PhenoRescue AI is a multi-modal artificial intelligence platform designed to accelerate phenotypic drug discovery for TP53 cancer mutations. By integrating cutting-edge generative AI models from NVIDIA with structural biology and high-content imaging data, PhenoRescue AI moves beyond traditional structure-based docking to predict a compound's ability to restore healthy cellular function.
This platform provides a comprehensive approach to identifying and prioritizing therapeutic candidates for specific TP53 mutations:
- Multi-Modal Prediction: Predicts compound efficacy by combining:
- Structural Layer: Assesses if a compound physically binds to the mutated p53 protein.
- Phenotypic Layer: Predicts the cellular consequences of binding by analyzing high-content imaging data (Cell Painting) to identify "rescued" cellular phenotypes.
- Unified Dashboard: A web-based interface for:
- Inputting TP53 mutations.
- Visualizing mutant protein structures.
- Viewing a ranked list of compounds with combined structural and phenotypic scores.
- Displaying predicted morphological profiles of rescued cells.
- Novel Compound Generation: Utilizes generative chemistry models to design and synthesize novel molecules tailored to rescue specific phenotypes.
- Actionable Prioritization: Delivers a prioritized shortlist of compounds supported by both structural binding evidence and predicted phenotypic rescue, guiding experimental validation.
- "TechBio" Approach: Focuses on the complex interplay between molecular interactions and cellular biology, offering a more holistic view of drug efficacy.
PhenoRescue AI leverages a robust blend of state-of-the-art AI, scientific computing, and web technologies.
- TP53 Variant Annotations: NCI TP53 Database, IARC TP53 Database.
- CRISPR Gene Effect Scores: DepMap Portal (
Achilles_gene_effect_CERES.csv). - Cell Painting Morphological Profiles: JUMP Cell Painting Consortium dataset (AWS S3 bucket
s3://cellpainting-gallery/cpg0016-jump/). - Cell Line Metadata: DepMap (
model_list.csv).
PhenoRescue AI integrates powerful NVIDIA generative AI models:
- GenMol (Generative Chemistry): Used for de-novo generation of novel chemical scaffolds and fragment-based molecule generation.
- MolMIM (Controlled Generation): Optimizes generated molecules for specific desired properties (e.g., binding affinity, solubility).
- DiffDock (Molecular Docking): Predicts 3D protein-ligand binding poses, including an "All-atom DiffDock Pocket" for detailed structural analysis.
The BioNeMo platform provides the underlying infrastructure for integrating and deploying the NVIDIA models:
- NVIDIA NIMs: Optimized, deployable microservices for GenMol, MolMIM, and DiffDock.
- BioNeMo Agent Toolkit: Facilitates the creation of AI agents to automate complex workflows, from compound generation and docking to phenotypic prediction and reporting.
- Backend: Python, FastAPI, Docker
- Frontend: TypeScript, React, Material UI
- ML Frameworks: PyTorch, Scikit-learn, XGBoost
- Scientific Computing: RDKit, ESMFold, AlphaFold
- Data Management: Firebase, Supabase
The platform is designed with a modular architecture, allowing for flexible component integration and refinement.
graph TD
A[User Input: TP53 Mutation] --> B(Frontend: React + Material UI);
B --> C{Backend API: FastAPI};
subgraph "NVIDIA NIMs (GPU-Accelerated)"
D[GenMol: Generate Novel Compounds]
E[MolMIM: Optimize for Properties]
F[DiffDock: Predict Binding Pose]
end
subgraph "Existing Pipelines"
G[ESMFold/AlphaFold: Mutant Structure]
H[Physics-Based Docking: Lennard-Jones, Coulomb]
I[RDKit: Cheminformatics]
end
subgraph "Phenotypic Prediction"
J[Cell Painting Profile Database]
K[Trained Classifier: Phenotype → Response]
end
C --> D;
C --> E;
C --> F;
C --> G;
C --> H;
C --> I;
C --> J;
C --> K;
D --> F;
E --> F;
G --> H;
H --> F;
F --> K;
K --> L[Phenotypic Rescue Score];
L --> M[Output: Ranked Compound Shortlist];
M --> B;