Human major histocompatibility complex (MHC) proteins are encoded by the human leukocyte antigen (HLA) gene complex. When exogenous peptide fragments form peptide-HLA (pHLA) complexes with HLA molecules on the outer surface of cells, they can be recognized by T cells and trigger an immune response. Therefore, determining whether an HLA molecule can bind to a given peptide can improve the efficiency of vaccine design and facilitate the development of immunotherapy. This paper regards peptide fragments as natural language, we combine textCNN and BiLSTM to build a deep neural network model to encode the sequence features of HLA and peptides. Results on independent and external test datasets demonstrate that our CcBHLA model outperforms the state-of-the-art known methods in detecting HLA class I binding peptides. And the method is not limited by the HLA class I allele and the length of the peptide fragment. Users can download the model for binding peptide screening or retrain the model with private data on github (https://github.com/hongliangduan/CcBHLA-pan-specific-peptide-HLA-class-I-binding-prediction-via-Convolutional-and-BiLSTM-features.git).

As a highly versatile therapeutic modality, cyclic peptides have gained signifi-cant attention due to their exceptional binding affinity, minimal toxicity and capaci-ty to target the surface of conventionally "undruggable" proteins. However, the de-velopment of cyclic peptides with therapeutic effects by targeting intracellular bio-logical targets has been hindered by the issue of limited membrane permeability. In this paper, we have conducted an extensive benchmarking analysis of a proprietary dataset consisting of 6941 cyclic peptides, employing machine learning and deep learning models. In addition, we propose an innovative multimodal model called Multi_CycGT which combines a Graph Convolutional Network (GCN) and a Trans-former to extract 1D and 2D features. These encoded features are then fused for the prediction of cyclic peptide permeability. The cross-validation experiments demon-strate that the proposed Multi_CycGT model achieved the highest level of accuracy on the test set, with an accuracy value of 0.8206 and an AUC value of 0.8650. This paper introduces a pioneering deep learning-based approach that demonstrates en-hanced effectiveness in predicting the membrane permeability of cyclic peptides. It also represents the first attempt in this field. We hope that this work will help to ac-celerate the design of cyclic peptide active drugs in medicinal chemistry and chem-ical biology applications.

Deep learning is widely used in chemistry and can rival human chemists in certain scenarios. Inspired by molecule generation in new drug discovery, we present a deep learning-based reaction generation approach to perform reaction generation with the Trans-VAE model in this study. To comprehend how exploratory and innovative the model is in reaction generation, we constructed the dataset by time-split. We applied the Michael addition reaction as the generation vehicle and took the reactions reported before a certain date as the training set and explored whether the model could generate reactions that were reported after the date. We took 2010 and 2015 as the time points for the splitting of the Michael addition reaction respectively. Among the generated reactions, 911 and 487 reactions were applied in the experiments after the respective split time points, accounting for 12.75% and 16.29% of all reported reactions after each time point. The generated results were in line with expectations and additionally generated a large quantity of new chemically feasible Michael addition reactions, which also demonstrated the learnability of the Trans-VAE model for reaction rules. Our research provides a reference for future novel reaction discovery using deep learning.

Deep learning is widely used in chemistry and can rival human chemists in certain scenarios. Inspired by molecule generation in new drug discovery, we present a deep learning-based reaction generation approach to perform reaction generation with the Trans-VAE model in this study. To comprehend how exploratory and innovative the model is in reaction generation, we constructed the dataset by time-split. We applied the Michael addition reaction as the generation vehicle and took the reactions reported before a certain date as the training set and explored whether the model could generate reactions that were reported after the date. We took 2010 and 2015 as the time points for the splitting of the Michael addition reaction respectively. Among the generated reactions, 911 and 487 reactions were applied in the experiments after the respective split time points, accounting for 12.75% and 16.29% of all reported reactions after each time point. The generated results were in line with expectations and additionally generated a large quantity of new chemically feasible Michael addition reactions, which also demonstrated the learnability of the Trans-VAE model for reaction rules. Our research provides a reference for future novel reaction discovery using deep learning.