Finding Chemical Structures Corresponding to a Set of Coordinates in Chemical Descriptor Space

Abstract

When chemical structures are searched based on descriptor values, or descriptors are interpreted based on values, it is important that corresponding chemical structures actually exist. In order to consider the existence of chemical structures located in a specific region in the chemical space, we propose to search them inside training data domains (TDDs), which are dense areas of a training dataset in the chemical space. We investigated TDDs' features using diverse and local datasets, assuming that GDB11 is the chemical universe. These two analyses showed that considering TDDs gives higher chance of finding chemical structures than a random search-based method, and that novel chemical structures actually exist inside TDDs. In addition to those findings, we tested the hypothesis that chemical structures were distributed on the limited areas of chemical space. This hypothesis was confirmed by the fact that distances among chemical structures in several descriptor spaces were much shorter than those among randomly generated coordinates in the training data range

Recurrent Neural Network Model for Constructive Peptide Design

“…This property is relevant and potentially promising for molecular de novo design and inverse QSAR modeling. 4,5 Although the overall concept of constructive modeling is not new, there has been considerable recent interest in this and related "deep learning" approaches, which is partly motivated by the availability of customized computer hard-and software solutions. 6−8 Long short-term memory (LSTM) models belong to the class of recurrent neural networks (RNNs) incorporating socalled memory units.…”

“…In this context, the umbrella term “constructive machine learning” describes an entire class of problem-solving methods for which the ultimate learning goal is not necessarily to find the optimal model of the training data but rather to identify new instances (e.g., molecules) from within the applicability domain of the model that are likely to exhibit the desired properties. − In contrast to models that are typically used to classify a given set of unlabeled domain instances post hoc, constructive machine learning models are generative and able to capture an infinite or exponentially large combinatorial search space. This property is relevant and potentially promising for molecular de novo design and inverse QSAR modeling. , Although the overall concept of constructive modeling is not new, there has been considerable recent interest in this and related “deep learning” approaches, which is partly motivated by the availability of customized computer hard- and software solutions. − …”

CReM: chemically reasonable mutations framework for structure generation

“…The conventional descriptors proved to be very difficult to reconstruct structures. Several approaches were developed in the past but gained low popularity due to many restrictions and limitations [25][26][27][28][29][30]. The recent advances in deep learning allowed to generate a latent representation of input compounds, find an optimal set of the latent variables associated with desired property values and sample structures from this latent subspace of variables [31].…”

CReM: chemically reasonable mutations framework for structure generation

“…Several approaches were developed in the past but gain low popularity due to many restrictions and limitations [17][18][19][20][21][22]. Recent advances in deep learning allowed to generate latent representation of input compounds, find an optimal set of the latent variables associated with desired property values and sample structures from this latent subspace of variables [23].…”