Fenglei Li

Drug design is a time-consuming and labor-intensive process. However, numerous artificial intelligence (AI) approaches have emerged to accelerate this process, including structure generation, quantitative structure-activity relationship (QSAR) prediction, and molecular generation, which have significantly advanced drug development. Despite these advancements, most AI methods rely solely on existing databases, overlooking the valuable experience of medicinal chemists, which represents a vast and underutilized resource. For unique or rare molecules, existing databases may be insufficient to support effective machine learning models due to limited samples. By incorporating expert knowledge as priors, specifically through Human-in-the-Loop (HITL) methodologies, these challenges can be alleviated. Furthermore, we mainly focus on the design aspect, using machine learning models to generate molecules with desired properties. While traditional machine learning methods incorporate predefined molecular properties, HITL approaches allow the integration of implicit human preferences, enabling the design of molecules that align more closely with the intuition of medicinal chemists.

In summary, we aim to leverage HITL methodologies, combining expert knowledge with machine learning models, to generate promising drug candidates and advance drug development.