Skip to content

README.md

Latest commit

 

History

History
132 lines (96 loc) · 4.53 KB

README.md

File metadata and controls

132 lines (96 loc) · 4.53 KB
Raw

Bayesian Optimization

This directory contains the Gaussian-process and candidate-selection pieces used to rank cyclodextrin variants from precomputed molecular descriptors and MD-derived binding data.

What Is Here

Core code

  • morganKernel.py Defines additive Gaussian-process models built with GPyTorch and BoTorch. The models split the feature vector into a "primary" block and a "secondary" block, then apply separate Matern kernels and add them together.
  • training.py Contains train_model(...), a small training loop for exact GP models using Adam and ExactMarginalLogLikelihood.
  • get_candidates_delta.ipynb Notebook for interactive analysis and candidate selection. This is the main place to combine the stored data objects with the GP models and inspect predicted rankings.
  • README Older placeholder note. The more complete directory documentation is in this file.

Data subdirectory

data/ holds the serialized inputs used by the notebook and training code.

  • data/training_data.pkl Training-ready data object used to fit the GP model.
  • data/md_data.pkl MD-derived measurements used to build or validate the training set.
  • data/chem_space.pkl Full screened chemical space with descriptors and metadata.
  • data/candidates.pkl Candidate subset selected from the larger chemical space.
  • data/dG_data_pfoa.pkl Free-energy dataset for PFOA-related runs.
  • data/dG_data_pfoa_2.pkl Alternate or updated PFOA free-energy dataset.
  • data/chem_space_structures/ PDB files for the screened chemical-space structures. Filenames are numeric IDs such as 00000.pdb.
  • data/prim_cleaved_structs/ Reference host structures with primary-side groups removed: acd_prim_removed.pdb, bcd_prim_removed.pdb, and gcd_prim_removed.pdb.
  • data/labels/ Label assets used by the data-generation workflow.
  • data/README.md Short note indicating that chem_space.pkl is generated from a fingerprint-generation script.

How The Code Fits Together

The intended flow is:

  1. Prepare the descriptor and MD data objects in data/.
  2. Load them in get_candidates_delta.ipynb.
  3. Instantiate one of the additive GP models from morganKernel.py.
  4. Train the model with train_model(...) from training.py.
  5. Score the full chemical space in data/chem_space.pkl.
  6. Save or inspect shortlisted candidates, for example in data/candidates.pkl.

The model classes assume the input feature vector is concatenated in a fixed order:

  • the first primary_dim columns correspond to the primary substituent features
  • the next secondary_dim columns correspond to the secondary substituent features

If you change the descriptor layout, you must update those dimensions when building the model.

How To Use It

1. Install the Python dependencies

At minimum, the code in this directory expects:

  • torch
  • gpytorch
  • botorch
  • jupyter for the notebook workflow

You will also need whatever packages were used to create the pickle files in data/.

2. Open the notebook workflow

From the repository root:

cd bayesianoptimization
jupyter notebook get_candidates_delta.ipynb

Use the notebook when you want the full interactive workflow: loading the pickles, fitting the GP, and exploring candidate rankings.

3. Reuse the GP model in a script

Minimal training example:

import torch
import gpytorch
from gpytorch.mlls import ExactMarginalLogLikelihood

from morganKernel import AdditiveGPModel_botorch
from training import train_model

train_x = torch.randn(20, 16)
train_y = torch.randn(20)
likelihood = gpytorch.likelihoods.GaussianLikelihood()

model = AdditiveGPModel_botorch(
    train_x=train_x,
    train_y=train_y,
    likelihood=likelihood,
    primary_dim=8,
    secondary_dim=8,
)

mll = ExactMarginalLogLikelihood(likelihood, model)
model = train_model(model, mll, train_x, train_y, max_iter=500, lr=0.1)

4. Work with the stored structures

If you need the 3D structures associated with a chemical-space ID, look in:

bayesianoptimization/data/chem_space_structures/<ID>.pdb

Example:

ls bayesianoptimization/data/chem_space_structures/00042.pdb

Notes And Assumptions

  • This directory is centered on model definition and candidate ranking, not raw descriptor generation.
  • The notebook is likely the main analysis entry point; the Python modules are lightweight helpers around that workflow.
  • The pickles are binary data products and are not self-describing from the command line. If you modify their schema, update this README and any loading code in the notebook.