Binding prediction between targets and drug-like compounds through Deep Neural Networks have generated promising results in recent years, outperforming traditional machine learning-based methods. However, the generalization capability of these classification models is still an issue to be addressed. In this work, we explored how di↵erent cross-validation strategies applied to data from di↵erent molecular databases a↵ect to the performance of binding prediction proteochemometrics models. These strategies are: (1) random splitting, (2) splitting based on K-means clustering (both 1 of actives and inactives), (3) splitting based on source database and (4) splitting based both in the clustering and in the source database. These schemas are applied to a Deep Learning proteochemometrics model and to a simple logistic regression model to be used as baseline. Additionally, two di↵erent ways of describing molecules in the model are tested: (1) by their SMILES and (2) by three fingerprints. The classification performance of our Deep Learning-based proteochemometrics model is comparable to the state of the art. Our results show that the lack of generalization of these models is due to a bias in public molecular databases and that a restrictive cross-validation schema based on compounds clustering leads to worse but more robust and credible results. Our results also show better performance when representing molecules by their fingerprints.

Binding prediction between targets and drug-like compounds through Deep Neural Networks have generated promising results in recent years, outperforming traditional machine learning-based methods. However, the generalization capability of these classification models is still an issue to be addressed. In this work, we explored how different cross-validation strategies applied to data from different molecular databases affect to the performance of binding prediction proteochemometrics models. These strategies are: (1) random splitting, (2) splitting based on K-means clustering (both of actives and inactives), (3) splitting based on source database and (4) splitting based both in the clustering and in the source database. These schemas are applied to a Deep Learning proteochemometrics model and to a simple logistic regression model to be used as baseline. Additionally, two different ways of describing molecules in the model are tested: (1) by their SMILES and (2) by three fingerprints. The classification performance of our Deep Learning-based proteochemometrics model is comparable to the state of the art. Our results show that the lack of generalization of these models is due to a bias in public molecular databases and that a restrictive cross-validation schema based on compounds clustering leads to worse but more robust and credible results. Our results also show better performance when representing molecules by their fingerprints.

The use of raw amino acid sequences as input for deep learning models for protein functional prediction has gained popularity in recent years. This scheme obliges to manage proteins with different lengths, while deep learning models require same-shape input. To accomplish this, zeros are usually added to each sequence up to a established common length in a process called zero-padding. However, the effect of different padding strategies on model performance and data structure is yet unknown. We propose and implement four novel types of padding the amino acid sequences. Then, we analysed the impact of different ways of padding the amino acid sequences in a hierarchical Enzyme Commission number prediction problem. Results show that padding has an effect on model performance even when there are convolutional layers implied. Contrastingly to most of deep learning works which focus mainly on architectures, this study highlights the relevance of the deemed-of-lowimportance process of padding and raises awareness of the need to refine it for better performance. The code of this analysis is publicly available at https ://githu b.com/b2sla b/paddi ng_bench mark. Since the breakthrough of deep learning (DL) 1 , deep neural networks are being successfully applied in computational biology 2,3. This is due to their capacity for automatically extracting meaningful features from raw data 4. Specifically, DL is useful in the context of biological sequences, such as proteins or RNA, because it can learn directly from the sequence and hence, capture nonlinear dependencies and interaction effects. Some examples of applications of DL on biological sequences include prediction of specifities of DNA and RNA binding proteins 5 , DNA function quantification 6 , de novo peptide design 7 , detection of conserved DNA fragments 8 , prediction of protein associated GO terms 9 or quantification of the impact of genetic variation on gene regulatory mechanisms 3. The specific DL architectures able to leverage the inner structure of sequential biological data are Convolutional Neural Networks (CNN) and Recurrent Neural Networks (RNN). CNNs entail translational invariance 10 and can be used to find relevant patterns with biological meaning 5,8,11,12. For their part, bidirectional RNNs (and the derived Long Short-Term Memory and Gated Recurrent Units) are appropiate for modelling biological sequences since they are suited for data with a sequential but non-causal structure, variable length, and long-range dependencies 13-16. Both architectures are usually combined, as in DEEPre 17 , where a CNN-RNN model performs a hierarchical classification of enzymes. Proteins are long linear sequences constituted by amino acid residues attached covalently. These amino acid residues are represented by letters that cannot be directly processed by the mathematical operations used by DL models. Choosing how to digitally encode amino acids is a crucial step in this context, since it can affect to the overall performance of the models 18. A comprehensive review and a...