Single-cell transcriptomics (scRNA-seq) is revolutionizing biological research, yet it faces challenges such as inefficient transcript capture and noise. To address these challenges, methods like neighbor averaging or graph diffusion are used. These methods often rely on k-nearest neighbor graphs from low-dimensional manifolds. However, scRNA-seq data suffer from the ‘curse of dimensionality’, leading to the over-smoothing of data when using imputation methods. To overcome this, sc-PHENIX employs a PCA-UMAP diffusion method, which enhances the preservation of data structures and allows for a refined use of PCA dimensions and diffusion parameters (e.g., k-nearest neighbors, exponentiation of the Markov matrix) to minimize noise introduction. This approach enables a more accurate construction of the exponentiated Markov matrix (cell neighborhood graph), surpassing methods like MAGIC. sc-PHENIX significantly mitigates over-smoothing, as validated through various scRNA-seq datasets, demonstrating improved cell phenotype representation. Applied to a multicellular tumor spheroid dataset, sc-PHENIX identified known extreme phenotype states, showcasing its effectiveness. sc-PHENIX is open-source and available for use and modification.

Early diagnosis and preventive measures are necessary to mitigate diseases’ impact on the yield of Colocasia esculenta (Taro). This study addresses the challenges of Taro disease identification by employing two key strategies: integrating explainable artificial intelligence techniques to interpret deep learning models and conducting a comparative analysis of advanced architectures Inception V3, ResNet-50, and Vision Transformers for classifying common Taro diseases, including leaf blight and mosaic virus, as well as identifying healthy leaves. The novelty of this work lies in the first-ever integration of SHapley Additive exPlanations (SHAP) with deep learning architectures to enhance model interpretability while providing a comprehensive comparison of state-of-the-art methods for this underexplored crop. The proposed models significantly improve the ability to recognize complex patterns and features, achieving high accuracy and robust performance in disease classification. The model’s efficacy was evaluated through multi-class statistical metrics, including accuracy, precision, F1 score, recall, specificity, Chohen’s kappa, and area under the curve. Among the architectures, Inception V3 exhibited superior performance in accuracy (0.9985), F1 score (0.9985), recall (0.9985), and specificity (0.9992). The explainability of Inception V3 was further enhanced using SHAP, which provides insights by dissecting the contributions of individual features in Taro leaves to the model’s predictions. This approach facilitates a deeper understanding of the disease classification process and supports the development of effective disease management strategies, ultimately contributing to improved Taro cultivation practices.

Single-cell transcriptomics (scRNA-seq) is becoming a technology that is transforming biological discovery in many fields of medicine. Despite its impact in many areas, scRNASeq is technologically and experimentally limited by the inefficient transcript capture and the high rise of noise sources. For that reason, imputation methods were designed to denoise and recover missing values. Many imputation methods (e.g., neighbor averaging or graph diffusion) rely on k nearest neighbor graph construction derived from a mathematical space as a low-dimensional manifold. Nevertheless, the construction of mathematical spaces could be misleading the representation of densities of the distinct cell phenotypes due to the negative effects of the curse of dimensionality. In this work, we demonstrated that the imputation of data through diffusion approach on PCA space favor over-smoothing when increases the dimension of PCA and the diffusion parameters, such k-NN ( k-nearest neighbors ) and t (value of the exponentiation of the Markov matrix) parameters. In this case, the diffusion on PCA space distorts the cell neighborhood captured in the Markovian matrix creating an artifact by connecting densities of distinct cell phenotypes, even though these are not related phenotypically. In this situation, over-smoothing of data is due to the fact of shared information among spurious cell neighbors. Therefore, it can not account for more information on the variability (from principal components) or nearest neighbors for a well construction of a cell-neighborhood. To solve above mentioned issues, we propose a new approach called sc-PHENIX( s ingle c ell- PHE notype recovery by N on-linear I mputation of gene e X pression ) which uses PCA-UMAP initialization for revealing new insights into the recovered gene expression that are masked by diffusion on PCA space. sc-PHENIX is an open free algorithm whose code and some examples are shown at https://github.com/resendislab/sc-PHENIX .

Motivation: Microbiota data suffers from technical noise (reflected as excess of zeros in the count matrix) and the curse of dimensionality. This complicates downstream data analysis and compromises the scientific discovery’s reliability. Data sparsity makes it difficult to obtain a well-cluster structure and distorts the abundance distributions. Currently, there is a rised need to develop new algorithms with improved capacities to reduce noise and recover missing information. Results: We present mb-PHENIX, an open-source algorithm developed in Python, that recovers taxa abundances from the noisy and sparse microbiota data. Our method deals with sparsity in the count matrix (in 16S microbiota and shotgun studies) by applying imputation via diffusion onto the supervised Uniform Manifold Approximation Projection (sUMAP) space. Our hybrid machine learning approach allows the user to denoise microbiota data. Thus, the differential abundance of microbes is more accurate among study groups, where abundance analysis fails. Availability: The mb-PHENIX algorithm is available at https://github.com/resendislab/mb-PHENIX . An easy-to-use implementation is available on Google Colab (see GitHub) Contact: Oresendis@inmegen.gob.mx Supplementary information: Supplementary data are available at Bioinformatics online.

Today, the intratumoral composition is a relevant factor associated with the progression and aggression of cancer. Although it suggests a metabolic interdependence among the subpopulations inside the tumor, a detailed map of how this interdependence contributes to the malignant phenotype is still lacking. To address this issue, we developed a systems biology approach integrating single-cell RNASeq and genome-scale metabolic reconstruction to map the metabolic cross-feeding among the subpopulations previously identified in the spheroids of MCF7 breast cancer. By calibrating our model with expression profiles and the experimental growth rate, we concluded that the reverse Warburg effect emerges as a mechanism to optimize community growth. Furthermore, through an in silico analysis, we identified lactate, alpha-ketoglutarate, and some amino acids as key metabolites whose disponibility alters the growth rate of the spheroid. Altogether, this work provides a strategy for assessing how space and intratumoral heterogeneity influence the metabolic robustness of cancer, issues suggesting that computational strategies should move toward the design of optimized treatments.

The COVID-19 pandemic, caused by SARS-CoV-2, has led to significant challenges worldwide, including diverse clinical outcomes and prolonged post-recovery symptoms known as Long COVID or Post-COVID-19 syndrome. Emerging evidence suggests a crucial role of metabolic reprogramming in the infection’s long-term consequences. This study employs a novel approach utilizing machine learning (ML) and explainable artificial intelligence (XAI) to analyze metabolic alterations in COVID-19 and Post-COVID-19 patients. By integrating ML with SHAP (SHapley Additive exPlanations) values, we aimed to uncover metabolomic signatures and identify potential biomarkers for these conditions. Our analysis included a cohort of 142 COVID-19, 48 Post-COVID-19 samples and 38 CONTROL patients, with 111 identified metabolites. Traditional analysis methods like PCA and PLS-DA were compared with advanced ML techniques to discern metabolic changes. Notably, XGBoost models, enhanced by SHAP for explainability, outperformed traditional methods, demonstrating superior predictive performance and providing different insights into the metabolic basis of the disease’s progression and its aftermath, the analysis revealed several metabolomic subgroups within the COVID-19 and Post-COVID-19 conditions, suggesting heterogeneous metabolic responses to the infection and its long-term impacts. This study highlights the potential of integrating ML and XAI in metabolomics research.