The exponential growth of the synthesizable chemical space represents a bottleneck in modern drug
discovery, even among in silico methods. The rapid development of artificial intelligence has brought a
drastic reduction in drug development times, but the most impactful advances, often outsourced to
leading tech companies, typically remain behind the walls of big pharma. The goal of this project is to
develop an open, AI-driven computational drug discovery ecosystem that (i) enables fast, preferably real-
time access to large chemical spaces, (ii) democratizes rational drug design for researchers without
programming expertise, and (iii) keeps pace with the explosive growth of the chemical universe. We have
previously shown that protein-ligand binding can be characterized by a finite number of unique interaction
patterns (pharmacophores). At this level, we hypothesize that virtual screening results are highly reusable
and can be queried in real time after a single calculation. Building on this, the project will expand our
existing pharmacophore database with support for non-enumerated chemical spaces and dynamic
pharmacophore representations that capture protein conformational flexibility. We will develop AI-driven
interfaces to this database, including fine-tuned chemical language models, natural language querying via
Retrieval-Augmented Generation (RAG), and multi-agent systems to decompose and execute arbitrarily
complex tasks. These components will be integrated into cascaded discovery workflows with AI-assisted
docking workflows and state-of-the-art co-folding models such as AlphaFold3, and ultimately validated
with biochemical experiments targeting oncology, neuropharmacology, and protein-protein interactions.
All tools will be released openly, establishing an interactive discovery engine that accelerates the
identification of new medicines.
The project is supported by the Momentum ("Lendület") Programme of the Hungarian Academy of Sciences.