MOSES introduces a comprehensive benchmarking platform for molecular generative models, offering standardized datasets, evaluation metrics, and baselines. By providing a unified measuring stick, it aims to resolve reproducibility challenges in chemical distribution learning.
ChemBERTa-3 provides a unified, scalable infrastructure for pretraining and benchmarking chemical foundation models. It addresses reproducibility gaps in previous studies like MoLFormer through standardized scaffold splitting and open-source tooling.
ChemDFM-R is a 14B-parameter chemical reasoning model that integrates a 101B-token dataset of atomized chemical knowledge. Using a mix-sourced distillation strategy and domain-specific reinforcement learning, it outperforms similarly sized models and DeepSeek-R1 on ChemEval.
This work investigates the scaling hypothesis for molecular transformers, training RoBERTa models on 77M SMILES from PubChem. It compares Masked Language Modeling (MLM) against Multi-Task Regression (MTR) pretraining, finding that MTR yields better downstream performance but is computationally heavier.
This paper introduces ChemBERTa, a RoBERTa-based model pretrained on 77M SMILES strings. It systematically evaluates the impact of pretraining dataset size, tokenization strategies, and input representations (SMILES vs. SELFIES) on downstream MoleculeNet tasks, finding that performance scales positively with data size.
MERMaid leverages fine-tuned vision models and VLM reasoning to mine chemical reaction data directly from PDF figures and tables. By handling context inference and coreference resolution, it builds high-fidelity knowledge graphs with 87% end-to-end accuracy.
Comprehensive evaluation of 8 optical chemical structure recognition tools using a newly curated dataset of 2,702 patent images. Proposes ChemIC, a ResNet-50 classifier to route images to specialized tools based on content type, demonstrating that no single tool excels at all tasks.
ChemReco automates the recognition of hand-drawn chemical structures using a synthetic data pipeline and an EfficientNet+Transformer architecture, achieving 96.90% accuracy on C-H-O molecules.
DECIMER.ai addresses the lack of open tools for Optical Chemical Structure Recognition (OCSR) by providing a comprehensive, deep-learning-based workflow. It features a novel data generation pipeline (RanDepict), a web application, and models for segmentation and recognition that rival or exceed proprietary solutions.
Proposes a new architecture (DGAT) to resolve global context loss in chemical structure recognition. Introduces Cascaded Global Feature Enhancement and Sparse Differential Global-Local Attention, achieving 84.0% BLEU-4 and handling complex chiral structures implicitly.
This paper presents an enhanced deep learning architecture for Optical Chemical Structure Recognition (OCSR) specifically optimized for hand-drawn inputs. By pairing an EfficientNetV2 encoder with a Transformer decoder and training on over 150 million synthetic images, the model achieves 73.25% exact match accuracy on a real-world hand-drawn benchmark of 5,088 images.
MMSSC-Net introduces a multi-stage cognitive approach for OCSR, utilizing a SwinV2 encoder and GPT-2 decoder to recognize atomic and bond sequences. It achieves 75-98% accuracy across benchmark datasets by handling varying image resolutions and noise through fine-grained perception of atoms and bonds.