A groundbreaking approach to language models that enables controlled bidirectional generation specifically designed for pharmaceutical applications.

• Combines strengths of multiple LLM architectures (MLM, CLM, Seq2Seq) while addressing their limitations
• Inspired by biological processes of growth and evolution to optimize molecular structures
• Demonstrates superior performance on viral and cancer drug optimization benchmarks
• Provides a flexible framework that balances controllability with generation capabilities

This research represents a significant advancement for medical drug discovery, potentially accelerating the development of targeted therapeutics by enabling more precise molecular optimization techniques.

ControllableGPT: A Ground-Up Designed Controllable GPT for Molecule Optimization