Img2Mol - Accurate SMILES Recognition from Molecular Graphical Depictions

Abstract

<p>Automatic recognition of the molecular content of a molecule’s graphical depiction is an extremely challenging problem that remains largely unsolved despite decades of research. Recent advances in neural machine translation enable the auto-encoding of molecular structures in a continuous vector space of fixed size (latent representation) with low reconstruction errors. In this paper, we present a fast and accurate model combining a deep convolutional neural network learning from molecule depictions and a pre-trained decoder that translates the latent representation into the SMILES representation of the molecules. This combination allows to precisely infer a molecular structure from an image. Our rigorous evaluation show that Img2Mol is able to correctly translate up to 88% of the molecular depictions into their SMILES representation. A pretrained version of Img2Mol is made publicly available on GitHub for non-commercial users.</p>

RanDepict - Random Chemical Structure Depiction Generator

“…Numerical ranges are split into three subranges which are then treated like categories. For example, if the bond width could be described by an integer with the possible values [1,2,3,4,5,6], this would be allocated to three positions in the fingerprint. These positions would be linked to the subsets [1,2], [3,4] and [5,6].…”

“…For example, if the bond width could be described by an integer with the possible values [1,2,3,4,5,6], this would be allocated to three positions in the fingerprint. These positions would be linked to the subsets [1,2], [3,4] and [5,6]. This means that the fingerprint does not always define an exact value for certain parameters but only specifies a range.…”

“…Since 2019, there has been a lot of development in the field of deep learning-based optical chemical structure recognition (OCSR) [1][2][3][4][5][6][7]. This indicates a paradigm shift as convolutional neural networks (CNN) as encoders in combination with recurrent neural networks (RNN) or transformers as decoders replace the rule-based systems that have previously defined the standard in the field [8].…”

“…Every single step in these workflows can be fine-tuned to achieve optimal results. In 2021, Clevert et al have shown that the openly available rule-based systems surprisingly fail on the common benchmark datasets when slight image perturbations like rotation and shearing are introduced [3]. This lack of robustness is a clear indication that these systems have been overfitted to the benchmark datasets and that there is a need for more diverse benchmark data.…”

Multi-task Proteochemometric Modelling

“…These are 512-dimensional descriptors learned in an unsupervised way using a recurrent autoencoder to translate between non-canonical SMILES and their canonical form. They have been shown to be very effective in QSAR modeling, as well as PCM modelling [21,12], inverse molecular problems such as optical chemical structure recognition [7] or reverse-engineering of molecular structures [14].…”