Qubigen Applies Federated AI to Generate Novel Candidates for First Drug Development Client

Qubigen reports the successful application of its Federated AI-Drug Design platform in collaboration with our first Drug Development Client, focusing on a confidential therapeutic target. This marks the first external validation of Qubigen’s privacy-preserving AI pipeline, drawing insights from proprietary Client data as well as Qubigen’s own molecular and ADMET knowledge database to accelerate the design of novel drug candidates.

Aims

Qubigen engaged with the Client to evaluate Qubigen’s Federated AI performance for an oncogenic target protein. The goals were:

1. To reproduce the training set of initial hits to validate that the AI is properly exploring the configurations of viable molecules.

2. To replicate the Client’s top molecules previously selected for further development, to assess the AI’s ability to generalize.

3. To additionally generate novel candidates with promising ADMET profiles, which were not initially considered by the Client.

Methods

Qubigen’s Federated AI was applied to a proprietary dataset of unique compounds, with access and validation managed by a secure Data Portal. The target-specific data included results for binding affinity and functional inhibition. The Federated AI method involved selecting one of Qubigen’s pretrained AI models, initially trained on Qubigen’s extensive proprietary ADMET dataset, and the AI was then fine-tuned on target-specific data to arrive at a stable generative AI.

Qubigen’ Federated AI allowed training on both datasets, Qubigen’s in-house data, and the Client’s data, to ‘virtually’ combine the insights into a multi-modal federated dataset, without ever moving or seeing the contributed data.

The resulting AI structures were then evaluated for binding in silico using an integrated AI-Virtual Screening pipeline, involving molecular mechanics and quantum chemistry screening, to select the top candidates. AI-based predictions of synthetic feasibility and ADMET properties were also provided.

Results

1. Dataset Reproducibility

Qubigen’s Federated AI reproduced 99% of active molecules from the training dataset, confirming that the generative AI could explore the valid chemical space of small molecules defined by the Client’s target protein using traditional SAR methods.

2. Blind Holdback Recovery

A set of 27 top candidates were held back from the AI training dataset and used to evaluate the AI’s ability to generalize. 100% of the compounds were reproduced by Qubigen’s Federated AI, representing an enrichment ratio of 6-fold over the original focused dataset. This validated the AI’s ability to prioritize relevant candidates.

3. Generation of Novel Compounds

The AI also produced 64 novel compounds structurally dissimilar to the training and holdback datasets, with 9 prioritized based on strong binding energy predictions, favorable ADMET profiles, and synthetic feasibility. These novel candidates exhibited drug-like properties on par with, or superior to, the 27 top candidates.

Conclusion

This successful collaboration demonstrates that Qubigen’s unique Federated AI platform can:

• Accurately reproduce known SAR chemical series,

• Generalize to unseen, high-quality candidates, and

• Design novel compounds with strong predicted drug-likeness profiles.

These results were achieved while maintaining full data privacy and sovereignty through the Federated AI process, underscoring its potential to drive collaborative drug discovery without compromising IP, while enabling rapid, secure, and scalable exploration of chemical space.

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 promising lead drug candidates - get in touch to explore how we can support your next discovery.