Revolutionizing AI in drug discovery via innovative molecular representation paradigms

Projektbeschreibung

Von KI erfasste chemische Informationen umwandeln

Die KI hat zu mehreren Durchbrüchen in der Wissenschaft geführt. In der Arzneimittelforschung hat die KI jedoch eher zu einer Evolution als zu einer Revolution geführt. Immer ausgefeiltere Algorithmen führen zu immer geringeren Gewinnen, was eine Revolution der KI in der Arzneimittelforschung erforderlich macht. In diesem Zusammenhang wird das ERC-finanzierte Projekt ReMINDER die molekularen Darstellungen, auf die sich Algorithmen des tiefen Lernens stützen, völlig neu überdenken. Diese Darstellungen haben sich in den vergangenen Jahrzehnten nicht wesentlich verändert und begrenzen die Menge und Qualität der chemischen Informationen, die von der KI erlernt werden können. Um effizientere Modelle zu entwickeln und offene wissenschaftliche Herausforderungen zu lösen, wird das Projekt eine neue „molekulare Sprache“ entwickeln, die anspruchsvollere chemische Informationen erfasst, die auf KI zugeschnitten sind. ReMINDER wird die Landschaft der molekularen KI verändern.

Ziel

Artificial intelligence (AI) in the form of deep learning is driving unprecedented progress in numerous fields, e.g. for protein structure prediction and organic reaction planning. In drug discovery and chemical biology, such progress is an “evolution” rather than a revolution: several tasks still await to be solved by AI, e.g. accurate structure-activity and activity-cliff prediction, and design of structurally innovative chemical matter. Increasingly complex deep learning approaches are leading to progressively smaller gains in model capabilities, calling for a revolution in AI for drug discovery. The springboard for this project is a striking observation: while novel deep learning algorithms are in continuous development, the input ‘raw’ molecular representations they rely on (e.g. SMILES strings and molecular graphs) have not considerably changed in the last four decades – limiting the amount and quality of chemical information learnable by AI. The potential of capturing more sophisticated chemical information better into a new ‘molecular language’ is still untapped and bears promise to revolutionize molecular AI. ReMINDER will break with traditional approaches and shift the object of study from increasingly complex algorithms to novel molecular representation paradigms for AI. ReMINDER will be an agent of change in the molecular AI landscape, by developing a new representation framework at the interface between method development and experimental validation. ReMINDER will disrupt the potential of AI to (a) navigate complex structure-activity landscapes, (b) design innovative bioactive molecules from scratch, (c) leverage binding pocket information for molecule discovery. By transforming the chemical information captured for AI, we open opportunities to develop more efficient models and solve open scientific challenges. ReMINDER will create the basis for exciting new technology in the field of deep learning for drug discovery, and chemistry at large.