Filtern
Erscheinungsjahr
Dokumenttyp
- Wissenschaftlicher Artikel (24)
- Konferenzveröffentlichung (13)
- Preprint (7)
Schlagworte
- OCSR (3)
- CDK (2)
- DECIMER (2)
- Deep Learning (2)
- OCSR, Optical Chemical Structure Recognition (2)
- Transformer (2)
- AI (1)
- Chemical image depiction (1)
- Chemical space (1)
- Chemical structure depictions (1)
Inspired by the super-human performance of deep learning models in playing the game of Go after being presented with virtually unlimited training data, we looked into areas in chemistry where similar situations could be achieved. Encountering large amounts of training data in chemistry is still rare, so we turned to two areas where realistic training data can be fabricated in large quantities, namely a) the recognition of machine-readable structures from images of chemical diagrams and b) the conversion of IUPAC(-like) names into structures and vice versa. In this talk, we outline the challenges, technical implementation and results of this study.
Optical Chemical Structure Recognition (OCSR): Vast amounts of chemical information remain hidden in the primary literature and have yet to be curated into open-access databases. To automate the process of extracting chemical structures from scientific papers, we developed the DECIMER.ai project. This open-source platform provides an integrated solution for identifying, segmenting, and recognising chemical structure depictions in scientific literature. DECIMER.ai comprises three main components: DECIMER-Segmentation, which utilises a Mask-RCNN model to detect and segment images of chemical structure depictions; DECIMER-Image Classifier EfficientNet-based classification model identifies which images contain chemical structures and DECIMER-Image Transformer which acts as an OCSR engine which combines an encoder-decoder model to convert the segmented chemical structure images into machine-readable formats, like the SMILES string.
DECIMER.ai is data-driven, relying solely on the training data to make accurate predictions without hand-coded rules or assumptions. The latest model was trained with 127 million structures and 483 million depictions (4 different per structure) on Google TPU-V4 VMs
Name to Structure Conversion: The conversion of structures to IUPAC(-like) or systematic names has been solved algorithmically or rule-based in satisfying ways. This fact, on the other side, provided us with an opportunity to generate a name-structure training pair at a very large scale to train a proof-of-concept transformer network and evaluate its performance.
In this work, the largest model was trained using almost one billion SMILES strings. The Lexichem software utility from OpenEye was employed to generate the IUPAC names used in the training process. STOUT V2 was trained on Google TPU-V4 VMs. The model's accuracy was validated through one-to-one string matching, BLEU scores, and Tanimoto similarity calculations. To further verify the model's reliability, every IUPAC name generated by STOUT V2 was analysed for accuracy and retranslated using OPSIN, a widely used open-source software for converting IUPAC names to SMILES. This additional validation step confirmed the high fidelity of STOUT V2's translations.
The DECIMER.ai Project
(2024)
Over the past few decades, the number of publications describing chemical structures and their metadata has increased significantly. Chemists have published the majority of this information as bitmap images along with other important information as human-readable text in printed literature and have never been retained and preserved in publicly available databases as machine-readable formats. Manually extracting such data from printed literature is error-prone, time-consuming, and tedious. The recognition and translation of images of chemical structures from printed literature into machine-readable format is known as Optical Chemical Structure Recognition (OCSR). In recent years, deep-learning-based OCSR tools have become increasingly popular. While many of these tools claim to be highly accurate, they are either unavailable to the public or proprietary. Meanwhile, the available open-source tools are significantly time-consuming to set up. Furthermore, none of these offers an end-to-end workflow capable of detecting chemical structures, segmenting them, classifying them, and translating them into machine-readable formats.
To address this issue, we present the DECIMER.ai project, an open-source platform that provides an integrated solution for identifying, segmenting, and recognizing chemical structure depictions within the scientific literature. DECIMER.ai comprises three main components: DECIMER-Segmentation, which utilizes a Mask-RCNN model to detect and segment images of chemical structure depictions; DECIMER-Image Classifier EfficientNet-based classification model identifies which images contain chemical structures and DECIMER-Image Transformer which acts as an OCSR engine which combines an encoder-decoder model to convert the segmented chemical structure images into machine-readable formats, like the SMILES string.
A key strength of DECIMER.ai is that its algorithms are data-driven, relying solely on the training data to make accurate predictions without any hand-coded rules or assumptions. By offering this comprehensive, open-source, and transparent pipeline, DECIMER.ai enables automated extraction and representation of chemical data from unstructured publications, facilitating applications in chemoinformatics and drug discovery.
The use of molecular string representations for deep learning in chemistry has been steadily increasing in recent years. The complexity of existing string representations, and the difficulty in creating meaningful tokens from them, lead to the development of new string representations for chemical structures. In this study, the translation of chemical structure depictions in the form of bitmap images to corresponding molecular string representations was examined. An analysis of the recently developed DeepSMILES and SELFIES representations in comparison with the most commonly used SMILES representation is presented where the ability to translate image features into string representations with transformer models was specifically tested. The SMILES representation exhibits the best overall performance whereas SELFIES guarantee valid chemical structures. DeepSMILES perform in between SMILES and SELFIES, InChIs are not appropriate for the learning task. All investigations were performed using publicly available datasets and the code used to train and evaluate the models has been made available to the public.