📅 Day 1: Target Structural Backbone Generation with RFdiffusion
Core Objective: Master score-based diffusion models to generate novel protein backbones, binder scaffolds, and target-conditioned functional sites.
- Principles of generative denoising diffusion probabilistic models (DDPMs) in 3D protein structure space
- De novo scaffold generation and motif scaffolding for functional site presentation
- Target-conditioned diffusion: setting hotspot residues, binding pockets, and symmetry constraints
- Generating diverse candidates and evaluating structural backbone quality
🛠️
Hands-on Lab: Setup and execute an RFdiffusion pipeline to generate a library of novel binder backbones targeted to a designated protein receptor pocket using Google Colab.
🧰
Tools Covered: RFdiffusion, PyMOL, Biopython, Google Colab
📅 Day 2: Sequence Generation & Inverse Folding with ProteinMPNN
Core Objective: Translate target structural backbones into physical amino acid sequences engineered for high-affinity target binding.
- Fundamentals of message-passing graph neural networks (GNNs) in sequence inverse folding
- Fixed-backbone sequence design and residue-level conditioning (temperature, fixed positions)
- Optimizing amino acid sequence recovery, solubility, and aggregation propensity
- Automating multi-sequence sampling per generated backbone scaffold
🛠️
Hands-on Lab: Pass generated RFdiffusion backbones into ProteinMPNN to sample candidate sequences, apply sequence-level filters, and prepare FASTA inputs using Google Colab.
🧰
Tools Covered: ProteinMPNN, ANARCI, FastA Utilities, Google Colab
📅 Day 3: Biomolecular Complex Validation with AlphaFold 3 & Boltz-1
Core Objective: Predict complex interactions, score binding interfaces, and perform final candidate selection using next-generation AI architectures.
- Next-generation structure prediction: from AlphaFold 2 Evoformer to AlphaFold 3 Pairformer & diffusion modules
- Modeling biomolecular complexes: protein-protein interactions, nucleic acids, and small-molecule ligands
- Evaluating interface confidence metrics: pLDDT, PAE (Predicted Aligned Error), and ipTM scores
- Lead candidate ranking and automated export for wet-lab gene synthesis and expression
🛠️
Hands-on Lab: Predict target-binder complex structures using open-source AlphaFold 3 implementations (Boltz-1), filter leads by ipTM/pLDDT thresholds, and export validated designs using Google Colab.
🧰
Tools Covered: AlphaFold 3 Server, Boltz-1, ESMFold, Google Colab
