Dataset Information

Multimodal representation learning for predicting molecule-disease relations.

ABSTRACT:

Motivation

Predicting molecule-disease indications and side effects is important for drug development and pharmacovigilance. Comprehensively mining molecule-molecule, molecule-disease and disease-disease semantic dependencies can potentially improve prediction performance.

Methods

We introduce a Multi-Modal REpresentation Mapping Approach to Predicting molecular-disease relations (M2REMAP) by incorporating clinical semantics learned from electronic health records (EHR) of 12.6 million patients. Specifically, M2REMAP first learns a multimodal molecule representation that synthesizes chemical property and clinical semantic information by mapping molecule chemicals via a deep neural network onto the clinical semantic embedding space shared by drugs, diseases and other common clinical concepts. To infer molecule-disease relations, M2REMAP combines multimodal molecule representation and disease semantic embedding to jointly infer indications and side effects.

Results

We extensively evaluate M2REMAP on molecule indications, side effects and interactions. Results show that incorporating EHR embeddings improves performance significantly, for example, attaining an improvement over the baseline models by 23.6% in PRC-AUC on indications and 23.9% on side effects. Further, M2REMAP overcomes the limitation of existing methods and effectively predicts drugs for novel diseases and emerging pathogens.

Availability and implementation

The code is available at https://github.com/celehs/M2REMAP, and prediction results are provided at https://shiny.parse-health.org/drugs-diseases-dev/.

Supplementary information

Supplementary data are available at Bioinformatics online.

SUBMITTER: Wen J

PROVIDER: S-EPMC9940625 | biostudies-literature | 2023 Feb

REPOSITORIES: biostudies-literature

ACCESS DATA

Publications

Multimodal representation learning for predicting molecule-disease relations.

Wen Jun J Zhang Xiang X Rush Everett E Panickan Vidul A VA Li Xingyu X Cai Tianrun T Zhou Doudou D Ho Yuk-Lam YL Costa Lauren L Begoli Edmon E Hong Chuan C Gaziano J Michael JM Cho Kelly K Lu Junwei J Liao Katherine P KP Zitnik Marinka M Cai Tianxi T

Bioinformatics (Oxford, England) 20230201 2

<h4>Motivation</h4>Predicting molecule-disease indications and side effects is important for drug development and pharmacovigilance. Comprehensively mining molecule-molecule, molecule-disease and disease-disease semantic dependencies can potentially improve prediction performance.<h4>Methods</h4>We introduce a Multi-Modal REpresentation Mapping Approach to Predicting molecular-disease relations (M2REMAP) by incorporating clinical semantics learned from electronic health records (EHR) of 12.6 mil ...[more]

PMID: 36805623

Similar Datasets

Project description:IntroductionAlzheimer's disease (AD) is a progressive neurodegenerative disorder. Current core cerebrospinal fluid (CSF) AD biomarkers, widely employed for diagnosis, require a lumbar puncture to be performed, making them impractical as screening tools. Considering the role of sleep disturbances in AD, recent research suggests quantitative sleep electroencephalography features as potential non-invasive biomarkers of AD pathology. However, quantitative analysis of comprehensive polysomnography (PSG) signals remains relatively understudied. PSG is a non-invasive test enabling qualitative and quantitative analysis of a wide range of parameters, offering additional insights alongside other biomarkers. Machine Learning (ML) gained interest for its ability to discern intricate patterns within complex datasets, offering promise in AD neuropathology detection. Therefore, this study aims to evaluate the effectiveness of a multimodal ML approach in predicting core AD CSF biomarkers.MethodsMild-moderate AD patients were prospectively recruited for PSG, followed by testing of CSF and blood samples for biomarkers. PSG signals underwent preprocessing to extract non-linear, time domain and frequency domain statistics quantitative features. Multiple ML algorithms were trained using four subsets of input features: clinical variables (CLINVAR), conventional PSG parameters (SLEEPVAR), quantitative PSG signal features (PSGVAR) and a combination of all subsets (ALL). Cross-validation techniques were employed to evaluate model performance and ensure generalizability. Regression models were developed to determine the most effective variable combinations for explaining variance in the biomarkers.ResultsOn 49 subjects, Gradient Boosting Regressors achieved the best results in estimating biomarkers levels, using different loss functions for each biomarker: least absolute deviation (LAD) for the Aβ42, least squares (LS) for p-tau and Huber for t-tau. The ALL subset demonstrated the lowest training errors for all three biomarkers, albeit with varying test performance. Specifically, the SLEEPVAR subset yielded the best test performance in predicting Aβ42, while the ALL subset most accurately predicted p-tau and t-tau due to the lowest test errors.ConclusionsMultimodal ML can help predict the outcome of CSF biomarkers in early AD by utilizing non-invasive and economically feasible variables. The integration of computational models into medical practice offers a promising tool for the screening of patients at risk of AD, potentially guiding clinical decisions.

Project description:MotivationAccurately predicting cancer survival is crucial for helping clinicians to plan appropriate treatments, which largely improves the life quality of cancer patients and spares the related medical costs. Recent advances in survival prediction methods suggest that integrating complementary information from different modalities, e.g. histopathological images and genomic data, plays a key role in enhancing predictive performance. Despite promising results obtained by existing multimodal methods, the disparate and heterogeneous characteristics of multimodal data cause the so-called modality gap problem, which brings in dramatically diverse modality representations in feature space. Consequently, detrimental modality gaps make it difficult for comprehensive integration of multimodal information via representation learning and therefore pose a great challenge to further improvements of cancer survival prediction.ResultsTo solve the above problems, we propose a novel method called cross-aligned multimodal representation learning (CAMR), which generates both modality-invariant and -specific representations for more accurate cancer survival prediction. Specifically, a cross-modality representation alignment learning network is introduced to reduce modality gaps by effectively learning modality-invariant representations in a common subspace, which is achieved by aligning the distributions of different modality representations through adversarial training. Besides, we adopt a cross-modality fusion module to fuse modality-invariant representations into a unified cross-modality representation for each patient. Meanwhile, CAMR learns modality-specific representations which complement modality-invariant representations and therefore provides a holistic view of the multimodal data for cancer survival prediction. Comprehensive experiment results demonstrate that CAMR can successfully narrow modality gaps and consistently yields better performance than other survival prediction methods using multimodal data.Availability and implementationCAMR is freely available at https://github.com/wxq-ustc/CAMR.Supplementary informationSupplementary data are available at Bioinformatics online.

Project description:Abstract Background Diagnosis of schizophrenia is based on a collection of symptoms which are heterogeneous from one patient to the other. Therefore, improving the reliability of this diagnosis is a currently unmet need. Schizophrenia risk is associated with genetic variation and with environmental factors potentially affecting neurodevelopment. Moreover, among the symptoms, cognitive abnormalities are heritable and predate its clinical onset. Multivariate techniques can leverage the high dimensionality of data in order to study the combined effect of multiple risk factors and symptoms on clinical predictions. The aim of the current study is therefore to assess the predictability of schizophrenia diagnosis applying machine learning techniques to an ensemble of genetic, early environmental and cognitive deficits variables. Methods 442 subjects (339 healthy controls – HC – and 103 patients with schizophrenia – SCZ) were recruited for the study. Participants underwent a full neuropsychological evaluation (Modality 1, assessment of working memory, verbal fluency, intelligence quotient, attention, speed of processing and cognitive control), a broad environmental assessment (Modality 2, investigation of urbanicity, obstetric complications, developmental anomalies, socio-economic parental status and age of parents at birth) and genome-wide genotyping (Modality 3). Following published procedures, we computed individual risk scores for each of the single nucleotide polymorphisms (SNPs) associated with risk for schizophrenia in the Psychiatric Genomics Consortium (PGC) study. Data from Modalities 1, 2 and 3 entered NeuroMiner v0.998 and underwent preprocessing procedures through scaling, pruning of non-informative variables and imputation of missing values through Euclidean distance-based nearest-neighbor search. Then, these three modalities were included in a Support Vector Machine HC vs. SCZ classification algorithm, which applied decision-based data fusion strategies to integrate the individual predictions of the three modalities in a nested cross-validation framework. Results Our cross-validated results revealed that Modality 1 (cognition) predicted schizophrenia diagnosis with the highest Balanced Accuracy (BAC, 87.3%) and that the most selected cognitive indices were intelligence quotient scores and attentional abilities. Modality 2 (environment) classified HC and SCZ with a BAC of 67.2%, and the most predictive environmental features were the parental socio-economic status, the presence of developmental anomalies during the first year of life and the age of father at birth. On the other hand, Modality 3 (genetics) predicted schizophrenia diagnosis with BAC=54,1%. The most informative SNPs were FUT9 rs117074560, TCF4 rs72934570 and STAG1 rs7432375. Decision-based fusion combining individual cognitive, environmental and genetic decision scores predicted the classification of SCZ from HC with a 78.9% BAC. Discussion Our results using a novel machine learning approach suggest that an ensemble of cognitive, early environmental and genetic features can predict schizophrenia with significant accuracy. Our results also give key information on cognitive and environmental factors that can be targeted in early identification programs and offer novel insights about genetic loci that may be prioritized in future investigations of the pathophysiology of the disease. However, the near chance-level predictive ability of the genetic modality alone calls for the implementation and testing of more complex models of interaction between multiple risk factors.

Dataset Information

Multimodal representation learning for predicting molecule-disease relations.

Motivation

Methods

Results

Availability and implementation

Supplementary information

Publications

Multimodal representation learning for predicting molecule-disease relations.

Similar Datasets

OmicsDI is part of the ELIXIR infrastructure

Tweets