UMAP-Based split

Training data, models, and python code for the manuscript: Machine learning models and performance dependency on 2D chemical descriptor space for retention time prediction of pharmaceuticals.

MOE descriptors of the METLIN SMRT dataset (original by Domingo-Almenara et. al. with RTs and structures available at: https://figshare.com/ndownloader/files/18130628), training scripts (python), UMAP, GMM, and SVR models with training splits results are within the SMRT_exp.zip file.

CSVs for feature importance for SVR models are standalone files.

Whole-brain functional connectivity (FC) measured with functional MRI (fMRI) evolves over time in meaningful ways at temporal scales going from years (e.g., development) to seconds [e.g., within-scan time-varying FC (tvFC)]. Yet, our ability to explore tvFC is severely constrained by its large dimensionality (several thousands). To overcome this difficulty, researchers often seek to generate low dimensional representations (e.g., 2D and 3D scatter plots) hoping those will retain important aspects of the data (e.g., relationships to behavior and disease progression). Limited prior empirical work suggests that manifold learning techniques (MLTs)—namely those seeking to infer a low dimensional non-linear surface (i.e., the manifold) where most of the data lies—are good candidates for accomplishing this task. Here we explore this possibility in detail. First, we discuss why one should expect tvFC data to lie on a low dimensional manifold. Second, we estimate what is the intrinsic dimension (ID; i.e., minimum number of latent dimensions) of tvFC data manifolds. Third, we describe the inner workings of three state-of-the-art MLTs: Laplacian Eigenmaps (LEs), T-distributed Stochastic Neighbor Embedding (T-SNE), and Uniform Manifold Approximation and Projection (UMAP). For each method, we empirically evaluate its ability to generate neuro-biologically meaningful representations of tvFC data, as well as their robustness against hyper-parameter selection. Our results show that tvFC data has an ID that ranges between 4 and 26, and that ID varies significantly between rest and task states. We also show how all three methods can effectively capture subject identity and task being performed: UMAP and T-SNE can capture these two levels of detail concurrently, but LE could only capture one at a time. We observed substantial variability in embedding quality across MLTs, and within-MLT as a function of hyper-parameter selection. To help alleviate this issue, we provide heuristics that can inform future studies. Finally, we also demonstrate the importance of feature normalization when combining data across subjects and the role that temporal autocorrelation plays in the application of MLTs to tvFC data. Overall, we conclude that while MLTs can be useful to generate summary views of labeled tvFC data, their application to unlabeled data such as resting-state remains challenging.

This entry contains the data used to implement the bachelor thesis. It was investigated how embeddings can be used to analyze supersecondary structures. Abstract of the thesis: This thesis analyzes the behavior of supersecondary structures in the context of embeddings. For this purpose, data from the Protein Topology Graph Library was provided with embeddings. This resulted in a structured graph database, which will be used for future work and analyses. In addition, different projections were made into the two-dimensional space to analyze how the embeddings behave there. In the Jupyter Notebook 1_data_retrival.ipynb the download process of the graph files from the Protein Topology Graph Library (https://ptgl.uni-frankfurt.de) can be found. The downloaded .gml files can also be found in graph_files.zip. These form graphs that represent the relationships of supersecondary structures in the proteins. These form the data basis for further analyses. These graph files are then processed in the Jupyter Notebook 2_data_storage_and_embeddings.ipynb and entered into a graph database. The sequences of the supersecondary and secondary structures from the PTGL can be found in fastas.zip. The embeddings were also calculated using the ESM model of the Facebook Research Group (huggingface.co/facebook/esm2_t12_35M_UR50D), which can be found in three .h5 files. These are then added there subsequently. The whole process in this notebook serves to build up the database, which can then be searched using Cypher querys. In the Jupyter Notebook 3_data_science.ipynb different visualizations and analyses are then carried out, which were made with the help of UMAP. For the installation of all dependencies, it is recommended to create a Conda environment and then install all packages there. To use the project, PyEED should be installed using the snapshot of the original repository (source repository: https://github.com/PyEED/pyeed). The best way to install PyEED is to execute the pip install -e . command in the pyeed_BT folder. The dependencies can also be installed by using poetry and the .toml file. In addition, seaborn, h5py and umap-learn are required. These can be installed using the following commands: pip install h5py==3.12.1 pip install seaborn==0.13.2 umap-learn==0.5.7

Passive Acoustic Monitoring (PAM) is emerging as a solution for monitoring species and environmental change over large spatial and temporal scales. However, drawing rigorous conclusions based on acoustic recordings is challenging, as there is no consensus over which approaches, and indices are best suited for characterizing marine and terrestrial acoustic environments. Here, we describe the application of multiple machine-learning techniques to the analysis of a large PAM dataset. We combine pre-trained acoustic classification models (VGGish, NOAA & Google Humpback Whale Detector), dimensionality reduction (UMAP), and balanced random forest algorithms to demonstrate how machine-learned acoustic features capture different aspects of the marine environment. The UMAP dimensions derived from VGGish acoustic features exhibited good performance in separating marine mammal vocalizations according to species and locations. RF models trained on the acoustic features performed well for labelled sounds in the 8 kHz range, however, low and high-frequency sounds could not be classified using this approach. The workflow presented here shows how acoustic feature extraction, visualization, and analysis allow for establishing a link between ecologically relevant information and PAM recordings at multiple scales. The datasets and scripts provided in this repository allow replicating the results presented in the publication. Methods Data acquisition and preparation We collected all records available in the Watkins Marine Mammal Database website listed under the “all cuts'' page. For each audio file in the WMD the associated metadata included a label for the sound sources present in the recording (biological, anthropogenic, and environmental), as well as information related to the location and date of recording. To minimize the presence of unwanted sounds in the samples, we only retained audio files with a single source listed in the metadata. We then labelled the selected audio clips according to taxonomic group (Odontocetae, Mysticetae), and species. We limited the analysis to 12 marine mammal species by discarding data when a species: had less than 60 s of audio available, had a vocal repertoire extending beyond the resolution of the acoustic classification model (VGGish), or was recorded in a single country. To determine if a species was suited for analysis using VGGish, we inspected the Mel-spectrograms of 3-s audio samples and only retained species with vocalizations that could be captured in the Mel-spectrogram (Appendix S1). The vocalizations of species that produce very low frequency, or very high frequency were not captured by the Mel-spectrogram, thus we removed them from the analysis. To ensure that records included the vocalizations of multiple individuals for each species, we only considered species with records from two or more different countries. Lastly, to avoid overrepresentation of sperm whale vocalizations, we excluded 30,000 sperm whale recordings collected in the Dominican Republic. The resulting dataset consisted in 19,682 audio clips with a duration of 960 milliseconds each (0.96 s) (Table 1). The Placentia Bay Database (PBD) includes recordings collected by Fisheries and Oceans Canada in Placentia Bay (Newfoundland, Canada), in 2019. The dataset consisted of two months of continuous recordings (1230 hours), starting on July 1st, 2019, and ending on August 31st 2029. The data was collected using an AMAR G4 hydrophone (sensitivity: -165.02 dB re 1V/µPa at 250 Hz) deployed at 64 m of depth. The hydrophone was set to operate following 15 min cycles, with the first 60 s sampled at 512 kHz, and the remaining 14 min sampled at 64 kHz. For the purpose of this study, we limited the analysis to the 64 kHz recordings. Acoustic feature extraction The audio files from the WMD and PBD databases were used as input for VGGish (Abu-El-Haija et al., 2016; Chung et al., 2018), a CNN developed and trained to perform general acoustic classification. VGGish was trained on the Youtube8M dataset, containing more than two million user-labelled audio-video files. Rather than focusing on the final output of the model (i.e., the assigned labels), here the model was used as a feature extractor (Sethi et al., 2020). VGGish converts audio input into a semantically meaningful vector consisting of 128 features. The model returns features at multiple resolution: ~1 s (960 ms); ~5 s (4800 ms); ~1 min (59’520 ms); ~5 min (299’520 ms). All of the visualizations and results pertaining to the WMD were prepared using the finest feature resolution of ~1 s. The visualizations and results pertaining to the PBD were prepared using the ~5 s features for the humpback whale detection example, and were then averaged to an interval of 30 min in order to match the temporal resolution of the environmental measures available for the area. UMAP ordination and visualization UMAP is a non-linear dimensionality reduction algorithm based on the concept of topological data analysis which, unlike other dimensionality reduction techniques (e.g., tSNE), preserves both the local and global structure of multivariate datasets (McInnes et al., 2018). To allow for data visualization and to reduce the 128 features to two dimensions for further analysis, we applied Uniform Manifold Approximation and Projection (UMAP) to both datasets and inspected the resulting plots. The UMAP algorithm generates a low-dimensional representation of a multivariate dataset while maintaining the relationships between points in the global dataset structure (i.e., the 128 features extracted from VGGish). Each point in a UMAP plot in this paper represents an audio sample with duration of ~ 1 second (WMD dataset), ~ 5 seconds (PBD dataset, humpback whale detections), or 30 minutes (PBD dataset, environmental variables). Each point in the two-dimensional UMAP space also represents a vector of 128 VGGish features. The nearer two points are in the plot space, the nearer the two points are in the 128-dimensional space, and thus the distance between two points in UMAP reflects the degree of similarity between two audio samples in our datasets. Areas with a high density of samples in UMAP space should, therefore, contain sounds with similar characteristics, and such similarity should decrease with increasing point distance. Previous studies illustrated how VGGish and UMAP can be applied to the analysis of terrestrial acoustic datasets (Heath et al., 2021; Sethi et al., 2020). The visualizations and classification trials presented here illustrate how the two techniques (VGGish and UMAP) can be used together for marine ecoacoustics analysis. UMAP visualizations were prepared the umap-learn package for Python programming language (version 3.10). All UMAP visualizations presented in this study were generated using the algorithm’s default parameters.
Labelling sound sources The labels for the WMD records (i.e., taxonomic group, species, location) were obtained from the database metadata. For the PBD recordings, we obtained measures of wind speed, surface temperature, and current speed from (Fig 1) an oceanographic buy located in proximity of the recorder. We choose these three variables for their different contributions to background noise in marine environments. Wind speed contributes to underwater background noise at multiple frequencies, ranging 500 Hz to 20 kHz (Hildebrand et al., 2021). Sea surface temperature contributes to background noise at frequencies between 63 Hz and 125 Hz (Ainslie et al., 2021), while ocean currents contribute to ambient noise at frequencies below 50 Hz (Han et al., 2021) Prior to analysis, we categorized the environmental variables and assigned the categories as labels to the acoustic features (Table 2). Humpback whale vocalizations in the PBD recordings were processed using the humpback whale acoustic detector created by NOAA and Google (Allen et al., 2021), providing a model score for every ~5 s sample. This model was trained on a large dataset (14 years and 13 locations) using humpback whale recordings annotated by experts (Allen et al., 2021). The model returns scores ranging from 0 to 1 indicating the confidence in the predicted humpback whale presence. We used the results of this detection model to label the PBD samples according to presence of humpback whale vocalizations. To verify the model results, we inspected all audio files that contained a 5 s sample with a model score higher than 0.9 for the month of July. If the presence of a humpback whale was confirmed, we labelled the segment as a model detection. We labelled any additional humpback whale vocalization present in the inspected audio files as a visual detection, while we labelled other sources and background noise samples as absences. In total, we labelled 4.6 hours of recordings. We reserved the recordings collected in August to test the precision of the final predictive model. Label prediction performance We used Balanced Random Forest models (BRF) provided in the imbalanced-learn python package (Lemaître et al., 2017) to predict humpback whale presence and environmental conditions from the acoustic features generated by VGGish. We choose BRF as the algorithm as it is suited for datasets characterized by class imbalance. The BRF algorithm performs under sampling of the majority class prior to prediction, allowing to overcome class imbalance (Lemaître et al., 2017). For each model run, the PBD dataset was split into training (80%) and testing (20%) sets. The training datasets were used to fine-tune the models though a nested k-fold cross validation approach with ten-folds in the outer loop, and five-folds in the inner loop. We selected nested cross validation as it allows optimizing model hyperparameters and performing model evaluation in a single step. We used the default parameters of the BRF algorithm, except for the ‘n_estimators’ hyperparameter, for which we tested

and error (error_*).

This archive contains the restricted 10 ensemble benchmark and the scripts used in the manifold learning techniques assessment. Files related to an ensemble are prefixed with ID1_ID2_, where ID1 is the first member in alphabetical order, and ID2 is the reference for the structural alignment.

The archive includes the following for each member of the benchmark: A _mm.pdb file containing the ensemble's conformations. A _aln.fa file, which is the multiple sequence alignment of the ensemble. A _rmsd.txt file with the all pairwise root mean squared deviation (RMSD) of the ensemble. A _raw_coords_ca.bin file with the raw coordinates in binary format. A _raw_coords_ca_mask.bin file with the binary format gap coordinates. A _features_pca.csv file detailing the positions of each sample in the ensemble's principal component space. A _dist_to_hull.csv file with the ID of each ensemble member, their label in the clustering in the PC space, and the squared distance of this sample to the convex hull formed by members of the other clusters. A _pca_errors.csv file containing the same information as the _dist_to_hull.csv file, but with the addition of the PCA reconstruction error, measured as the RMSD between the predicted and ground truth structures. The prediction of a sample is done by fitting the PCA to all clusters except the one being evaluated. Three _XXX_kcpa_errors.json files with the kPCA reconstruction errors for each ensemble member, measured as the RMSD between the predicted and ground truth structures, using kPCA at different sigma and alpha parameters from the grid search. The XXX indicates the kernel used. The prediction of a sample is done by fitting the kPCA to all clusters except the one being evaluated. A _umap_errors.json file with the UMAP reconstruction errors for each ensemble member, measured as the RMSD between the predicted and ground truth structures, using UMAP at different n_neigh and min_dist parameters from the grid search. The prediction of a sample is done by fitting the UMAP to all clusters except the one being evaluated. UMAP could be run only on a subset of the ensembles. A _rbf_kpca_default_sigma.json file containing the kPCA reconstruction errors for each ensemble member, measured as the RMSD between the predicted and ground truth structures, using kPCA with RBF kernel at the default alpha and sigma parameters. The prediction of a sample is done by fitting the kPCA to all clusters except the one being evaluated. A _rbf_kpca_errors_real.json file with the kPCA reconstruction errors for each ensemble member, measured as the RMSD between the predicted and ground truth structures, using kPCA with RBF kernel with a predicted optimal sigma parameter and alpha parameters of 1.0, 1e-5, and 1e-6. The prediction of a sample is done by fitting the kPCA to all clusters except the one being evaluated. The scripts used to generate the convex hull and for the PCA-kPCA comparison are as follows: dist_to_hull.py computes the coordinates in the PC space of each member, divides the members into clusters, and computes the distance of each member to the convex hull formed by members of the other clusters in the PC space. This script uses polytope_app.cpp with a Python binding to compute the squared distance of each member to the convex hull. polytope_module.so is the compiled C++ module called by the Python script. interpol_apase.py computes the interpolation in the ATPase latent space, and outputs the .pdb files of the trajectories. pca_kpca.py calculates the reconstruction error for both PCA, kPCA, and UMAP for each ensemble member by fitting the PCA, kPCA, or UMAP to all members of other clusters, excluding the cluster of the member currently being evaluated. A procheck folder containing summary tables of the procheck analysis on original and reconstructed structures. The stats.csv file contains descriptive information about the benchmark. Please consult the related documentation to understand the meaning of each column in this file.

We introduce the GAMMA (Galactic Attributes of Mass, Metallicity, and Age) dataset, a comprehensive collection of galaxy data tailored for Machine Learning applications. This dataset offers detailed 2D maps and 3D cubes of 11 727 galaxies, capturing essential attributes: stellar age, metallicity, and mass. Together with the dataset we publish our code to extract any other stellar or gaseous property from the raw simulation suite to extend the dataset beyond these initial properties, ensuring versatility for various computational tasks. Ideal for feature extraction, clustering, and regression tasks, GAMMA offers a unique lens for exploring galactic structures through computational methods and is a bridge between astrophysical simulations and the field of scientific machine learning (ML). As a first benchmark, we apply Principal Component Analysis (PCA) on this dataset. We find that PCA effectively captures the key morphological features of galaxies with a small number of components. We achieve a dimensionality reduction by a factor of ∼200 (∼3650) for 2D images (3D cubes) with a reconstruction accuracy below 5%. We calculate UMAP (Uniform Manifold Approximation and Projection for Dimension Reduction) on the lower dimensional PCA scores of the 2D images to visualize the image space. An interactive version of this plot can be accessed using an online Dashboard (hover over a point to see the galaxy image and the IllustrisTNG Subhalo ID). All the code to generate this dataset and load the data structure is publicly available on GitHub, with an additional documentation page hosted on ReadTheDocs.

Single-cell RNA sequencing (scRNA-seq) is a high-throughput sequencing technology performed at the level of an individual cell, which can have a potential to understand cellular heterogeneity. However, scRNA-seq data are high-dimensional, noisy, and sparse data. Dimension reduction is an important step in downstream analysis of scRNA-seq. Therefore, several dimension reduction methods have been developed. We developed a strategy to evaluate the stability, accuracy, and computing cost of 10 dimensionality reduction methods using 30 simulation datasets and five real datasets. Additionally, we investigated the sensitivity of all the methods to hyperparameter tuning and gave users appropriate suggestions. We found that t-distributed stochastic neighbor embedding (t-SNE) yielded the best overall performance with the highest accuracy and computing cost. Meanwhile, uniform manifold approximation and projection (UMAP) exhibited the highest stability, as well as moderate accuracy and the second highest computing cost. UMAP well preserves the original cohesion and separation of cell populations. In addition, it is worth noting that users need to set the hyperparameters according to the specific situation before using the dimensionality reduction methods based on non-linear model and neural network.

Comparison of machine-learning methods by different measurements for CyTOF Dataset 1 (13 biomarkers, 24 labeled cell types).

Interactive UMAP plot of the French Polynesia recordings.

Two CyTOF benchmark data sets for analysis.

Comparison of methods for averaging performance in the identification of known cell types in training and testing data by different measurements for CyTOF1 and CyTOF2 datasets.

Contingency table for calculating the receiver operating characteristic curve.

Ischemic stroke’s complex pathophysiology demands therapeutic approaches targeting multiple pathways simultaneously, yet current treatments remain limited. We developed an innovative drug discovery pipeline combining a deep learning approach with experimental validation to identify natural compounds with comprehensive neuroprotective properties. Our computational framework integrated SELFormer, a transformer-based deep learning model, and multiple deep learning algorithms to predict NC bioactivity against seven crucial stroke-related targets (ACE, GLA, MMP9, NPFFR2, PDE4D, and eNOS). The pipeline encompassed IC50 predictions, clustering analysis, quantitative structure–activity relationship (QSAR) modeling, and uniform manifold approximation and projection (UMAP)-based bioactivity profiling followed by molecular docking studies and experimental validation. Analysis revealed six distinct NC clusters with unique molecular signatures. UMAP projection identified 11 medium-activity (6 < pIC50 ≤ 7) and 57 high-activity (pIC50 > 7) compounds, with molecular docking confirming strong correlations between binding energies and predicted pIC50 values. In vitro studies using NGF-differentiated PC12 cells under oxygen-glucose deprivation demonstrated significant neuroprotective effects of four high-activity compounds: feruloyl glucose, l-hydroxy-l-tryptophan, mulberrin, and ellagic acid. These compounds enhanced cell viability, reduced acetylcholinesterase activity and lipid peroxidation, suppressed TNF-α expression, and upregulated BDNF mRNA levels. Notably, mulberrin and ellagic acid showed superior efficacy in modulating oxidative stress, inflammation, and neurotrophic signaling. This study establishes a robust deep learning-driven framework for identifying multitarget natural therapeutics for ischemic stroke. The validated compounds, particularly mulberrin and ellagic acid, are promising for stroke treatment development. Our findings demonstrate the effectiveness of integrating computational prediction with experimental validation in accelerating drug discovery for complex neurological disorders.

Statistical report containing details of all data pre-processing steps to create the dataset for primary decision-makers, as well as initial data visualisation using UMAP.

Viruses have evolved the ability to bind and enter cells through interactions with a wide variety of cell macromolecules. We engineered peptide-modified adeno-associated virus (AAV) capsids that transduce the brain through the introduction of de novo interactions with 2 proteins expressed on the mouse blood–brain barrier (BBB), LY6A or LY6C1. The in vivo tropisms of these capsids are predictable as they are dependent on the cell- and strain-specific expression of their target protein. This approach generated hundreds of capsids with dramatically enhanced central nervous system (CNS) tropisms within a single round of screening in vitro and secondary validation in vivo thereby reducing the use of animals in comparison to conventional multi-round in vivo selections. The reproducible and quantitative data derived via this method enabled both saturation mutagenesis and machine learning (ML)-guided exploration of the capsid sequence space. Notably, during our validation process, we determined that nearly all published AAV capsids that were selected for their ability to cross the BBB in mice leverage either the LY6A or LY6C1 protein, which are not present in primates. This work demonstrates that AAV capsids can be directly targeted to specific proteins to generate potent gene delivery vectors with known mechanisms of action and predictable tropisms.

End stage renal disease (ESRD) engenders detrimental effects in the Immune system, manifested as quantitative alterations of lymphocyte subpopulations, akin, albeit not identical to those observed during the ageing process. We performed dimensionality reduction of an extended lymphocyte phenotype panel of senescent and exhaustion related markers in ESRD patients and controls with Principal Component Analysis (PCA) and Uniform Manifold Approximation and Projection (UMAP). The plane defined by the first two principal components of PCA showed two fuzzy clusters, for patients and controls, respectively, with loadings of non-senescent markers pointing towards the controls’ centroid. Naive lymphocytes were reduced in ESRD patients compared to controls (CD4+CD45RA+CCR7+ 200(150-328) vs. 426(260-585cells/μl respectively, P = 0.001, CD19+IgD+CD27- 54(26-85) vs. 130(83-262)cells/μl respectively, P < 0.001). PCA projections of the multidimensional ESRD immune phenotype suggested a more senescent phenotype in hemodialysis compared to hemodiafiltration treated patients. Lastly, clustering based on UMAP revealed three distinct patient groups, exhibiting gradual changes for naive, senescent, and exhausted lymphocyte markers. Machine learning algorithms can distinguish ESRD patients from controls, based on their immune-phenotypes and also, unveil distinct immunological groups within patients’ cohort, determined possibly by dialysis prescription.

Accuracy of classification to individuals using SVM based on S-UMAP reduced data.