Qubigen Successfully Completes Initial Testing of Generative AI for Molecular Candidates

Qubigen is excited to announce the successful completion of our initial research project which used Generative AI to accelerate novel drug design. This innovative research evaluated the performance of Federated AI versus traditional, centralized AI training, with both models trained exclusively on binding data for a specific oncogenic target. Qubigen is dedicated to advancing smarter, more efficient and privacy-conscious approaches to drug design.

Aims

The primary aim of the research was to evaluate whether Federated AI could match or exceed centralized AI training in generating novel, potent molecule candidates. Our study evaluated the outputs of both training methods based on:

• Reproducibility: How well the models explored known chemical space.

• Novelty: Whether the models proposed genuinely new molecular structures with promising properties.

Methods

To assess performance, we trained and evaluated Generative AI models using Qubigen's proprietary datasets. Each model was optimized for:

• Reproducibility: whether the AI was correctly able to explore the configurations of valid, potent molecules.

• Novelty: whether the AI was additionally able to propose molecules that were sufficiently different from the training data.

• Diversity: whether the AI was able to propose a wide range of different molecular scaffolds, rather than merely replicating a constrained set of compounds.

Virtual screening was applied to AI-generated molecules after training the AI on multiple datasets looking at various molecular properties (binding assays, functional assays, and ADMET assays) to evaluate the binding potential and drug-likeness of novel molecules.

We ran three benchmark experiments, each starting from different in-house data sources, to generate 10,000 candidate molecules per scenario.

Results

Federated AI yielded comparable, and in some cases superior, results to traditional centralized AI training, whilst still preserving data privacy:

• Reproducibility: Older AI model architectures lead to better coverage of the training data, whilst newer architectures still maintained good reproducibility.

• Novelty: Older AI model architectures generated fewer novel molecule candidates; newer architectures generated a good level of novelty.

• Diversity: Older AI model architectures generated a more constrained set of molecular scaffolds; newer architectures were less likely to repeatedly generate similar molecules.

One of the top binders generated using new model architectures, including Federated AI, shared some structural and binding characteristics with a positive candidate known to be in commercial pharmaceutical development, despite the original candidate not being present in any training dataset. The molecular scaffold of the newly generated candidate was of a particularly novel shape that had not previously been generated using older AI methods in the field.

The new candidate was innovative enough to be considered unique IP, highlighting the ability of Qubigen Federated AI to generate intellectually novel, high-value drug candidates.

Conclusion

Qubigen’s AI algorithms demonstrated potential to accelerate drug design by generating novel, IP-eligible compounds that bind to specific therapeutic targets. These results highlight Qubigen’s ability to produce innovative, commercially promising, and biologically relevant molecules, all without compromising data privacy.

Qubigen - Accelerate Drug Design Without Exposing Secrets

Whether you're advancing active programs, reviving dormant data, or starting from scratch, Qubigen’s secure AI platform and virtual screening capabilities can help you identify, optimize, and accelerate the path to novel drug candidates, get in touch to explore how we can support your next discovery.

Image Credit: Zeng & Treutlein