Search

The rational design of PROTACs is difficult due to their obscure structure-activity relationship. This study introduces a deep neural network model - DeepPROTACs to help design potent PROTACs molecules. It can predict the degradation capacity of a proposed PROTAC molecule based on structures of given target protein and E3 ligase. The experimental dataset is mainly collected from PROTAC-DB and appropriately labeled according to the DC50 and Dmax values. In the model of DeepPROTACs, the ligands as well as the ligand binding pockets are generated and represented with graphs and fed into Graph Convolutional Networks for feature extraction. While SMILES representations of linkers are fed into a Bidirectional Long Short-Term Memory layer to generate the features. Experiments show that DeepPROTACs model achieves 77.95% average prediction accuracy and 0.8470 area under receiver operating characteristic curve on the test set. DeepPROTACs is available online at a web server (https://bailab.siais.shanghaitech.edu.cn/services/deepprotacs/) and at github (https://github.com/fenglei104/DeepPROTACs).

Grinding force is the most important characteristic parameter in the study of grinding machining mechanism, and it has an important influence on the quality of the machined surface and the service life of the grinding wheel. Accurate prediction of grinding force is significant to process improvement, machine tool design, and condition diagnosis. This paper uses the ABAQUS software to simulate the end grinding process of Si3N4 ceramics. In the abrasive removal model, in order to express the complex grinding conditions of the fine-grained grinding wheel better, this paper adopts the truncated polyhedral abrasive grain model, and the abrasive grains are randomly rotated multiple times to verify the model, which reduces the simulation grinding force error caused by the randomness of irregular abrasive grain model and improves the prediction accuracy. This paper establishes an abrasive grain distribution model, the unit number of abrasive grains is calculated from the equivalent volume of abrasive grains, and the grinding force of a single abrasive grain extends to the end surface; in the process simulation model, considering the influence of the change of radial linear velocity on the grinding force, the final grinding force is obtained by partitioning and summing. The experimental comparison shows that the grinding force model has high prediction accuracy in the end grinding with the fine-grained grinding wheel.

When automated pattern matching tools are used to execute user-knowledge pattern matching (UKPM) in the expert knowledge recommendation system (EKRS), user-knowledge matching is uncertain and the matching efficiency is low. To solve the above problems, the dynamic UKPM mathematical model is established and the “Entropy-Beta” method of crowdsourcing task assignment is designed to solve the model in the study. Firstly, the concept of Entropy is combined with crowdsourcing. The uncertainty of user-knowledge matching results is measured and the magnitude of the uncertainty is calculated. Secondly, based on the Beta distribution function, the accuracy of matching results is measured. The optimal matching results are selected and the matching results were sent to EKRS according to the matching probability. Thirdly, the knowledge recommendation process of UKPM is dynamically adjusted according to the matching probability. Finally, the comparison results of several algorithms showed that the Entropy-Beta algorithm could largely improve the accuracy, efficiency, dynamic regulation, and other performances of EKRS.

TPN729 is a novel phosphodiesterase 5 (PDE5) inhibitor used to treat erectile dysfunction in men. Our previous study shows that the plasma exposure of metabolite M3 (N-dealkylation of TPN729) in humans is much higher than that of TPN729. In this study, we compared its metabolism and pharmacokinetics in different species and explored the contribution of its main metabolite M3 to pharmacological effect. We conducted a combinatory approach of ultra-performance liquid chromatography/quadrupole time-of-flight mass spectrometry-based metabolite identification, and examined pharmacokinetic profiles in monkeys, dogs, and rats following TPN729 administration. A remarkable species difference was observed in the relative abundance of major metabolite M3: i.e., the plasma exposure of M3 was 7.6-fold higher than that of TPN729 in humans, and 3.5-, 1.2-, 1.1-fold in monkeys, dogs, and rats, respectively. We incubated liver S9 and liver microsomes with TPN729 and CYP3A inhibitors, and demonstrated that CYP3A was responsible for TPN729 metabolism and M3 formation in humans. The inhibitory activity of M3 on PDE5 was 0.78-fold that of TPN729 (The IC(50) values of TPN729 and M3 for PDE5A were 6.17 ± 0.48 and 7.94 ± 0.07 nM, respectively.). The plasma protein binding rates of TPN729 and M3 in humans were 92.7% and 98.7%, respectively. It was astonishing that the catalyzing capability of CYP3A4 in M3 formation exhibited seven-fold disparity between different species. M3 was an active metabolite, and its pharmacological contribution was equal to that of TPN729 in humans. These findings provide new insights into the limitation and selection of animal model for predicting the clinical pharmacokinetics of drug candidates metabolized by CYP3A4.