary endpoint, assessed using RECIST version 1.1. Progression-free survival (PFS), overall survival (OS), disease control rate (DCR) and safety were the key secondary endpoints. This study was registered with ClinicalTrials.gov, NCT04350190. Results: This study enrolled 26 patients with RM-NPC between January 14, 2021 and September 15, 2021. At data cutoff (March 31, 2023), the median duration of follow-up was 16 months (ranging from 1 to 26 months). The ORR was 38.5% (10/26), the disease control rate (DCR) was 61.5% (16/26), and the median PFS was 6 months (IQR 3.0-20.0). The median OS was 14 months (IQR 6.0-21.25). Treatment-related grade 3 or 4 adverse events occurred in seven (26.9%) patients, and comprised anemia (7.7%), stomatitis (3.8%), headache (3.8%), pneumonia (7.7%), and myocarditis (3.8%). There were no serious treatment-related adverse events or treatment-related deaths. Conclusion: In patients with RM-NPC, apatinib plus camrelizumab showed promising antitumor activity and manageable toxicities.

Prediction model for synergistic anti-tumor multi-compound combinations from traditional Chinese medicine based on extreme gradient boosting, targets and gene expression data.

Traditional Chinese medicine (TCM) is characterized by synergistic therapeutic effect involving multiple compounds and targets, which provide potential new therapy for the treatment of complex cancer conditions. However, the main contributors and the underlying mechanisms of synergistic TCM cancer therapies remain largely undetermined. Machine learning now provides a new approach to determine synergistic compound combinations from complex components of TCM. In this study, a prediction model based on extreme gradient boosting (XGBoost) algorithm was constructed by integrating gene expression data of different cancer cell lines, targets information of natural compounds and drug response data. Radix Paeoniae Rubra (RPR) was selected as a model herbal sample to evaluate the reliability of the constructed model. The optimal XGBoost prediction model achieved a good performance with Mean Square Error (MSE) of 0.66, Mean Absolute Error (MAE) of 0.61, and the Root Mean Squared Error (RMSE) of 0.81 on test dataset. The superior synergistic anti-tumor combinations of D15 (Paeonol[Formula: see text][Formula: see text][Formula: see text]Ethyl gallate) and D13 (Paeoniflorin[Formula: see text][Formula: see text][Formula: see text]Paeonol) were successfully predicted from RPR and experimentally validated on MCF-7 cells. Moreover, the combination of D13 could work as a main contributor to a synergistic anti-proliferative activity in the compatibility of RPR and Cortex Moutan (CM). Our XGBoost model could be a reliable tool for the efficient prediction of synergistic anti-tumor multi-compound combinations from TCM.

A Strategy for the Effective Optimization of Pharmaceutical Formulations Based on Parameter-Optimized Support Vector Machine Model.

Engineering pharmaceutical formulations is governed by a number of variables, and the finding of the optimal preparation is intricately linked to the exploration of a multiparametric space through a variety of optimization tasks. As a result, making such optimization activities simpler is a significant undertaking. For the purposes of this study, we suggested a prediction model that was based on least square support vector machine (LSSVM) and whose parameters were optimized using the particle swarm optimization algorithm (PSO-LSSVM model). Other in silico optimization methods were used and compared, including the LSSVM and the back propagation (BP) neural networks algorithm. PSO-LSSVM demonstrated the highest performance on the test dataset, with the lowest mean square error. In addition, two dosage forms, quercetin solid dispersion and apigenin nanoparticles, were selected as model formulations due to the wide range of formulation compositions and manufacturing factors used in their production. Three different models were used to predict the ideal formulations of two different dosage forms, and in real world, the Taguchi orthogonal design arrays were used to optimize the formulations of each dosage form. It is clear that the predicted performance of two formulations using PSO-LSSVM was both consistent with the outcomes of the Taguchi orthogonal planned experiment, demonstrating the model's good reliability and high usefulness. Together, our PSO-LSSVM prediction model has the potential to accurately predict the best possible formulations, reduce the reliance on experimental effort, accelerate the process of formulation design, and provide a low-cost solution to drug preparation optimization.

Deep learning-based multi-drug synergy prediction model for individually tailored anti-cancer therapies.

While synergistic drug combinations are more effective at fighting tumors with complex pathophysiology, preference compensating mechanisms, and drug resistance, the identification of novel synergistic drug combinations, especially complex higher-order combinations, remains challenging due to the size of combination space. Even though certain computational methods have been used to identify synergistic drug combinations in lieu of traditional in vitro and in vivo screening tests, the majority of previously published work has focused on predicting synergistic drug pairs for specific types of cancer and paid little attention to the sophisticated high-order combinations. The main objective of this study is to develop a deep learning-based approach that integrated multi-omics data to predict novel synergistic multi-drug combinations (DeepMDS) in a given cell line. To develop this approach, we firstly created a dataset comprising of gene expression profiles of cancer cell lines, target information of anti-cancer drugs, and drug response against a large variety of cancer cell lines. Based on the principle of a fully connected feed forward Deep Neural Network, the proposed model was constructed using this dataset, which achieved a high performance with a Mean Square Error (MSE) of 2.50 and a Root Mean Squared Error (RMSE) of 1.58 in the regression task, and gave the best classification accuracy of 0.94, an area under the Receiver Operating Characteristic curve (AUC) of 0.97, a sensitivity of 0.95, and a specificity of 0.93. Furthermore, we utilized three breast cancer cell subtypes (MCF-7, MDA-MD-468 and MDA-MB-231) and one lung cancer cell line A549 to validate the predicted results of our model, showing that the predicted top-ranked multi-drug combinations had superior anti-cancer effects to other combinations, particularly those that were widely used in clinical treatment. Our model has the potential to increase the practicality of expanding the drug combinational space and to leverage its capacity to prioritize the most effective multi-drug combinational therapy for precision oncology applications.

Loss of PDK4 expression promotes proliferation, tumorigenicity, motility and invasion of hepatocellular carcinoma cells.

Although the roles and underlying mechanisms of other PDK family members (i.e., PDK1, PDK2 and PDK3) in tumor progression have been extensively investigated and are well understood, the functions and underlying molecular mechanisms of pyruvate dehydrogenase kinase 4 (PDK4) in the tumorigenesis and progression of various cancers [including hepatocellular carcinoma (HCC)] remain largely unknown. In this study, we examined the expression profile of PDK4 in HCC clinical tissue specimens and the roles of PDK4 in the proliferation, tumorigenicity, motility and invasion of HCC cells. The immunohistochemistry (IHC) and quantitative real-time PCR (qRT-PCR) results revealed that PDK4 was significantly downregulated in the cohort of HCC clinical specimens. Additionally, PDK4 protein was found in both the nucleus and cytoplasm of HCC cells based on an immunofluorescence (ICC) assay, and PDK4 protein was also found in the nucleus and cytoplasm of cancer cells contained in HCC clinical specimens based on IHC. The CCK-8 assay and cell colony formation assay demonstrated that stable depletion of endogenous PDK4 by lentivirus-mediated RNA interference (RNAi) markedly promoted the proliferation of HCC cell lines (i.e., BEL-7402 and BEL-7404 cells) in vitro, while PDK4 silencing significantly enhanced the tumorigenic ability of BEL-7404 cells in vivo. In addition to enhance proliferation and tumorigenesis induced by PDK4 silencing, additional studies demonstrated that knockdown of PDK4 led to increase migration and invasion of BEL-7402 and BEL-7404 cells in vitro. Taken together, these findings suggest that the loss of PDK4 expression contributes to HCC malignant progression.

miR-19 regulates the expression of interferon-induced genes and MHC class I genes in human cancer cells.

MicroRNA-19 (miR-19) is identified as the key oncogenic component of the miR-17-92 cluster. When we explored the functions of the dysregulated miR-19 in lung cancer, microarray-based data unexpectedly demonstrated that some immune and inflammatory response genes (i.e., IL32, IFI6 and IFIT1) were generally down-regulated by miR-19 overexpression in A549 cells, which prompted us to fully investigate whether the miR-19 family (i.e., miR-19a and miR-19b-1) was implicated in regulating the expression of immune and inflammatory response genes in cancer cells. In the present study, we observed that miR-19a or miR-19b-1 overexpression by miRNA mimics in the A549, HCC827 and CNE2 cells significantly downregulated the expression of interferon (IFN)-regulated genes (i.e., IRF7, IFI6, IFIT1, IFITM1, IFI27 and IFI44L). Furthermore, the ectopic miR-19a or miR-19b-1 expression in the A549, HCC827, CNE2 and HONE1 cells led to a general downward trend in the expression profile of major histocompatibility complex (MHC) class I genes (such as HLA-B, HLA-E, HLA-F or HLA-G); conversely, miR-19a or miR-19b-1 inhibition by the miRNA inhibitor upregulated the aforementioned MHC Class I gene expression, suggesting that miR-19a or miR-19b-1 negatively modulates MHC Class I gene expression. The miR-19a or miR-19b-1 mimics reduced the expression of interleukin (IL)-related genes (i.e., IL1B, IL11RA and IL6) in the A549, HCC827, CNE2 or HONE1 cells. The ectopic expression of miR-19a or miR-19b-1 downregulated IL32 expression in the A549 and HCC827 cells and upregulated IL32 expression in CNE2 and HONE1 cells. In addition, enforced miR-19a or miR-19b-1 expression suppressed IL-6 production by lung cancer and nasopharyngeal carcinoma (NPC) cells. Taken together, these findings demonstrate, for the first time, that miR-19 can modulate the expression of IFN-induced genes and MHC class I genes in human cancer cells, suggesting a novel role of miR-19 in linking inflammation and cancer, which remains to be fully characterized.

Optimization Strategy for Generating Gene-edited Tibet Minipigs by Synchronized Oestrus and Cytoplasmic Microinjection.

The Tibet minipig is a rare highland pig breed worldwide and has many applications in biomedical and agricultural research. However, Tibet minipigs are not like domesticated pigs in that their ovulation number is low, which is unfavourable for the collection of zygotes. Partly for this reason, few studies have reported the successful generation of genetically modified Tibet minipigs by zygote injection. To address this issue, we described an efficient way to generate gene-edited Tibet minipigs, the major elements of which include the utilization of synchronized oestrus instead of superovulation to obtain zygotes, optimization of the preparation strategy, and co-injection of clustered regularly interspaced short palindromic repeat sequences associated protein 9 (Cas9) mRNA and single-guide RNAs (sgRNAs) into the cytoplasm of zygotes. We successfully obtained allelic TYR gene knockout (TYR-/-) Tibet minipigs with a typical albino phenotype (i.e., red-coloured eyes with light pink-tinted irises and no pigmentation in the skin and hair) as well as TYR-/-IL2RG-/- and TYR-/-RAG1-/- Tibet minipigs with typical phenotypes of albinism and immunodeficiency, which was characterized by thymic atrophy and abnormal immunocyte proportions. The overall gene editing efficiency was 75% for the TYR single gene knockout, while for TYR-IL2RG and TYR-RAG1 dual gene editing, the values were 25% and 75%, respectively. No detectable off-target mutations were observed. By intercrossing F0 generation minipigs, targeted genetic mutations can also be transmitted to gene-edited minipigs' offspring through germ line transmission. This study is a valuable exploration for the efficient generation of gene-edited Tibet minipigs with medical research value in the future.

Loss of Cirbp expression is correlated with the malignant progression and poor prognosis in nasopharyngeal carcinoma.

Purpose: The correlation of cold-inducible RNA-binding protein (Cirbp) expression with clinicopathological features including patient prognosis in nasopharyngeal carcinoma (NPC) was investigated. Methods: The expression of Cirbp in NPC cell lines and tissue specimens was examined by qRT-PCR or immunohistochemistry (IHC). Results: Immunohistochemistry (IHC) results showed that high Cirbp expression was detected in 61 of 61 non-cancerous nasopharyngeal squamous epithelial biopsies, whereas the significantly reduced expression of Cirbp was observed in NPC specimens. In addition, IHC assay for Cirbp protein illustrated that the cells of 177 NPC samples and nasopharyngeal squamous epithlial cells displayed strong signals in nuclei and faint signals in cytoplasm, whereas Cirbp protein is mainly detected in the cell's cytoplasm in many other cancers. More importantly, TNM classification displayed that the low expression of Cirbp was more frequently observed in T3-T4, N2-N3, M1 and III-IV NPC biopsies, and undifferentiated carcinoma (UDC) than T1-T2, N0-N1, M0 and I-II tumors, and differentiated nonkeratinizing carcinoma (DNKC), suggesting that Cirbp loss is a key molecular event in advanced cases of NPC. Kaplan-Meier survival analysis indicated that NPC patients showing lower Cirbp expression had a significantly shorter overall survival time than those with high Cirbp expression. Multivariate analysis suggested that the level of Cirbp expression was an independent prognostic indicator for NPC survival. Finally, we revealed a significant positive association between Cirbp expression and E-cadherin, and a notable negative correlation between Cirbp expression and Ki67 labeling index in NPC biopsies. Conclusion: Collectively, these findings demonstrate that loss of Cirbp expression is correlated with malignant progression and poor prognosis in NPC.

miR-26a promotes hepatocellular carcinoma invasion and metastasis by inhibiting PTEN and inhibits cell growth by repressing EZH2.

A previous study revealed that therapeutic miR-26a delivery suppresses tumorigenesis in a murine liver cancer model, whereas we found that forced miR-26a expression increased hepatocellular carcinoma (HCC) cell migration and invasion, which prompted us to characterize the causes and mechanisms underlying enhanced invasion due to ectopic miR-26a expression. Gain-of-function and loss-of-function experiments demonstrated that miR-26a promoted migration and invasion of BEL-7402 and HepG2 cells in vitro and positively modulated matrix metalloproteinase (MMP)-1, MMP-2, MMP-9, and MMP-10 expression. In addition, exogenous miR-26a expression significantly enhanced the metastatic ability of HepG2 cells in vivo. miR-26a negatively regulated in vitro proliferation of HCC cells, and miR-26a overexpression suppressed HepG2 cell tumor growth in nude mice. Further studies revealed that miR-26a inhibited cell growth by repressing the methyltransferase EZH2 and promoted cell migration and invasion by inhibiting the phosphatase PTEN. Furthermore, PTEN expression negatively correlated with miR-26a expression in HCC specimens from patients with and without metastasis. Thus, our findings suggest for the first time that miR-26a promotes invasion/metastasis by inhibiting PTEN and inhibits cell proliferation by repressing EZH2 in HCC. More importantly, our data also suggest caution if miR-26a is used as a target for cancer therapy in the future.

Gene Expression Data Based Deep Learning Model for Accurate Prediction of Drug-Induced Liver Injury in Advance.

Drug-induced liver injury (DILI), one of the most common adverse effects, leads to drug development failure or withdrawal from the market in most cases, showing an emerging challenge that is to accurately predict DILI in the early stage. Recently, the vast amount of gene expression data provides us valuable information for distinguishing DILI on a genomic scale. Moreover, the deep learning algorithm is a powerful strategy to automatically learn important features from raw and noisy data and shows great success in the field of medical diagnosis. In this study, a gene expression data based deep learning model was developed to predict DILI in advance by using gene expression data associated with DILI collected from ArrayExpress and then optimized by feature gene selection and parameters optimization. In addition, the previous machine learning algorithm support vector machine (SVM) was also used to construct another prediction model based on the same data sets, comparing the model performance with the optimal DL model. Finally, the evaluation test using 198 randomly selected samples showed that the optimal DL model achieved 97.1% accuracy, 97.4% sensitivity, 96.8% specificity, 0.942 matthews correlation coefficient, and 0.989 area under the ROC curve, while the performance of SVM model only reached 88.9% accuracy, 78.8% sensitivity, 99.0% specificity, 0.794 matthews correlation coefficient, and 0.901 area under the ROC curve. Furthermore, external data sets verification and animal experiments were conducted to assess the optimal DL model performance. Finally, the predicted results of the optimal DL model were almost consistent with experiment results. These results indicated that our gene expression data based deep learning model could systematically and accurately predict DILI in advance. It could be a useful tool to provide safety information for drug discovery and clinical rational drug use in early stage and become an important part of drug safety assessment.

Abnormalities of hair structure and skin histology derived from CRISPR/Cas9-based knockout of phospholipase C-delta 1 in mice.

BACKGROUND: Hairless mice have been widely applied in skin-related researches, while hairless pigs will be an ideal model for skin-related study and other biomedical researches because of the similarity of skin structure with humans. The previous study revealed that hairlessness phenotype in nude mice is caused by insufficient expression of phospholipase C-delta 1 (PLCD1), an essential molecule downstream of Foxn1, which encouraged us to generate PLCD1-deficient pigs. In this study, we plan to firstly produce PLCD1 knockout (KO) mice by CRISPR/Cas9 technology, which will lay a solid foundation for the generation of hairless PLCD1 KO pigs. METHODS: Generation of PLCD1 sgRNAs and Cas 9 mRNA was performed as described (Shao in Nat Protoc 9:2493-2512, 2014). PLCD1-modified mice (F0) were generated via co-microinjection of PLCD1-sgRNA and Cas9 mRNA into the cytoplasm of C57BL/6J zygotes. Homozygous PLCD1-deficient mice (F1) were obtained by intercrossing of F0 mice with the similar mutation. RESULTS: PLCD1-modified mice (F0) showed progressive hair loss after birth and the genotype of CRISPR/Cas9-induced mutations in exon 2 of PLCD1 locus, suggesting the sgRNA is effective to cause mutations that lead to hair growth defect. Homozygous PLCD1-deficient mice (F1) displayed baldness in abdomen and hair sparse in dorsa. Histological abnormalities of the reduced number of hair follicles, irregularly arranged and curved hair follicles, epidermal hyperplasia and disturbed differentiation of epidermis were observed in the PLCD1-deficient mice. Moreover, the expression level of PLCD1 was significantly decreased, while the expression levels of other genes (i.e., Krt1, Krt5, Krt13, loricrin and involucrin) involved in the differentiation of hair follicle were remarkerably increased in skin tissues of PLCD1-deficient mice. CONCLUSIONS: In conclusion, we achieve PLCD1 KO mice by CRISPR/Cas9 technology, which provide a new animal model for hair development research, although homozygotes don't display completely hairless phenotype as expected.