Skip to content

README.md

Latest commit

 

History

History
207 lines (149 loc) · 18.5 KB

README.md

File metadata and controls

207 lines (149 loc) · 18.5 KB
Raw

🚀 synth-pdb Incubator: Frontier Explorations

This document serves as a roadmap and brainstorming space for experimental, high-impact use cases for synth-pdb. The goal is to push the boundaries of synthetic protein generation into areas where traditional structural biology (X-ray, NMR) reaches its limits.

🏗️ 1. The Cryo-EM "Standard Candle" Simulator

The Vision: Generate atomic-resolution synthetic density maps for massive complexes to benchmark Cryo-EM refinement software.

  • The Scientific Gap: Cryo-EM is often limited by "local resolution" issues. We lack a "perfect" ground truth for how a 1.2 Å map should look for complex proteins like Apoferritin when subjected to specific noise profiles.
  • The synth-pdb Angle: We can generate thousands of slightly varied "thermal states" of a protein at tensor-speed, which can be used to simulate the "blurring" seen in Cryo-EM maps.
  • Current Status: 🧪 Prototype Stage (Need to integrate map-to-density conversion tools).

☁️ 2. IDP "Ensemble-First" Validation

The Vision: Create the first automated pipeline for generating and validating "Conformational Ensembles" for Intrinsically Disordered Proteins (IDPs).

  • The Scientific Gap: IDPs are "invisible" to X-ray and Cryo-EM. NMR provides data, but interpreting that data requires a massive ensemble of potential shapes that is computationally expensive to generate.
  • The synth-pdb Angle: Our BatchedGenerator can create 10,000+ physically plausible random-coil conformations in seconds, providing the raw material for ensemble-averaging calculations (PRE, RDC, SAXS).
  • Primary Candidate: Alpha-synuclein (N-terminal fragment)
    • Sequence: MDVFMKGLSKAKEGVVAA
    • Context: Associated with Parkinson's; no fixed structure in solution.
  • Secondary Candidate: TDP-43 C-terminal Domain
    • Sequence: GNWGNQGSGNSNYGSGGSGNDQNQSGNSNYGSGGSGNDQNQSGNSNY
    • Context: Involved in ALS; extreme low-complexity disordered region.
  • Current Status: 🛠️ Active Development (See examples/interactive_tutorials/idp_ensemble_validation.ipynb and incubator/dark_proteome/).

🌌 3. Mapping the "Dark Proteome" (AI Benchmarking)

The Vision: Use synth-pdb to generate "Hard Decoys" for AI-predicted structures that have no experimental solution.

  • The Scientific Gap: AlphaFold-3 and ESM-Fold have predicted millions of structures, but we don't know where they might be "hallucinating" (e.g., in flexible loops or orphan proteins).
  • The synth-pdb Angle: We can generate "near-native" decoys—structures that look like the AI prediction but have subtle, physically plausible changes—to test if the AI's confidence scores (pLDDT) can actually distinguish them.
  • High-Priority Target: CASP16 Target T1207 (Viral 2A protein)
    • Sequence: (Check CASP16 Target List for exact sequence)
    • Context: A "Hard" target from 2024; few prediction groups were able to solve it correctly due to its novel fold and lack of homologs.
  • Orphan Protein Candidate: Human APELA (O95782)
    • Sequence: MKPMPFHLTLLLFLLVLLSPLLLSLAPAMAGSPSFPRRSRLPDLHPCL
    • Context: Human orphan protein with no known homologs; blind spot for MSA-based AI tools.
  • Current Status: 🧪 Prototype Stage (See incubator/dark_proteome/).

🧪 4. De Novo Miniprotein Forge

The Vision: A rapid prototyping environment for synthetic biologists to "visualize" their custom-designed sequences before they ever hit the lab.

  • The Scientific Gap: Designing small, stable peptides (miniproteins) often requires heavy Rosetta simulations.
  • The synth-pdb Angle: synth-pdb provides a "lightweight" alternative to quickly generate and check the geometry (bond angles, clashes) of a new design.
  • Current Status: 💡 Concept Stage.

💊 5. Pharmacophore Clash Filter for Peptide Drug Discovery

The Vision: Use synth-pdb's distogram and clash scoring to rapidly pre-filter peptide drug candidates against a known binding pocket before any docking simulation.

  • The Scientific Gap: Physics-based docking (AutoDock, Glide) is expensive — typically 10–60 minutes per compound. Millions of peptide candidates cannot be screened this way. A fast geometric pre-filter that eliminates obvious clashes would dramatically shrink the search space.
  • The synth-pdb Angle: scoring.calculate_clash_score() and distogram.calculate_distogram() are already implemented. Combined with the docking module's write_pqr() helper, we could build a tiered funnel: geometric clash → distogram similarity to known binders → full docking. The drug_discovery_pipeline.ipynb tutorial covers late-stage docking but skips this fast pre-filter step entirely.
  • Key Insight: For peptide drugs, the dominant failure mode is steric exclusion from the binding site — exactly what clash scoring detects cheaply.
  • Graduation Path: Extend drug_discovery_pipeline.ipynb with a "pre-screen" section, then graduate to a standalone notebook.
  • Current Status: 💡 Concept Stage. All required building blocks (scoring, distogram, docking) are implemented.

🗺️ 6. Contact Map Fingerprinting for Fold Classification

The Vision: Generate a large library of synthetic structures with known folds (α, β, α/β) and train a lightweight classifier on their contact maps — creating a fast "fold fingerprint" that categorises unknown structures without sequence alignment.

  • The Scientific Gap: Sequence-based fold classification (SCOP, CATH) breaks down in the "twilight zone" (<30% identity). Structural classifiers (DALI, TM-align) are slow. A contact-map-based approach sits in the middle: structure-derived, but significantly faster.
  • The synth-pdb Angle: contact.compute_contact_map() is implemented and returns a binary matrix. BatchedGenerator can produce thousands of labelled structures (labelled by the --structure string used to generate them) in seconds. The contact matrices are naturally 2D arrays — perfect input for a CNN or simple SVM classifier. This is a clean end-to-end ML story: generate → featurise → classify.
  • Graduation Path: Working prototype graduates to examples/ml_integration/ alongside the distogram and orientogram notebooks.
  • Current Status: 🧪 Prototype Stage. contact.py and BatchedGenerator are ready; no classifier code exists yet.

🧬 7. Synthetic SAXS Curve Generator

The Vision: Compute synthetic Small-Angle X-ray Scattering (SAXS) curves from generated structures to provide ground-truth data for benchmarking SAXS analysis software (ATSAS, BioXTAS RAW).

  • The Scientific Gap: SAXS is the only technique that measures IDPs and large complexes in solution at near-physiological conditions — but software validation requires structures whose SAXS curve is "known". Experimental SAXS always has instrument noise; synthetic curves from exact coordinates would be perfectly clean.
  • The synth-pdb Angle: A SAXS curve is computed from inter-atomic distances using the Debye formula, which is a direct function of distogram.calculate_distogram(). The IDP ensemble machinery already computes pairwise distances. Adding the Debye summation is a tractable extension of existing code with no new dependencies.
  • Connects To: Incubator ideas #2 (IDP ensembles) and #1 (Cryo-EM standard candles) — SAXS curves are used alongside both techniques for cross-validation.
  • Current Status: ✅ Graduated (See synth_pdb/saxs.py and scripts/build_multimodal_dataset.py).

🏗️ 16. High-Performance Multi-Modal Dataset Builder

The Vision: Consolidate PDB generation, Cryo-EM simulation, SAXS curves, and NMR observables into a single high-throughput "Dataset Factory" for AI training.

  • The Scientific Gap: Modern deep learning models (RoseTTAFold-All-Atom, CryoCloud) require massive, synchronized datasets. Manually running multiple simulation packages for every structure is a major bottleneck.
  • The synth-pdb Angle: Our vectorized BatchedGenerator combined with the new cryo_em and saxs modules can generate thousands of labeled, multi-modal data points in minutes. A single command produces a research-ready directory with 3D volumes, 1D curves, and 2D tables, all linked by a master metadata CSV.
  • Graduation Path: scripts/build_multimodal_dataset.py is the reference implementation for this factory.
  • Current Status: ✅ Graduated (See scripts/build_multimodal_dataset.py).

🔬 8. GFP Chromophore Engineering Sandbox

The Vision: Extend the existing GFP chromophore implementation into an interactive engineering environment — mutate residues around the chromophore, regenerate the structure, and predict spectral shifts using geometric proxies.

  • The Scientific Gap: Fluorescent protein engineering requires iterative mutagenesis to shift emission wavelength. Full computational prediction requires quantum chemistry (TD-DFT). However, geometric proxies — chromophore planarity, H-bond partner count, cavity burial depth — are known spectral correlates that can be computed quickly from structure.
  • The synth-pdb Angle: special_chemistry.find_gfp_chromophore_motif() and form_gfp_chromophore() are implemented. packing.optimize_sidechains() can repack the local environment after mutation. calculate_residue_sasa() can estimate chromophore burial depth. Combining these gives a lightweight "spectral proxy" without quantum chemistry.
  • Connects To: The existing gfp_molecular_forge.ipynb tutorial — this would be a direct advanced extension of that notebook at the Advanced tier.
  • Current Status: 🧪 Prototype Stage. Core chemistry primitives are ready; the mutagenesis loop and spectral proxy scoring are not yet implemented.

🧩 9. Co-evolutionary Fitness Landscape Explorer

The Vision: Combine MSA generation (evolution, msa) with SASA burial and rotamer quality to build a "fitness landscape" heat map showing which single-point mutations are structurally tolerated vs. destabilising.

  • The Scientific Gap: Deep Mutational Scanning (DMS) experiments measure mutational fitness but require large-scale wet lab work. Computational landscapes (EVmutation, ESM-1v) exist but are opaque black boxes with no structural interpretability.
  • The synth-pdb Angle: evolution.generate_msa_sequences() provides sequence variation. burial_ratio + rotamer quality from get_quality_report() provide structural scoring. The workflow: generate all single-point mutants → score each with the defensibility scorecard → visualise as a 2D heat map (position × amino acid). This is transparent and directly connects sequence to structural consequence.
  • Scalability Note: With BatchedGenerator, scanning 20 mutations × 50 positions = 1,000 structures takes seconds, making this genuinely interactive.
  • Connects To: The protein_quality_assessment.ipynb scorecard added in this session (Step 6) would be the core scoring primitive for this explorer.
  • Current Status: 💡 Concept Stage. All components are individually implemented; no integration code exists yet.

🏥 10. Synthetic Ensemble for NMR Order Parameter Validation

The Vision: Generate ensembles of physically plausible backbone conformations with known conformational diversity and use them as synthetic ground truth to validate Model-Free (Lipari-Szabo) analysis software.

  • The Scientific Gap: Software packages like FAST, relax, and Modelfree need test cases with known S² values to validate their fitting routines. Real NMR data is noisy; a synthetic ensemble with controlled S² (via known RMSF) would be a perfect benchmark with a provably correct answer.
  • The synth-pdb Angle: relaxation.predict_order_parameters() is implemented. BatchedGenerator can produce ensembles with tunable conformational diversity (via drift and rmsd_max). A calibration loop — vary input diversity → compute S² per residue → plot vs. drift magnitude — gives a transparent S² ground-truth curve.
  • Connects To: Directly extends virtual_nmr_spectrometer.ipynb and alphafold_vs_nmr_dynamics.ipynb, providing the scientific validation backbone that both currently lack.
  • Current Status: 💡 Concept Stage.

💧 11. Explicit Solvent "Water Box" Simulator

The Vision: Surround generated structures with explicit water molecules (TIP3P model) to simulate realistic solvent-protein interactions.

  • The Scientific Gap: Implicit solvent models (GBSA) fail to capture "water-bridging" interactions—highly ordered water molecules that stabilize specific loops or binding sites.
  • The synth-pdb Angle: Our core generation logic is fast; the bottleneck is the physics of thousands of water molecules. By implementing a "sparse" water box (only hydrating the surface or active site), we can maintain high performance while improving biophysical realism.
  • Current Status: 💡 Concept Stage.

🏗️ 12. Heterogeneous BMRB Dataset Integration

The Vision: Extend the ML dataset factory to produce mixed training sets containing both perfectly clean synthetic structures and noisy, experimental BMRB/PDB data pairs.

  • The Scientific Gap: AI models trained only on synthetic data often fail on real, noisy experimental data.
  • The synth-pdb Angle: synth_nmr already has a BMRB downloader. We can bridge this into dataset.py to create "Heterogeneous Batches"—half synthetic, half experimental—to improve the generalization of downstream GNNs and transformers.
  • Current Status: 💡 Concept Stage.

🔄 13. Ensemble-Average NMR observables (RDC & RPF)

The Vision: Shift from single-structure predictions to ensemble-averaged RDCs and RPF scores, reflecting the dynamic nature of proteins in solution.

  • The Scientific Gap: Proteins are not static. NMR observables like RDCs are an average across all states in an ensemble. Standard predictors only look at a single PDB file.
  • The synth-pdb Angle: Since we can generate ensembles at scale, we can compute observables for each member and average them (optionally weighted by energy). This provides a "Virtual NMR" experience that matches the complexity of real lab data.
  • Current Status: 🛠️ Active Development (RDC shim implemented; RPF integration in progress).

📉 14. Interactive Q-Factor Validation Panel

The Vision: Add a real-time "Q-factor" agreement score to the interactive quality tutorials, teaching users how to measure structural "truth" against RDC data.

  • The Scientific Gap: Understanding RDC agreement (Q-factor) is mathematically non-trivial for many students.
  • The synth-pdb Angle: Extend the protein_quality_assessment.ipynb tutorial to show how Q-factor changes as you "drift" a structure away from its native fold. It provides a visual, interactive bridge between raw coordinates and experimental validation metrics.
  • Current Status: 🧪 Prototype Stage (Linked to Vision #13).

🏗️ 15. The "Structural Interactome" & Complex Validation

The Vision: Transform synth-pdb from a single-chain generator into a world-class validator for protein-protein assemblies and the "Structural Interactome."

  • The Scientific Gap: While single-protein folding (AlphaFold-2) is largely solved, the current frontier is modeling how proteins interact (the "Interactome"). Tools like AlphaFold-Multimer and ESM-Fold-Multimer often produce structures that look plausible individually but contain subtle physical errors at the interface (e.g., steric clashes, poor buried surface area, or missing ionic complementarity).
  • The synth-pdb Angle: We are implementing rigorous Interface Analytics to measure Buried Surface Area (BSA), inter-chain steric clashes, and electrostatic complementarity (inter-chain salt bridges). By providing a fast, physics-based "scorecard" for multi-chain assemblies, we allow researchers to rapidly benchmark the physical fidelity of AI-predicted complexes.
  • Key Reference: Levy (2010, JMB) established that BSA and the hydrophobic effect at the interface are the most reliable predictors of biological relevance.
  • Current Status: 🧪 Prototype Stage (Interface Metrics implemented in PDBValidator).

☣️ 17. Pathological Stress Tests & Physical Limits

The Vision: Use extreme, non-biological, or metamorphic sequences to test the absolute boundaries of protein physics and AI prediction.

  • The "Lead Pipe": Poly-Tryptophan (20x W)
    • Sequence: WWWWWWWWWWWWWWWWWWWW
    • Goal: Force into a helix to observe maximal steric clashing and energy frustration.
  • The "Shape-Shifter": Metamorphic XCL1
    • Sequence: VGSEVSDKRTCVSLTTQRLPVSRIKTYTITEGSLRAVIFITKRGLKVCADPQATWVRDVVRSMDRKSNTRNNMIQTKPTGTQQSTNTAVTLTG
    • Goal: Demonstrate a sequence with two valid, competing folds.
  • The "Viral Ghost": Pandoravirus ORFan
    • Sequence: MASTSRRLLARLLRRRLLRRLRRRLLARLLRRRLLRRLRRRLLARLLRRR
    • Goal: Explore the stability of giant virus sequences that have no evolutionary signal.
  • The "Anti-Hydrophobic" Monster
    • Sequence: KKKKKKKKKKKKKKKKKKKK (Poly-Lysine)
    • Goal: Create a compact globular state that is 100% polar, resulting in a Hydrophobic Burial Ratio of zero.
  • Current Status: 🧪 Prototype Stage (Anti-Hydrophobic and Cryo-EM tests in progress).

📝 How to Contribute an Exploration

  1. Draft a Vision: Define the "What If?".
  2. Define the Gap: Why can't traditional tools do this easily?
  3. Code a Prototype: Add a script in a subfolder within /incubator/.
  4. Validate & Graduate: If the idea proves stable and useful, it "graduates" to the core examples/ directory.

🧬 18. NCAA Support: Expanding the Chemical Alphabet

The Vision: Extend synth-pdb beyond the 20 standard L-amino acids to support Non-Canonical Amino Acids (NCAAs) for peptidomimetics and drug discovery.

  • The Scientific Gap: Most therapeutic peptides and many biological systems (post-translational modifications, expanded genetic codes) use residues that aren't in the standard 20. Current tools often require complex manual setup for each new residue.
  • The synth-pdb Angle: By generalizing the NeRF backbone logic and implementing a custom residue template loader, we can generate structures with arbitrary chemistry at high speed.
  • Current Status: 💡 Concept Stage (See incubator/ncaa_roadmap.md).