Advances in cryo-electron tomography (cryo-ET) have enabled the visualization of molecules within their native cellular environments in three-dimensions (3D). These visualizations are essential for studying the functions of biological entities in their natural conditions. Recently, deep learning techniques have shown significant success in tackling the challenge of particle detection in cryo-ET data. However, accurately identifying and classifying multi-class molecules remain challenging due to factors like low signal-to-noise ratios and the wide range of particle sizes. In this study, we introduce a novel framework CFNPicker for 3D object detection applied to cryo-ET analysis. A major advantage of our method is the design of central feature network (CFN) to integrate central features across multiple scales, allowing for the accurate detection of both small ( $⩽$200) and large ( $⩾$600) molecules. Additionally, we propose an adaptive weighted sampling training strategy to distinguish the complex noise distribution in the background, reducing false positive particles. We also construct the localization label to explicitly utilize the size and position variations of multi-class protein structures. Compared with existing methods, CFN improves the F1 score for classification by 3.6%, 7.3%, 6.6%, and 5.1% for the four smallest molecules tested respectively, while preserving similar or higher F1 scores for other molecules analyzed.

Mass spectrometry plays a crucial role in biomedicine by detecting isotopes, contributing significantly to research, diagnostics, and therapy development. This study introduces IsoFusion, a deep learning model designed to address isotope detection in raw mass spectra. Rather than directly applying convolutional layers to all signal and noise peaks, IsoFusion employs a trial-and-error strategy. First, it investigates all potential charge states (trials) and collects signal peaks around expected m/z values for each trial. Then, convolutional layers extract features from each trial, which are fused to identify the correct one. Finally, the reparameterization trick predicts isotope features based on this correct trial. A key advantage of IsoFusion is shared model parameters across all trials, enhancing feature learning for less common charge states using data from prevalent ones. Our results show a significant accuracy improvement for charge state 5, reaching 99.42%, compared to DeepIso’s 43.36%. Moreover, IsoFusion achieves a 97.33% detection accuracy for isotopes, with 2.4% of detected isotopes previously unidentified by four commonly used methods.

The effectiveness of AI-driven drug discovery can be enhanced by pretraining on small molecules. However, the conventional masked language model pretraining techniques are not suitable for molecule pretraining due to the limited vocabulary size and the non-sequential structure of molecules. To overcome these challenges, we propose FragAdd, a strategy that involves adding a chemically implausible molecular fragment to the input molecule. This approach allows for the incorporation of rich local information and the generation of a high-quality graph representation, which is advantageous for tasks like virtual screening. Consequently, we have developed a virtual screening protocol that focuses on identifying estrogen receptor alpha binders on a nucleus receptor. Our results demonstrate a significant improvement in the binding capacity of the retrieved molecules. Additionally, we demonstrate that the FragAdd strategy can be combined with other self-supervised methods to further expedite the drug discovery process.