Skip to content

⚙️ Installation

ChemicalDice Integrator (CDI) depends on advanced scientific and deep learning libraries such as RDKit, PyTorch, and HDF5. To avoid dependency conflicts and unnecessary overhead, CDI is designed with a modular installation approach.

[!TIP] Install only what you need based on your use case.

🛠️ Environment Setup

To setup an environment to run ChemicalDice, you can install Miniconda using the following commands:

wget https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh
bash Miniconda3-latest-Linux-x86_64.sh

Follow the prompts after the above commands to install Conda. Create a separate environment named chemicaldice using the conda create command:

conda create -n chemicaldice python=3.9
conda activate chemicaldice

📦 Install Packages

To use the ChemicalDice package, you need to install it along with its dependencies. You can install ChemicalDice and its dependencies using the following commands:

🧩 Minimal Installation

Purpose: Lightweight usage for:

  • Remote API inference
  • Structural validation
  • Schema-level operations

No heavy deep learning frameworks are required.

pip install ChemicalDice
pip install numpy pandas rdkit tqdm requests 
pip install scikit-learn xgboost lightgbm

Installs:

  • numpy
  • pandas
  • rdkit
  • tqdm
  • requests
  • scikit-learn
  • xgboost
  • lightgbm

🧬 Descriptor Generation

Purpose: Compute external modal descriptors (e.g., Mordred, Quantum, Grover) directly from SMILES strings using the CDI feature extraction modules.

pip install ChemicalDice[descriptors]

Installs (in addition to minimal setup):

  • mordred
  • transformers
  • tokenizers
  • networkx

[!NOTE] Some descriptors like MOPAC or 3D-MORSE require pre-compiled binaries. You can automatically configure them by running: 

cdi setup mopac

🧠 Training Environment

Purpose: Enables deep model optimization through two primary frameworks:

  • CDI-Basic: Multi-modal representation learning and unsupervised autoencoder dimensionality reduction.
  • CDI-Generalised: Supervised SMILES modeling utilizing Mamba (smi_ssed) structured state-spaces.
pip install ChemicalDice[training]

Installs (in addition to minimal setup):

  • torch
  • h5py
  • huggingface-hub
  • smi_ssed

[!NOTE] Initializing the Mamba kernels and model weights requires a one-time configuration: 

cdi setup gen

🚀 Deployment Server

Purpose: Production-ready setup for:

  • High-throughput inference
  • Asynchronous API serving
  • SMILES-to-embedding translation
pip install ChemicalDice[deployment]

Installs (in addition to minimal setup):

  • fastapi
  • uvicorn
  • pydantic
  • torch
  • smi_ssed

⚙️ Environment Configuration (cdi setup)

After installing the Python package, you must configure the external environment parts to enable 3D descriptors and deep learning architectures.

Command Target Description
cdi setup gen Generalized CDI Installs Mamba kernels and fetches SMI-SSED model weights from HuggingFace.
cdi setup mopac Quantum Descriptors Downloads and configures MOPAC and 3D-MORSE binaries.
cdi setup all Full Environment Runs both of the above (default).

[!IMPORTANT] The cdi setup gen command requires a working CUDA installation and git to clone the weight repositories.