LEAD-k: Latent Error Adversarial Discriminator for kinase

LEAD is a deep-learning multi-task framework designed to accurately predict drug targets for novel compounds. The architecture features a dynamic Autoencoder (with multi-decoder latent reconstruction) and a dual-discriminator that calibrates classification confidence and manages out-of-distribution (OOD) risk in novel chemical spaces.

---

🛠️ Quick Start Instructions

To clone, set up, and deploy LEAD locally or on a remote cluster, follow these steps:

Step 1: Clone the Repository

git clone https://github.com/mitkeng/leadk.git
cd leadk

Step 2: Establish the Virtual Environment

It is highly recommended to use an isolated conda or virtualenv space to prevent dependency collisions.

# Creating and activating conda environment
conda create -n lead_env python=3.10 -y
conda activate lead_env

Or using virtualenv:

python3 -m venv lead_env
source lead_env/bin/activate

Step 3: Install Required Dependencies

Ensure dependencies are updated and installed securely:

pip install --upgrade pip
pip install -r requirements.txt

---

📖 Command Line Manual (CLI User Guide)

This application is split into two primary modular pipelines: Training (train.py) and Inference (predict.py).

1. Training Pipeline (train.py)

Use this module to retrain your autoencoder networks and adversarial discriminators on custom tabular feature profiles.

Positional & Optional CLI Arguments:

• --train_pos (Required): String. Path to the CSV file representing active/positive compound configurations.
• --train_neg (Required): String. Path to the CSV file representing negative/unlabelled background configurations.
• --save_dir (Optional): String. Output folder to write model checkpoints and transforms (Default: models).
• --epochs (Optional): Integer. Maximum number of training epochs (Default: 250).
• --model_lr (Optional): Float. Adjust learning rate parameters for model gradients (Default: 0.0003).
• --decoders (Optional): Integer. The number of independent reconstruction decoders for your feature space (Default: 5).
• --adv_weight (Optional): Float. Adversarial classification loss regularization scale parameter (Default: 0.03).
• --seed (Optional): Integer. Anchor global seed parameters for perfect result reproduction (Default: 42).

Training CLI Command Example:

python train.py \
  --train_pos "Cancer_Drug_Train_Sanitized.csv" \
  --train_neg "Cancer_Drug_Test_Negative.csv" \
  --save_dir "models" \
  --epochs 250 \
  --model_lr 0.0003 \
  --decoders 5

---

2. Inference & Target Screening Pipeline (predict.py)

Predict kinase target affinities for newly designed target molecules with active dual-discriminator logit shielding.

Positional & Optional CLI Arguments:

• --input (Required): String. Path to input screening CSV containing custom descriptor values.
• --checkpoint (Optional): String. Target model parameters checkpoint path (Default: models/lead_checkpoint.pth).
• --transforms (Optional): String. Saved preprocessing transform configurations path (Default: models/transforms.pkl).
• --output (Optional): String. Filepath to store the prediction log results (Default: predictions_output.csv).
• --penalty_max (Optional): Float. Maximum temperature penalty factor allowed (Default: 6.0).

Inference CLI Command Example:

python predict.py \
  --input "Cancer_Drug_Test_Strictly_Novel.csv" \
  --checkpoint "models/lead_checkpoint.pth" \
  --output "Test_Prediction_Logs.csv" \
  --penalty_max 6.0