Neural networks for anatomical therapeutic chemical (ATC) classification

Prediction of drug's anatomical therapeutic chemical (ATC) code by constructing biological profiles of ATC codes

BACKGROUND

The Anatomical Therapeutic Chemical (ATC) classification system, proposed and maintained by the World Health Organization, is among the most widely used drug classification schemes. Recently, it has become a key research focus in drug repositioning. Computational models often pair drugs with ATC codes to explore drug-ATC code associations. However, the limited information available for ATC codes constrains these models, leaving significant room for improvement.

RESULTS

This study presents an inference method to identify highly related target proteins, structural features, and side effects for each ATC code, constructing comprehensive biological profiles. Association networks for target proteins, structural features, and side effects are established, and a random walk with restart algorithm is applied to these networks to extract raw associations. A permutation test is then conducted to exclude false positives, yielding robust biological profiles for ATC codes. These profiles are used to construct new ATC code kernels, which are integrated with ATC code kernels from the existing model PDATC-NCPMKL. The recommendation matrix is subsequently generated using the procedures of PDATC-NCPMKL. Cross-validation results demonstrate that the new model achieves AUROC and AUPR values exceeding 0.96.

CONCLUSION

The proposed model outperforms PDATC-NCPMKL and other previous models. Analysis of the contributions of the newly added ATC code kernels confirms the value of biological profiles in enhancing the prediction of drug-ATC code associations.

PDATC-NCPMKL: Predicting drug's Anatomical Therapeutic Chemical (ATC) codes based on network consistency projection and multiple kernel learning

The development and discovery of new drugs is time-consuming and needs lots of human and material resources. Therefore, discovery of novel effects of existing drugs is an important alternative way, which can accelerate the process of designing "new" drugs. The anatomical Therapeutic Chemical (ATC) classification system recommended by World Health Organization (WHO) is a basic research area in this regard. A novel ATC code of an existing drug suggests its novel effects. Some computational models have been proposed, which can predict the drug-ATC code associations. However, their performance is not very high. There still exist spaces for improvement. In this study, a new recommendation system (named PDATC-NCPMKL), which incorporated network consistency projection and multi-kernel learning, was designed to identify drug-ATC code associations. For drugs or ATC codes, several kernels were constructed, which were fused by a multiple kernel learning method and an additional kernel integration scheme. To enhance the performance, the drug-ATC code association adjacency matrix was reformulated by a variant of weighted K nearest known neighbors (WKNKN). The reformulated adjacency matrix, drug and ATC code kernels were fed into network consistency projection to generate the association score matrix. The proposed recommendation system was tested on the ATC codes at the second, third and fourth levels in drug ATC classification system using ten-fold cross-validation. The results indicated that all AUROC and AUPR values were close to or exceeded 0.96. Such performance was higher than some existing computational models. Some additional tests were conducted to prove the utility of adjacency matrix reformulation and to analyze the importance of drug and ATC code kernels.

Ensemble of Networks for Multilabel Classification

Multilabel learning goes beyond standard supervised learning models by associating a sample with more than one class label. Among the many techniques developed in the last decade to handle multilabel learning best approaches are those harnessing the power of ensembles and deep learners. This work proposes merging both methods by combining a set of gated recurrent units, temporal convolutional neural networks, and long short-term memory networks trained with variants of the Adam optimization approach. We examine many Adam variants, each fundamentally based on the difference between present and past gradients, with step size adjusted for each parameter. We also combine Incorporating Multiple Clustering Centers and a bootstrap-aggregated decision trees ensemble, which is shown to further boost classification performance. In addition, we provide an ablation study for assessing the performance improvement that each module of our ensemble produces. Multiple experiments on a large set of datasets representing a wide variety of multilabel tasks demonstrate the robustness of our best ensemble, which is shown to outperform the state-of-the-art.

Identifying the kind behind SMILES-anatomical therapeutic chemical classification using structure-only representations

Anatomical Therapeutic Chemical (ATC) classification for compounds/drugs plays an important role in drug development and basic research. However, previous methods depend on interactions extracted from STITCH dataset which may make it depend on lab experiments. We present a pilot study to explore the possibility of conducting the ATC prediction solely based on the molecular structures. The motivation is to eliminate the reliance on the costly lab experiments so that the characteristics of a drug can be pre-assessed for better decision-making and effort-saving before the actual development. To this end, we construct a new benchmark consisting of 4545 compounds which is with larger scale than the one used in previous study. A light-weight prediction model is proposed. The model is with better explainability in the sense that it is consists of a straightforward tokenization that extracts and embeds statistically and physicochemically meaningful tokens, and a deep network backed by a set of pyramid kernels to capture multi-resolution chemical structural characteristics. Its efficacy has been validated in the experiments where it outperforms the state-of-the-art methods by 15.53% in accuracy and by 69.66% in terms of efficiency. We make the benchmark dataset, source code and web server open to ease the reproduction of this study.