Chemical structure-aware molecular image representation learning.

Current methods of molecular image-based drug discovery face two major challenges: (1) work effectively in absence of labels, and (2) capture chemical structure from implicitly encoded images. Given that chemical structures are explicitly encoded by molecular graphs (such as nitrogen, benzene rings and double bonds), we leverage self-supervised contrastive learning to transfer chemical knowledge from graphs to images. Specifically, we propose a novel Contrastive Graph-Image Pre-training (CGIP) framework for molecular representation learning, which learns explicit information in graphs and implicit information in images from large-scale unlabeled molecules via carefully designed intra- and inter-modal contrastive learning. We evaluate the performance of CGIP on multiple experimental settings (molecular property prediction, cross-modal retrieval and distribution similarity), and the results show that CGIP can achieve state-of-the-art performance on all 12 benchmark datasets and demonstrate that CGIP transfers chemical knowledge in graphs to molecular images, enabling image encoder to perceive chemical structures in images. We hope this simple and effective framework will inspire people to think about the value of image for molecular representation learning.

Disease Quantification of Liver Lymphoma in CT Images without Lesion Segmentation.

AIM: This study aimed to automatically implement liver disease quantification (DQ) in lymphoma using CT images without lesion segmentation. BACKGROUND: Computed Tomography (CT) imaging manifestations of liver lymphoma include diffuse infiltration, blurred boundaries, vascular drift signs, and multiple lesions, making liver lymphoma segmentation extremely challenging. METHODS: The method includes two steps: liver recognition and liver disease quantification. We use the transfer learning technique to recognize the diseased livers automatically and delineate the livers manually using the CAVASS software. When the liver is recognized, liver disease quantification is performed using the disease map model. We test our method in 10 patients with liver lymphoma. A random grouping cross-validation strategy is used to evaluate the quantification accuracy of the manual and automatic methods, with reference to the ground truth. RESULTS: We split the 10 subjects into two groups based on lesion size. The average accuracy for the total lesion burden (TLB) quantification is 91.76%±0.093 for the group with large lesions and 95.57%±0.032 for the group with small lesions using the manual organ (MO) method. An accuracy of 85.44%±0.146 for the group with larger lesions and 81.94%±0.206 for the small lesion group is obtained using the automatic organ (AO) method, with reference to the ground truth. CONCLUSION: Our DQ-MO and DQ-AO methods show good performance for varied lymphoma morphologies, from homogeneous to heterogeneous, and from single to multiple lesions in one subject. Our method can also be extended to CT images of other organs in the abdomen for disease quantification, such as Kidney, Spleen and Gallbladder.

Deep-learning based approach to identify substrates of human E3 ubiquitin ligases and deubiquitinases.

E3 ubiquitin ligases (E3s) and deubiquitinating enzymes (DUBs) play key roles in protein degradation. However, a large number of E3 substrate interactions (ESIs) and DUB substrate interactions (DSIs) remain elusive. Here, we present DeepUSI, a deep learning-based framework to identify ESIs and DSIs using the rich information present in protein sequences. Utilizing the collected golden standard dataset, key hyperparameters in the process of model training, including the ones relevant to data sampling and number of epochs, have been systematically assessed. The performance of DeepUSI was thoroughly evaluated by multiple metrics, based on internal and external validation. Application of DeepUSI to cancer-associated E3 and DUB genes identified a list of druggable substrates with functional implications, warranting further investigation. Together, DeepUSI presents a new framework for predicting substrates of E3 ubiquitin ligases and deubiquitinates.

Artificial Intelligence Applications in Rare Skin Diseases / 罕见病研究

In clinical practice, early diagnosis, accurate assessment and effective management of rare skin diseases are difficult. The Big Data gives rise to the exponential growth of biomedical data, including medical images, multi-omics information and electronic health records. Artificial intelligence (AI), particularly machine learning, has its advantage in processing complex and abundant information. Researches have applied AI in the field of rare skin diseases. In this paper, we briefly describe, discuss, and foresee the research on AI based image data, multi-omics data & text data and AI in assisting rare skin disease drug development, in order to improve the awareness of dermatologist understanding of this field and actively promote the development of AI usage on rare skin diseases.

An efficient method for building a database of diatom populations for drowning site inference using a deep learning algorithm.

Seasonal or monthly databases of the diatom populations in specific bodies of water are needed to infer the drowning site of a drowned body. However, existing diatom testing methods are laborious, time-consuming, and costly and usually require specific expertise. In this study, we developed an artificial intelligence (AI)-based system as a substitute for manual morphological examination capable of identifying and classifying diatoms at the species level. Within two days, the system collected information on diatom profiles in the Huangpu and Suzhou Rivers of Shanghai, China. In an animal experiment, the similarities of diatom profiles between lung tissues and water samples were evaluated through a modified Jensen-Shannon (JS) divergence measure for drowning site inference, reaching a prediction accuracy of 92.31%. Considering its high efficiency and simplicity, our proposed method is believed to be more applicable than existing methods for seasonal or monthly water monitoring of diatom populations from sections of interconnected rivers, which would help police narrow the investigation scope to confirm the identity of an immersed body.

Building a Compact Convolutional Neural Network for Embedded Intelligent Sensor Systems Using Group Sparsity and Knowledge Distillation.

As artificial intelligence (AI)- or deep-learning-based technologies become more popular,the main research interest in the field is not only on their accuracy, but also their efficiency, e.g., theability to give immediate results on the users' inputs. To achieve this, there have been many attemptsto embed deep learning technology on intelligent sensors. However, there are still many obstacles inembedding a deep network in sensors with limited resources. Most importantly, there is an apparenttrade-off between the complexity of a network and its processing time, and finding a structure witha better trade-off curve is vital for successful applications in intelligent sensors. In this paper, wepropose two strategies for designing a compact deep network that maintains the required level ofperformance even after minimizing the computations. The first strategy is to automatically determinethe number of parameters of a network by utilizing group sparsity and knowledge distillation (KD)in the training process. By doing so, KD can compensate for the possible losses in accuracy causedby enforcing sparsity. Nevertheless, a problem in applying the first strategy is the unclarity indetermining the balance between the accuracy improvement due to KD and the parameter reductionby sparse regularization. To handle this balancing problem, we propose a second strategy: a feedbackcontrol mechanism based on the proportional control theory. The feedback control logic determinesthe amount of emphasis to be put on network sparsity during training and is controlled based onthe comparative accuracy losses of the teacher and student models in the training. A surprising facthere is that this control scheme not only determines an appropriate trade-off point, but also improvesthe trade-off curve itself. The results of experiments on CIFAR-10, CIFAR-100, and ImageNet32 X 32datasets show that the proposed method is effective in building a compact network while preventingperformance degradation due to sparsity regularization much better than other baselines.