molai — Molecular AI Toolkit (v0.1) STILL UNDER DEVELOPMENT

molai is a modular Python toolkit for molecular property prediction and de novo molecule generation using SMILES-based deep learning models.

It is designed for:

• QSAR modeling (e.g. pIC50 prediction)
• inverse molecular design
• generative chemistry research
• small to medium datasets (∼10²–10⁴ molecules)

Design philosophy
Simple building blocks → composable pipelines → research-ready workflows

---

Features

Prediction

• SMILES → property regression
• LSTM-based predictors
• Fingerprint-based baselines (MACCS, Morgan)

Generation

• SMILES LSTM language models
• SMILES Variational Autoencoder (VAE)
• Property-guided molecule generation

Optimization

• Filtering-based inverse QSAR
• Reinforcement learning (policy gradient)
• Latent-space optimization (VAE + latent predictor)

Chemistry utilities

• RDKit SMILES validation
• QED / SA / novelty filtering
• PhysChem constraints (MW, logP, scaffolds)

---

Core Concepts

molai cleanly separates modeling, training, and application logic: