CXML - ChemXploreML

Appearance

Dimensionality Reduction ​

High-dimensional molecular embeddings, while rich in information, are difficult to visualize and interpret directly. The Dimensionality Reduction module in ChemXploreML provides a suite of powerful algorithms to project these embeddings into a lower-dimensional space (typically 2D or 3D), making it possible to visualize the chemical space, identify clusters, and understand the relationships between molecules.

Overview ​

Dimensionality Reduction Overview

Workflow ​

The process of applying dimensionality reduction involves the following steps:

  1. Select Embedder: Choose the molecular embedding that you want to reduce. These are the embeddings you generated in the "Embed Molecule" section.
  2. Choose a Method: Select a dimensionality reduction algorithm from the available tabs. Each method has its own strengths and is suited for different types of analysis.
  3. Configure Parameters: Adjust the parameters for the selected algorithm. You can hover over each parameter to get a description of what it does.
  4. Save Parameters: Before running the reduction, you must save your parameter configuration. This creates a reusable settings file and ensures your workflow is reproducible.
  5. Run Reduction: Click the "Run" button to start the dimensionality reduction process. The application will save the resulting low-dimensional embeddings as a new .npy file.
  6. Visualize: The reduced embeddings can then be used for visualization and further analysis.

Supported Methods ​

Below is a detailed description of each supported dimensionality reduction method and its configurable parameters.

PCA (Principal Component Analysis) ​

  • Description: A linear technique that transforms the data into a new coordinate system, where the axes (principal components) are orthogonal and capture the maximum variance in the data.
  • Local Structure: ⚠️ Moderate
  • Global Structure: ✅ Excellent
  • Scalability: ✅ Fast

Parameters:

  • n_components: The number of principal components to keep. Default: 70.
  • random_state: Seed for the random number generator. Default: 42.

UMAP (Uniform Manifold Approximation and Projection) ​

  • Description: A non-linear technique excellent for preserving both local and some global structure. It is often used for visualizing clusters in high-dimensional data.
  • Local Structure: ✅ Excellent
  • Global Structure: ⚠️ Moderate
  • Scalability: ✅ Fast

Parameters:

  • n_neighbors: Controls how UMAP balances local versus global structure. Default: 15.
  • min_dist: The minimum distance between embedded points. Default: 0.1.
  • n_components: The dimension of the space to embed into. Default: 2.
  • metric: The distance metric to use. Default: 'euclidean'.
  • random_state: Seed for the random number generator. Default: 42.

t-SNE (t-Distributed Stochastic Neighbor Embedding) ​

  • Description: A non-linear technique that is particularly good at revealing the underlying cluster structure in the data. It excels at preserving local structure.
  • Local Structure: ✅ Excellent
  • Global Structure: ❌ Poor
  • Scalability: ❌ Slow

Parameters:

  • n_components: The dimension of the space to embed into. Default: 2.
  • perplexity: Related to the number of nearest neighbors considered for each point. Default: 30.
  • random_state: Seed for the random number generator. Default: 42.

KernelPCA ​

  • Description: An extension of PCA that uses kernel methods to perform non-linear dimensionality reduction.
  • Local Structure: ✅ Good
  • Global Structure: ✅ Good
  • Scalability: ⚠️ Slower

Parameters:

  • n_components: Number of components to keep. Default: 2.
  • kernel: The kernel function to use ('linear', 'poly', 'rbf', 'sigmoid', 'cosine'). Default: 'rbf'.
  • gamma: Kernel coefficient. Default: null.

PHATE (Potential of Heat-diffusion for Affinity-based Trajectory Embedding) ​

  • Description: A visualization method that captures both local and global nonlinear structure using a heat-diffusion-based affinity metric. It is particularly well-suited for visualizing trajectories and progressions in data.
  • Local Structure: ✅ Good
  • Global Structure: ✅ Excellent
  • Scalability: ⚠️ Medium

Parameters:

  • n_components: Number of dimensions to reduce to. Default: 2.
  • knn: Number of nearest neighbors. Default: 5.
  • decay: Controls the decay of the kernel. Default: 40.
  • t: Diffusion time scale. Default: 'auto'.
  • random_state: Seed for reproducibility. Default: 42.

ISOMAP ​

  • Description: A non-linear technique that preserves the geodesic distances between points on a manifold.
  • Local Structure: ✅ Excellent
  • Global Structure: ✅ Good
  • Scalability: ⚠️ Medium

Parameters:

  • n_components: Number of coordinates for the manifold. Default: 2.
  • n_neighbors: Number of neighbors to consider for each point. Default: 5.

Laplacian Eigenmaps ​

  • Description: A spectral method that preserves local manifold information by constructing a graph from the data and embedding it in a lower-dimensional space.
  • Local Structure: ✅ Excellent
  • Global Structure: ❌ Poor
  • Scalability: ✅ Fast

Parameters:

  • n_components: Dimension of the embedding space. Default: 2.
  • n_neighbors: Number of neighbors for constructing the neighborhood graph. Default: 10.

TriMap ​

  • Description: A non-linear method that uses triplet constraints to preserve both local and global structure in the data.
  • Local Structure: ✅ Excellent
  • Global Structure: ✅ Good
  • Scalability: ✅ Fast

Parameters:

  • n_dims: Output dimensionality. Default: 2.
  • n_inliers: Number of inlier points per triplet. Default: 10.
  • n_outliers: Number of outlier points per triplet. Default: 5.
  • n_random: Number of random triplets per point. Default: 5.
  • distance: The distance metric to use. Default: 'euclidean'.

Factor Analysis ​

  • Description: A linear statistical method used to describe variability among observed, correlated variables in terms of a potentially lower number of unobserved variables called factors.
  • Local Structure: ⚠️ Limited
  • Global Structure: ✅ Good
  • Scalability: ✅ Fast

Parameters:

  • n_components: Number of latent factors to extract. Default: 2.

Next Steps ​

After reducing the dimensionality of your embeddings, you can:

  • Use the 2D or 3D embeddings to create scatter plots and visualize your chemical space.
  • Use the reduced-dimension embeddings as features for training a machine learning model.