Periodic Reporting for period 1 - ReMINDER (Revolutionizing AI in drug discovery via innovative molecular representation paradigms)
Reporting period: 2023-01-01 to 2025-06-30
Artificial intelligence (AI) is transforming science and technology, from language translation to medical imaging. In chemistry and drug discovery, however, AI has not yet brought the same level of breakthrough. While deep learning (DL) tools like AlphaFold revolutionized protein structure prediction, their impact on small molecule discovery has been more incremental. A key reason is that, despite advances in algorithms, the basic “language” of molecules used by AI – such as SMILES strings and molecular graphs – has remained essentially unchanged for decades.
ReMINDER aims to address this gap by fundamentally rethinking how molecules are represented for AI applications. Instead of focusing solely on developing more complex algorithms, ReMINDER introduces entirely new molecular representation paradigms tailored for deep learning. These representations are designed to more richly encode chemical and biological information, allowing AI systems to better navigate molecular complexity and design novel therapeutics.
The project develops and benchmarks new representations that explicitly account for chemical properties, three-dimensional dynamics, and protein-ligand interaction features. These representations will be tested and validated across several drug discovery tasks, including de novo molecular design, structure-activity prediction, and structure-based screening. In doing so, ReMINDER seeks to build a new foundation for AI in chemistry, enabling models that are more efficient, generalizable, and able to tackle real-world challenges such as drug selectivity, polypharmacology, and molecular novelty.
ReMINDER aims to address this gap by fundamentally rethinking how molecules are represented for AI applications. Instead of focusing solely on developing more complex algorithms, ReMINDER introduces entirely new molecular representation paradigms tailored for deep learning. These representations are designed to more richly encode chemical and biological information, allowing AI systems to better navigate molecular complexity and design novel therapeutics.
The project develops and benchmarks new representations that explicitly account for chemical properties, three-dimensional dynamics, and protein-ligand interaction features. These representations will be tested and validated across several drug discovery tasks, including de novo molecular design, structure-activity prediction, and structure-based screening. In doing so, ReMINDER seeks to build a new foundation for AI in chemistry, enabling models that are more efficient, generalizable, and able to tackle real-world challenges such as drug selectivity, polypharmacology, and molecular novelty.
During the first period of the project, several new molecular representation strategies have been developed, each tailored to tackle different challenges in drug discovery. These include representations designed for generating new molecules from scratch, for learning from the three-dimensional shape and flexibility of compounds, and for capturing information about how small molecules interact with proteins.
Some of these representations have already been tested in a range of benchmarks, showing promising results in early evaluations. Others are still in development, but initial findings suggest they could open the door to new applications, such as structure-based design or modeling molecular interactions over time.
Although the project experienced some delays in recruitment, a strong and interdisciplinary team is now in place. Collaborations with academic and industry partners are helping to accelerate progress and explore new use cases, including enzyme design and protein interaction modeling.
Alongside technical developments, the project is also contributing to the broader research community by:
- Developing tools and formats to make new representations more accessible.
- Establishing collaborations across disciplines.
- Laying the groundwork for systematic evaluation and comparison of molecular representations.
Some of these representations have already been tested in a range of benchmarks, showing promising results in early evaluations. Others are still in development, but initial findings suggest they could open the door to new applications, such as structure-based design or modeling molecular interactions over time.
Although the project experienced some delays in recruitment, a strong and interdisciplinary team is now in place. Collaborations with academic and industry partners are helping to accelerate progress and explore new use cases, including enzyme design and protein interaction modeling.
Alongside technical developments, the project is also contributing to the broader research community by:
- Developing tools and formats to make new representations more accessible.
- Establishing collaborations across disciplines.
- Laying the groundwork for systematic evaluation and comparison of molecular representations.
Traditional molecular representations used in AI are often limited in the type and quality of chemical information they can express. As a result, even advanced algorithms may struggle to learn complex patterns or generalize to new tasks.
ReMINDER introduces a shift in perspective: rather than pushing existing models to their limits, it focuses on improving the information that these models receive in the first place. By doing so, it aims to make AI more effective, data-efficient, and capable of solving challenges that are currently out of reach.
The new representations developed in the project are designed to address key limitations of current approaches — for example, by incorporating richer chemical context, accounting for molecular flexibility, or enabling learning directly from protein binding sites. Some of the findings, including unexpected results on how molecular dynamics affect model performance, have already inspired new directions in the research team and beyond.
These innovations are expected to benefit not only drug discovery, but also adjacent fields such as materials science, chemical biology, and synthetic chemistry. By creating open-source tools and a conceptual framework, the project ensures that others can build on its results, fostering long-term impact and uptake by both academia and industry.
ReMINDER introduces a shift in perspective: rather than pushing existing models to their limits, it focuses on improving the information that these models receive in the first place. By doing so, it aims to make AI more effective, data-efficient, and capable of solving challenges that are currently out of reach.
The new representations developed in the project are designed to address key limitations of current approaches — for example, by incorporating richer chemical context, accounting for molecular flexibility, or enabling learning directly from protein binding sites. Some of the findings, including unexpected results on how molecular dynamics affect model performance, have already inspired new directions in the research team and beyond.
These innovations are expected to benefit not only drug discovery, but also adjacent fields such as materials science, chemical biology, and synthetic chemistry. By creating open-source tools and a conceptual framework, the project ensures that others can build on its results, fostering long-term impact and uptake by both academia and industry.