Advanced Features
This document explores the advanced capabilities of synth-pdb for machine learning and complex structural generation.
1. Multiple Sequence Alignments (MSA)
For training co-evolutionary models, you can generate synthetic MSAs that maintain a consistent structure:
The --gen-msa flag uses a Markov Chain Monte Carlo (MCMC) algorithm to sample sequence variations that are biophysically compatible with the starting structure, simulating the constraints of natural selection.
2. Hard Decoy Generation
Machine learning models for structure prediction benefit from training on "hard decoys"—structures that are close to the correct fold but contain subtle errors.
The --hard flag enables threading, sequence shuffling, and torsion drift to create realistic decoy ensembles that challenge structure-quality classifiers.
3. Dataset Factory
You can generate bulk datasets for training neural networks in one command:
python -m synth_pdb.main \
--mode dataset \
--num-samples 1000 \
--min-length 10 \
--max-length 100 \
--dataset-format npz \
--train-ratio 0.8
This will generate 1000 structures, split them into training (80%) and validation (20%) sets, and save them as compressed NumPy arrays (.npz) for efficient loading in PyTorch or TensorFlow.
4. AI-Driven Interpolation
synth-pdb supports interpolating between two structures, which is useful for visualizing conformational transitions or latent space traversals:
python -m synth_pdb.main \
--mode ai \
--ai-op interpolate \
--start-pdb helix.pdb \
--end-pdb sheet.pdb \
--steps 10
This generates a series of intermediate PDB files that represent a linear (or smoothed) transition between the two endpoint structures.
5. Side-Chain Optimization
For high-resolution structures, you can perform Monte Carlo side-chain optimization:
This algorithm searches through rotamer libraries for each residue to find the combination that minimizes steric clashes, resulting in a more energetically favorable side-chain packing.