Predicting Microbe-Disease Association Based On Sparse Representation And Few-Shot Learning

Research has shown that there are close interactions between the human host and many microbe communities,including bacteria,archaea,viruses,protozoa,and fungi.Increasing clinical evidence indicates that the microbe within the human body plays a significant role in the development,progression,and exacerbation of various complex diseases in humans.Therefore,identifying potential disease-associated microbes can provide deeper insights into the mechanisms of human diseases,and facilitate disease prevention,diagnosis,and treatment.However,traditional experimental-based methods are both costly and time-consuming.In recent years,with the advancement of machine learning and artificial intelligence technologies,researchers have started exploring the application of these technologies to address the prediction of microbe-disease associations.However,predicting potential interactions between novel microbes or diseases using computational methods remains a challenging task.Mainly includes:(1)The current publicly available microbe-disease association datasets have small sample sizes,and the microbe-disease association matrix is sparse,resulting in limited richness in the representation of microbes and diseases,which affects the accuracy of association prediction.Learning from a small amount of data has become a significant challenge due to the small sample sizes and sparsity of the microbe-disease association matrix.(2)Traditional matrix completion-based algorithms typically use matrix factorization methods to learn latent feature vectors of microbes and diseases,and then multiply these latent feature vectors to obtain the microbe-disease association matrix.However,this matrix completion-based approach often fails to capture the non-linear relationships between microbe and diseases,resulting in insufficient prediction accuracy.(3)In the past,some algorithms have overlooked the importance of label propagation in the inference of microbe-disease associations.They solely rely on the feature representations of microbes and diseases to predict their associations,without fully utilizing the label information for assisted prediction.Other methods fail to effectively integrate feature representations and label propagation,and separating the processes of feature representation and label propagation may significantly impact the accuracy of predictions.Therefore,integrating feature representations and label propagation is one of the important approaches to improve the accuracy of microbe-disease association predictions.Regarding the above-mentioned issue,this paper conducts research on the prediction of microbe-disease associations and key techniques,and proposes a microbe-disease association prediction algorithm based on sparse representation and few-shot learning.The main innovative achievements of this paper are as follows:Proposed a microbe-disease association prediction algorithm based on few-shot learning and neural inductive matrix completion(NIGMDA),aiming to address the issues of data imbalance and small sample size in microbe-disease association prediction.Firstly,a similarity network is constructed by integrating microbe functional similarity,disease semantic similarity,and Gaussian interaction profile kernel similarity to address the issue of insufficient semantic information contained in the feature representations of microbes and diseases.Finally,the learned low-rank feature representations are embedded into the neural inductive matrix completion model in this algorithm.In this model,the feature projection matrix used in traditional inductive matrix completion process is replaced with a non-linear neural network structure,which can learn any arbitrary function from data.This is done to capture the non-linear relationships between microbe and disease feature representations,and thereby improve the accuracy of association prediction.(2)Proposed a microbe-disease association prediction algorithm based on sparse representation and label propagation,named as GSNMDA,aiming to address the sparsity issue in the microbe-disease association matrix.Firstly,a similarity network is constructed for microbes and diseases.Then,graph autoencoders are utilized to encode the microbe and disease spaces separately,in order to obtain sparse feature representations and capture global similarity relationships.Meanwhile,the label propagation process is simulated by reconstructing the association matrix from the rating matrix.This end-to-end collaborative training of two graph autoencoders integrates the sparse representation and label propagation process,thereby enhancing the robustness and accuracy of the integration process.In addition,the model incorporates self-attention mechanism,allowing the model to adaptively handle the weight relationships between nodes in the graph structure,thereby improving the performance of the model and providing a theoretical basis and biological applications for attention-based neural networks.This paper conducted extensive experiments on microbiome-disease association datasets such as HMDAD and Disbiome,and evaluated the results using multiple metrics.The experimental results show that the proposed model in this paper outperforms some of the current state-of-the-art models in microbe-disease association prediction tasks,and reveals more potential links between microbes and diseases.Meanwhile,the reliability and predictive performance of the model were further confirmed through case studies.Therefore,the proposed model in this study has the capability to uncover potential associations between microbes and diseases,providing valuable insights for biological research.