ToxiVerse: A Public Platform for Chemical Toxicity Data Sharing and Customizable Predictive Modeling

ToxiVerse is a freely accessible, user-friendly web platform that integrates curated toxicity datasets, chemical bioprofiling, and automated machine learning tools to enable researchers without programming expertise to predict chemical toxicity and support drug development and environmental safety.

The Gist

Imagine you are a chef trying to figure out if a new, mysterious spice is safe to eat. In the old days, you might have had to feed it to a bunch of animals to see what happened. That takes a long time, costs a lot of money, and raises ethical questions.

Today, scientists use computers to predict toxicity (how poisonous something is) instead. But here's the problem: most of these computer tools are like high-end, professional kitchens. They are powerful, but you need to be a master chef (a computer programmer) to even turn them on. If you don't know how to code, you're stuck.

ToxiVerse is the solution. Think of it as a user-friendly, all-in-one "Toxicity Kitchen" that anyone can walk into, no matter their cooking skills. It's a free website designed to help researchers, students, and regulators check if chemicals are safe without needing to write a single line of code.

Here is how ToxiVerse works, broken down into its three main "stations":

1. The Bioprofiler: The "Super-Sniffer"

Imagine you have a new chemical, but you don't know much about it. The Bioprofiler is like a super-sniffer dog that goes to the world's biggest library of chemical tests (called PubChem) to see if this chemical has ever been tested before.

• The Problem: Sometimes the library has gaps. Maybe the chemical was tested on 100 things, but not on the 101st thing you care about.
• The ToxiVerse Fix: The system uses a "smart guesser" (Machine Learning). If it sees that Chemical A and Chemical B look very similar, and Chemical A failed a specific test, the system guesses that Chemical B will probably fail it too. It fills in the missing blanks so you get a complete picture of the chemical's behavior.

2. The Database: The "Organized Filing Cabinet"

Most scientific data is messy. It's like a library where books are thrown on the floor, written in different languages, and some pages are torn out.

• The ToxiVerse Fix: The Database module is a super-organized filing cabinet. The team has cleaned up about 50,000 chemicals, making sure every file is labeled correctly, the data is consistent, and the "pages" aren't missing.
• What you get: You can browse through this cabinet to find data on specific dangers, like "Will this hurt the liver?" or "Is this cancer-causing?" You can download these clean files with a single click, ready to use.

3. The Cheminformatics Module: The "DIY Prediction Workshop"

This is the most powerful part. Usually, you have to go to a factory and buy a pre-made prediction machine. If you want to test a new type of chemical, the factory won't help you.

• The ToxiVerse Fix: This module is a DIY workshop. You can bring your own raw ingredients (your own data files) or grab some from the filing cabinet.
• The Process:
  1. Clean: You drop your data in, and the system automatically washes off the dirt (fixes chemical errors).
  2. Visualize: It draws a 3D map showing where your chemicals sit in relation to each other, like a constellation map.
  3. Build: You tell the system, "I want to predict if these chemicals are toxic." You click a button, and the system builds a custom prediction model for your specific data.
  4. Predict: You can then feed new chemicals into this custom model to see if they are safe.

Why is this a big deal?

Before ToxiVerse, if you wanted to build a custom safety model, you needed a PhD in computer science and a team of programmers. ToxiVerse removes the "code barrier."

• It's flexible: You aren't stuck with pre-made answers; you can build your own.
• It's smart: It combines chemical structure with biological activity (how the chemical actually behaves in a living system).
• It's accessible: It's free, runs in your web browser, and comes with a tutorial so you can learn how to use it in minutes.

In short: ToxiVerse takes the complex, messy world of chemical safety testing and turns it into a simple, drag-and-drop experience. It empowers anyone to ask, "Is this chemical safe?" and get a scientifically backed answer without needing to be a coding wizard.

Technical Summary

1. Problem Statement

Chemical toxicity assessment is critical for drug development and environmental safety but currently faces significant bottlenecks:

• Limitations of Current Methods: Regulatory frameworks rely heavily on animal testing, which is time-consuming, costly, and ethically controversial.
• Data Quality Issues: Public toxicity databases often suffer from inconsistent annotations, missing metadata, non-standardized formats, and species discrepancies, leading to bias in modeling.
• Accessibility Barriers: Existing computational tools (e.g., ADMETlab, ProTox, VEGA-QSAR) often rely on pre-trained models with fixed endpoints. They lack flexibility for users to input custom datasets, build new models, or perform batch processing. Furthermore, many tools require programming expertise, excluding researchers without computational backgrounds.
• The Gap: There is a lack of user-friendly, web-based platforms that integrate curated data, automated bioprofiling, and customizable machine learning (ML) modeling without requiring coding skills.

2. Methodology and Platform Architecture

ToxiVerse is a modular, web-based platform developed using Flask (Python 3.11) and deployed in a Docker environment. It utilizes SQLite for data storage, RDKit for cheminformatics, and scikit-learn for machine learning. The platform consists of three integrated modules:

A. Bioprofiler Module

• Function: Generates comprehensive chemical descriptors by combining experimental bioactivity data with ML-based imputation.
• Data Source: Leverages a local SQLite database derived from PubChem (bioactivities and compound-to-InChIKey mappings), updated every six months.
• Workflow:
  1. Initial Bioprofile: Constructs a chemical-bioactivity matrix where active/inactive outcomes are encoded.
  2. Assay Selection: Uses Mutual Information (MI) scores to rank assays based on their ability to distinguish overall active vs. inactive chemicals.
  3. Gap Filling: For chemicals with missing data in selected assays, Random Forest (RF) classifiers are trained using ECFP6 fingerprints (radius 3, 2048 bits) to predict outcomes.
  4. Output: Produces a complete bioprofile (hybrid descriptors) that integrates structural and biological information, enabling mechanistic read-across.

B. Database Module

• Content: Hosts approximately 50,000 curated unique chemicals covering over 50 toxicity endpoints (e.g., hepatotoxicity, carcinogenicity, developmental toxicity).
• Data Curation: Data is sourced from previous studies and public resources, standardized with unique ToxiVerse-IDs, canonical SMILES, and PubChem CIDs.
• Features:
  • Visualization: Interactive histograms of activity distributions and Principal Component Analysis (PCA) plots to visualize chemical space.
  • Bioassay Ranking: Identifies relevant PubChem bioassays for specific endpoints using a Bayesian smoothing approach to adjust active rates, ranking assays by their relevance to the target toxicity.

C. Cheminformatics Module

• Function: Enables end-to-end QSAR modeling for custom or curated datasets.
• Data Ingestion: Supports CSV/SDF uploads or direct retrieval from PubChem via Assay IDs (AIDs).
• Structure Curation: Uses the ChEMBL Structure Pipeline (via RDKit) to clean, standardize, and validate structures (handling stereochemistry, valence, salts, etc.).
• Modeling:
  • Descriptors: Calculates RDKit descriptors, ECFP6, and FCFP6 fingerprints.
  • Algorithms: Supports Random Forest (RF), Support Vector Machines (SVM), and k-Nearest Neighbors (k-NN) for both classification and regression.
  • Optimization: Hyperparameters are tuned via grid search with 5-fold cross-validation.
• Prediction: Users can upload new chemicals (CSV/SDF/SMILES) to predict toxicity using trained models. Results are downloadable as CSV/SDF.

3. Key Contributions

1. User-Friendly Accessibility: A no-code, web-based interface that allows researchers without programming expertise to perform complex tasks like data curation, bioprofiling, and QSAR modeling.
2. Customizable Modeling: Unlike tools restricted to pre-trained models, ToxiVerse allows users to build custom models on their own datasets or curated ToxiVerse data.
3. Hybrid Bioprofiling: The unique integration of PubChem high-throughput screening (HTS) data with ML-based gap filling creates "biologically enriched" descriptors, overcoming the sparsity of public bioactivity data.
4. Comprehensive Data Curation: A rigorously curated database of ~50,000 chemicals with standardized formats and linked bioassays, addressing issues of data inconsistency found in other public repositories.
5. Batch Processing: Supports the processing of large datasets and the generation of standardized modeling reports.

4. Results and Validation

• Implementation: The platform is fully functional and publicly accessible at www.toxiverse.com.
• Performance:
  • The Bioprofiler successfully generates complete bioprofiles for target chemicals by imputing missing assay data using RF models trained on ECFP6 fingerprints.
  • In a case study using an estrogen receptor antagonist dataset (888 compounds), the system successfully built and evaluated nine classification models (combinations of descriptors and algorithms).
  • Performance metrics (Accuracy, AUC, F1-score, Precision, Recall, ROC AUC for classification; R², MSE, MAPE for regression) are automatically generated and visualized.
• Visualization: The platform successfully renders 3D PCA plots and heatmaps (e.g., a heatmap of 1,622 assays for 6,543 compounds), allowing users to explore chemical clusters and activity distributions.

5. Significance

ToxiVerse represents a significant advancement in computational toxicology by bridging the gap between data availability and usability.

• Democratization: It empowers a broader range of researchers (including those in toxicology and biology without coding skills) to utilize advanced ML for chemical risk assessment.
• Regulatory Impact: By providing a transparent, customizable, and reproducible platform, it supports the "3Rs" (Replacement, Reduction, Refinement) of animal testing by facilitating robust in silico alternatives.
• Scientific Utility: The ability to combine structural data with imputed biological activity data offers a more holistic view of chemical toxicity, potentially improving the accuracy of predictive models and facilitating the discovery of mechanisms of action.
• Future-Proofing: The modular design and integration with public databases (PubChem) ensure the platform remains current and adaptable to new toxicity endpoints and data sources.