Abstract Background Coronary heart disease (CHD) is a major cause of morbidity and mortality worldwide. Identifying key risk factors is essential for effective risk assessment and prevention. A data-driven approach using machine learning (ML) offers advanced techniques to analyze complex, nonlinear, and high-dimensional datasets, uncovering novel predictors of CHD that go beyond the limitations of traditional models, which rely on predefined variables. Objective This study aims to evaluate the contribution of various risk factors to CHD, focusing on both established and novel markers using ML techniques. Methods The study recruited 7672 participants aged 30-84 years from Suita City, Japan, between 1989 and 1999. Over an average of 15 years, participants were monitored for cardiovascular events. A total of 7260 participants and 28 variables were included in the analysis after excluding individuals with missing outcome data and eliminating unnecessary variables. Five ML models—logistic regression, random forest (RF), support vector machine, Extreme Gradient Boosting, and Light Gradient-Boosting Machine—were applied for predicting CHD incidence. Model performance was evaluated using accuracy, sensitivity, specificity, precision, area under the curve, F₁-score, calibration curves, observed-to-expected ratios, and decision curve analysis. Additionally, Shapley Additive Explanations (SHAPs) were used to interpret the prediction models and understand the contribution of various risk factors to CHD. Results Among 7260 participants, 305 (4.2%) were diagnosed with CHD. The RF model demonstrated the highest performance, with an accuracy of 0.73 (95% CI 0.64‐0.80), sensitivity of 0.74 (95% CI 0.62‐0.84), specificity of 0.72 (95% CI 0.61‐0.83), and an area under the curve of 0.73 (95% CI 0.65‐0.80). RF also showed excellent calibration, with predicted probabilities closely aligning with observed outcomes, and provided substantial net benefit across a range of risk thresholds, as demonstrated by decision curve analysis. SHAP analysis elucidated key predictors of CHD, including the intima-media thickness (IMT_cMax) of the common carotid artery, blood pressure, lipid profiles (non–high-density lipoprotein cholesterol, high-density lipoprotein cholesterol, and triglycerides), and estimated glomerular filtration rate. Novel risk factors identified as significant contributors to CHD risk included lower calcium levels, elevated white blood cell counts, and body fat percentage. Furthermore, a protective effect was observed in women, suggesting the potential necessity for gender-specific risk assessment strategies in future cardiovascular health evaluations. Conclusions We developed a model to predict CHD using ML and applied SHAP methods for interpretation. This approach highlights the multifactor nature of CHD risk evaluation, aiming to support health care professionals in identifying risk factors and formulating effective prevention strategies.

Aims Health food products (HFPs) are foods and products related to maintaining and promoting health. HFPs may sometimes cause unforeseen adverse health effects by interacting with drugs. Considering the importance of information on the interactions between HFPs and drugs, this study aimed to establish a workflow to extract information on Drug‐HFP Interactions (DHIs) from open resources. Methods First, Information on drugs, enzymes, their interactions, and known DHIs was collected from multiple public databases and literature sources. Next, a network consisted of enzymes, HFP, and drugs was constructed, assuming enzymes as candidates for hubs in Drug‐HFP interactions (Method 1). Furthermore, we developed methods to analyze the biomedical context of each drug and HFP to predict potential DHIs out of the DHIs obtained in Method 1 by applying BioWordVec, a widely used biomedical terminology quantifier (Method 2‐1 and 2‐2). Results 44,965 DHIs (30% known) were identified in Method 1, including 38 metabolic enzymes, 157 HFPs, and 1256 drugs. Method 2‐1 selected 7401 DHIs (17% known) from the DHIs of Method 1, while Method 2‐2 chose 2819 DHIs (30% known). Based on the different assumptions in these methods where Method 2‐1 specifically selects HFPs interacting with specific enzymes and Method 2‐2 specifically selects HFPs with similar function with drugs, the propsed methods resulted in extracting a wide variety of DHIs. Conclusions By integrating the results of language processing techniques with those of the network analysis, a workflow to efficiently extract unknown and known DHIs was constructed.

Abstract Motivation Enzymes are key targets to biosynthesize functional substances in metabolic engineering. Therefore, various machine learning models have been developed to predict Enzyme Commission numbers, one of the enzyme annotations. However, the previously reported models predict the sequences with numerous consecutive identical amino acids, which are found within unannotated sequences, as enzymes. Results Here, we propose EnzymeNet for prediction of complete Enzyme Commission numbers using Residual Neural Networks. EnzymeNet can exclude the exceptional sequences described above. Several EnzymeNet models were built and optimized to explore the best conditions for removing such sequences. As a result, the models exhibited higher prediction accuracy with Macro F1 score up to 0.850 than previously reported models. Moreover, even the enzyme sequences with low similarity to training data, which were difficult to predict using the reported models, could be predicted extensively using EnzymeNet models. The robustness of EnzymeNet models will lead to discover novel enzymes for biosynthesis of functional compounds using microorganisms. Availability The source code of EnzymeNet models is freely available at and https://github.com/nwatanbe/enzymenet. Supplementary information Supplementary data are available at Bioinformatics Advances online.

The number of unannotated protein sequences is explosively increasing due to genome sequence technology. A more comprehensive understanding of protein functions for protein annotation requires the discovery of new features that cannot be captured from conventional methods. Deep learning can extract important features from input data and predict protein functions based on the features. Here, protein feature vectors generated by 3 deep learning models are analyzed using Integrated Gradients to explore important features of amino acid sites. As a case study, prediction and feature extraction models for UbiD enzymes were built using these models. The important amino acid residues extracted from the models were different from secondary structures, conserved regions and active sites of known UbiD information. Interestingly, the different amino acid residues within UbiD sequences were regarded as important factors depending on the type of models and sequences. The Transformer models focused on more specific regions than the other models. These results suggest that each deep learning model understands protein features with different aspects from existing knowledge and has the potential to discover new laws of protein functions. This study will help to extract new protein features for the other protein annotations.

Omics data was acquired, and the development and research of metabolic simulation and analysis methods using them were also actively carried out. However, it was a laborious task to acquire such data each time the medium composition, culture conditions, and target organism changed. Therefore, in this study, we aimed to extract and estimate important variables and necessary numbers for predicting metabolic flux distribution as the state of cell metabolism by flux sampling using a genome-scale metabolic model (GSM) and its analysis. Acetic acid production from glucose in Escherichia coli with GSM iJO1366 was used as a case study. Flux sampling obtained by OptGP using 1000 pattern constraints on substrate, product, and growth fluxes produced a wider sample than the default case. The analysis also suggested that the fluxes of iron ions, O2, CO2, and NH4+, were important for predicting the metabolic flux distribution. Additionally, the comparison with the literature value of 13C-MFA using CO2 emission flux as an example of an important flux suggested that the important flux obtained by this method was valid for the prediction of flux distribution. In this way, the method of this research was useful for extracting variables that were important for predicting flux distribution, and as a result, the possibility of contributing to the reduction of measurement variables in experiments was suggested.

Abstract Engineering the microbial production of secondary metabolites is limited by the known reactions of correctly annotated enzymes. Therefore, the machine learning discovery of specialized enzymes offers great potential to expand the range of biosynthesis pathways. Benzylisoquinoline alkaloid production is a model example of metabolic engineering with potential to revolutionize the paradigm of sustainable biomanufacturing. Existing bacterial studies utilize a norlaudanosoline pathway, whereas plants contain a more stable norcoclaurine pathway, which is exploited in yeast. However, committed aromatic precursors are still produced using microbial enzymes that remain elusive in plants, and additional downstream missing links remain hidden within highly duplicated plant gene families. In the current study, machine learning is applied to predict and select plant missing link enzymes from homologous candidate sequences. Metabolomics-based characterization of the selected sequences reveals potential aromatic acetaldehyde synthases and phenylpyruvate decarboxylases in reconstructed plant gene-only benzylisoquinoline alkaloid pathways from tyrosine. Synergistic application of the aryl acetaldehyde producing enzymes results in enhanced benzylisoquinoline alkaloid production through hybrid norcoclaurine and norlaudanosoline pathways.