Please join Yale Chemistry for a thesis seminar with Anton Morgunov, Batista Lab.
Title: Structure Precedes Quantity: Generative Foundations for Molecular Design and Synthesis Planning
Summary: While AlphaFold demonstrated that structural mastery is the prerequisite for solving biology, chemistry remains stuck in a loop of regression on small, noisy datasets. This thesis argues that the path to true Artificial Chemical Intelligence lies not in predicting properties, but in mastering the generative grammar of synthesis.
We present a unified framework for this transition across three pillars: Design, Planning, and Rigorous Evaluation. First, ChemSpaceAL introduces active learning “beacons” to navigate chemical space, enabling high-efficiency generation of protein-specific binder candidates. Second, DirectMultiStep (DMS) reformulates retrosynthesis as a sequence-to-sequence translation task, replacing brittle search heuristics with a holistic chemical language model. Finally, RetroCast exposes the field’s “validity crisis,” proving that current benchmarks incentivize hallucination, and establishes the engineering standards required to build reproducible, foundation-scale models.
Collectively, this work shifts the objective function of the field: from the retrospective retrieval of known chemistry to the prospective, generative planning of the adjacent possible.
