Skip to content

Project Roadmap & Future Horizons

Project: synth-pdb
Focus: Biophysical Realism, Computational Efficiency, AI Research Support
Strategy: Weighted by Highest Reward / Lowest Risk

✅ Recently Completed

The following features have been successfully implemented and verified: - Batched Generation (GPU-First): Vectorized SIMD/GPU-ready data factory. - D-Amino Acid Support: Mirror-image peptide generation and validation. - Hard Decoy Generation: Gaussian drift and sequence threading for AI training. - Numba JIT Support: Core geometry kernels optimized for high-performance. - Cyclic Peptide Support: Head-to-tail cyclization for therapeutic modeling.

🚀 Next Horizons (Prioritized)

Sorted by High-Value / Low-Risk first.

1. ML Integration Examples 🤖 (High Value / Low Risk)

  • Concept: Provide Jupyter notebooks/scripts demonstrating zero-copy handover from BatchedPeptide to PyTorch/JAX/MLX.
  • Why: Closes the loop for AI researchers. Shows how to use the "contiguous tensor" output in real training pipelines.
  • Implementation: Create an examples/ml_loading/ directory with clear demonstrations.

2. Hardware Benchmarking Suite ⏱️ (Medium-High Value / Low Risk)

  • Concept: A formal tool to compare Serial vs. Batched performance across hardware (CPU vs. Apple Silicon M-series).
  • Why: Quantifies the efficiency gains of the "GPU-First" architecture and justifies its use for large-scale data production.
  • Implementation: New benchmarks/ module with automated performance plotting.

3. "Chromophore" Educational Case (GFP) 🌈 (Medium-High Value / Medium Risk)

  • Concept: Specialized test case for a Green Fluorescent Protein (GFP) fragment, focusing on the Ser-Tyr-Gly cyclization.
  • Why: Adds coordination chemistry and post-translational modification (PTM) education.
  • Implementation: Extend cofactors.py or create special_chemistry.py to handle the unique SYG ring closure.

🔬 Section 1: Biophysical Realism

Goal: Generate structures that are indistinguishable from experimental data.

1. Explicit Solvent "Water Box" 💧

  • Concept: Simulate the protein in a box of explicit water molecules (TIP3P model) rather than implicit solvent.
  • Use Case: Capturing "water-bridging" interactions that stabilize specific conformations.
  • Risk: Medium. Increases minimization time significantly.

2. ✅ Macrocycle Design Lab (Cyclic Peptides)

  • Result: Implemented modeling for Oxytocin (disulfide) and Cyclosporine A (cyclic/D-amino).
  • Demo: hormone_lab.ipynb
  • Value: Macrocycles are at the forefront of drug discovery. An AI model that can design cyclic peptides needs a training set of realistic closed loops, which standard PDBs lack in sufficient quantity.

🤖 Section 2: AI Research Support

Goal: Provide data that trains better models (AlphaFold, RFDiffusion, etc.).

1. Co-Evolutionary Constraints (MSA) 🧬

  • Concept: Generate a family of sequences that imply a specific fold, simulating evolution.
  • Implementation: Enhance dataset.py to generate correlated mutations based on the contact map.

2. ✅ Full Orientogram Factory (\(\omega, \theta, \phi\) angles)

  • Result: Implemented vectorized 6D inter-residue orientations with virtual GLY \(C\beta\) support.
  • Demo: orientogram_lab.ipynb

3. ✅ Bulk Dataset Factory (NPZ Pipeline)

4. ✅ Multi-Modal Transformer Training (Structure + NMR)

  • Result: Integrated synchronized structural generation with synthetic NMR observables.
  • Demo: neural_nmr_pipeline.ipynb
  • Value: Demonstrates how synth-pdb can train models to predict experimental observables (like NMR shifts) directly from sequence or 3D geometry.

5. 📉 Torsion Angle Drift & Distribution Analysis

  • Objective: Statistically analyze how adding "Drift" to Ramachandran angles affects the overall fold and the radius of gyration.
  • Key Features: --drift, dihedral_backbone, and BatchedGenerator.
  • Value: Useful for researchers testing the robustness of their GNNs (Graph Neural Networks) to small backbone perturbations.

📉 Long-Term / High Risk

  • Protein-Protein Docking: Complex multi-chain physics.
  • Metadynamics Simulation: Better suited for dedicated MD engines.
  • Fragment Assembly (Rosetta-style): Massive database dependencies.