AI-Powered Structural Biology: AlphaFold-Based Protein Prediction, Validation & Biological Discovery

Aim

Workshop Objectives

Workshop Structure

Day 1 : AI Protein Structure Prediction: From Model Generation to Confidence

Core Topics

• Evolution of AI-driven protein structure prediction
• AlphaFold, RoseTTAFold and protein language model-based approaches
• Experimental structures vs AI-predicted structures
• PDB and AlphaFold DB for structural retrieval
• Understanding pLDDT, PAE and residue-level confidence
• Local vs global structural confidence
• Why a high-confidence prediction is not automatically a biologically correct model
• Common limitations: intrinsically disordered regions, flexible loops, multi-domain proteins and unusual conformations
• Introduction to trustworthy structural AI and model validation

🛠️ Hands-on : AI Structure Confidence Analysis

Google Colab + PDB + AlphaFold DB + Mol*

Participants will:

1. Select a protein with both experimental and AlphaFold structures.
2. Retrieve the structures from PDB and AlphaFold DB.
3. Visualize the predicted structure and confidence information.
4. Generate a pLDDT profile and PAE heatmap using a prepared Colab notebook.
5. Identify high- and low-confidence regions.
6. Compare the confidence pattern with the experimentally determined structure.

Skill gained: Reading AI structural confidence instead of treating an AI prediction as ground truth.

Day 2 : Structural Validation, Uncertainty & Explainable Protein AI

Core Topics

• What makes an AI-generated protein structure trustworthy?
• Structural validation beyond confidence scores
• RMSD, TM-score and residue-level structural deviation
• Experimental-vs-predicted structure benchmarking
• Detecting local structural errors
• Uncertainty quantification and confidence assessment
• AI failure modes in protein structure prediction
• Explainable AI (XAI) for biological models
• Interpreting model outputs rather than simply reporting prediction scores
• Reproducibility and transparent computational workflows

🛠️ Hands-on : Experimental vs AI Structure Benchmarking

Google Colab + PDB + AlphaFold DB + TM-align

Participants will:

1. Load an experimental PDB structure and corresponding AI prediction.
2. Align the two structures using a prepared notebook/workflow.
3. Calculate RMSD and TM-score.
4. Generate a per-residue deviation plot.
5. Locate regions with the largest structural differences.
6. Compare those regions against Day 1 pLDDT/PAE results.
7. Build a simple interpretation:
   High confidence + low deviation → stronger structural agreement
   Low confidence + high deviation → region requiring caution

Skill gained: Quantitatively validating an AI prediction and connecting uncertainty with structural error.

Day 3 : Biological Interpretation & Trustworthy AI for Drug Discovery

Core Topics

• From predicted structure to biological interpretation
• AI-assisted protein function prediction
• Sequence–structure–function relationships
• Protein language models and structure-aware representations
• Functional domains and conserved regions
• UniProt, InterPro and Pfam annotation
• Mapping biological annotations onto predicted structures
• Using structural AI for:
  • therapeutic target discovery
  • protein engineering
  • functional annotation
  • drug discovery
• Trustworthy AI considerations for pharmaceutical and biomedical applications
• FAIR and reproducible computational workflows
• Emerging directions: protein language models, multimodal protein AI, structure-aware models and foundation models

These topics align well with recent MDPI research trends, including protein language models, multimodal structure/function prediction, structure-aware representations and AI-assisted protein function prediction.

🛠️ Hands-on : Build a Protein Trustworthiness Report

Google Colab + UniProt + InterPro + Pfam + AlphaFold DB

Participants will:

1. Enter a UniProt protein ID into a prepared Colab notebook.
2. Retrieve the corresponding AlphaFold structure.
3. Automatically obtain functional/domain annotations.
4. Map conserved domains onto the predicted structure.
5. Combine:
   • structural confidence
   • structural validation
   • functional annotation
   • domain information
6. Generate a one-page protein AI assessment report.
7. Produce a simple Green / Amber / Red trustworthiness summary based on predefined criteria.

Skill gained: Moving from AI prediction → validation → biological interpretation → evidence-based confidence assessment.

Who Should Enrol?

Important Dates

Workshop Outcomes