COMPRESS

Compression Of Molecular Physical fields into Reduced Spatial Sites

COMPRESS is an optimization framework that maps an all-atom (AA) molecule of M atoms to K physically parameterized sites (K < M):

S ∈ R^{M×6}  →  V ∈ R^{K×6}
  (x, y, z, q, σ, ε)

Each site is defined by three spatial coordinates and three non-bonded interaction parameters - partial charge (q), Lennard-Jones radius (σ), and well depth (ε). The K sites are optimized to reproduce the density, electrostatic, and van der Waals (vdW) fields of the original AA molecule on a 3D face-centered cubic (FCC) grid. This yields a fixed-size, directly physically interpretable molecular representation whose compression level is controlled by K.

---

Directory Structure

COMPRESS/
├── COMPRESS.py          # Main entry point
├── README.md
├── pyproject.toml       # Package metadata and dependencies
├── example/
│   └── test.smi         # Example input
└── script/
    ├── __init__.py
    ├── extract_params.py    # SMI/PDB → ACPYPE → params CSV
    ├── init.py              # AA and COMPRESS(CG) grid initialization
    ├── grid.py              # Grid class (field computation)
    ├── loss.py              # Loss functions
    ├── update_features.py   # L-BFGS optimization loop
    └── write_file.py        # Save results

---

Installation

Option 1: pip (recommended)

All dependencies including OpenBabel are installed automatically:

git clone https://github.com/username/COMPRESS.git
cd COMPRESS
pip install -e .

Option 2: pip from GitHub (no clone needed)

# Latest
pip install git+https://github.com/username/COMPRESS.git

# Specific version
pip install git+https://github.com/username/COMPRESS.git@v0.1.0

After installation, compress command is available anywhere:

compress -t smi -n benzene -s 12

---

Requirements

• Python >= 3.10
• PyTorch >= 2.0.0
• NumPy >= 1.26.0
• pandas >= 2.0.0
• scikit-learn >= 1.3.0
• RDKit >= 2024.03.0
• acpype >= 2023.10.27
• openbabel-wheel >= 3.1.1

All of the above are installed automatically via pip install.

---

Usage

compress -t <type> -n <name> -s <n_sites> [options]

Required Arguments

Argument	Description
-t, --type	Input file type: smi or pdb
-n, --name	Molecule name (must match filename, e.g. benzene → benzene.smi)
-s, --site	Number of COMPRESS sites

Optional Arguments

Argument	Default	Description
--steps	50	Number of optimization steps
--grid_interval	0.3	Grid spacing (Å)
--grid_buffer	5.0	Grid buffer around molecule (Å)
--lr_T	1.0	Langevin temperature
--lr_noise_scale	1e-7	Langevin noise scale
--decay_T	0.5	Temperature decay factor
--decay_T_interval	3	Temperature decay interval (steps)
--tau_density	0.2 0.5	Tau values for density field
--tau_charge	0.2 1.0	Tau values for charge field
--tau_epsilon	0.2 2.0	Tau values for VDW epsilon field

---

Example

An example input is provided in example/test.smi:

c1ccccc1  benzene

Case 1: SMILES string directly

No input file needed - SMILES is written to test.smi automatically:

compress -t smi -n test -s 12 --smiles "c1ccccc1"

Case 2: SMILES file

Run from the directory containing test.smi:

cd example
compress -t smi -n test -s 12

Case 3: PDB file

If you already have a PDB file, run from the directory containing test.pdb:

compress -t pdb -n test -s 12

You can also use "-s all" option to generate COMPRESS rerpesentations from K=1 to K=M

In all three cases, the pipeline runs automatically:

1. Generate test.pdb from SMILES via RDKit (Cases 1 & 2 only)
2. Run ACPYPE → test.acpype/
3. Extract atomic parameters → test_params.csv
4. Initialize AA and COMPRESS (CG) grids
5. Optimize COMPRESS sites via L-BFGS
6. Save results → test_s12_COMPRESS.pt

If test_params.csv already exists (e.g. rerunning with a different site count), steps 1–3 are skipped automatically:

compress -n test -s 6   # reuses test_params.csv

Expected output:

>> ----------------------------------------
>> Name      : test
>> Input     : /path/to/example/test.smi
>> Sites     : 12
>> Device    : cuda
>> Output    : /path/to/example/test_s12_COMPRESS.pt
>> ----------------------------------------
>> Input file found: test.smi
>> Generating PDB from SMILES: test.smi
>> Saving PDB file: test.pdb
>> Running Acpype...
>> Acpype finished successfully!
>> Extracting params from: test.acpype
>> Params saved: test_params.csv
>> ----------------------------------------
>> Running COMPRESS...
>> ----------------------------------------
>> Optimizing 50 steps...
>> Step    1 | Grid: 0.8241 | Total: 1.2034
>> Step    2 | Grid: 0.7193 | Total: 1.0871
...
>> Step   50 | Grid: 0.1023 | Total: 0.2341
>> Results saved: test_s12_COMPRESS.pt

---

Output

Results are saved as a PyTorch .pt file ({name}_s{n_sites}_COMPRESS.pt):

import torch

data = torch.load("test_s12_COMPRESS.pt")

data["AA_pos"]   # All-atom positions      (N_aa, 3)
data["AA_chg"]   # All-atom charges        (N_aa,)
data["AA_sig"]   # All-atom sigma          (N_aa,)
data["AA_eps"]   # All-atom epsilon        (N_aa,)

data["pos"]      # COMPRESS site positions       (N_cg, 3)
data["chg"]      # COMPRESS site charges         (N_cg,)
data["sig"]      # COMPRESS site sigma           (N_cg,)
data["eps"]      # COMPRESS site epsilon         (N_cg,)

data["loss"]     # Final loss dict

---

Pipeline Overview

example/test.smi
   │
   ▼ RDKit (if smi)
example/test.pdb
   │
   ▼ ACPYPE
example/test.acpype/
   │
   ▼ extract_params
example/test_params.csv
   │
   ▼ COMPRESS (init → optimize)
example/test_s12_COMPRESS.pt

---

Data

The COMPRESS_Drugs-GEOM_K25 dataset is available on Zenodo at:

https://doi.org/10.5281/zenodo.20072228