Ready to Work Together?
Therapeutic Drug Characterisation
Multi-Agent Framework for Multi-Objective Molecular Optimisation
Modifying a molecule’s structure to improve characteristics such as efficacy, stability, or reduced toxicity is a crucial yet challenging task in drug discovery. Despite substantial research efforts, molecular optimisation frequently depends on iterative trial and error, which can yield suboptimal ligands. The challenge lies in balancing multiple interdependent properties, where even a small structural alteration can markedly influence molecular behaviour. Achieving the desired property improvements while maintaining scaffold integrity is essential.
Intractābilis has developed a collaborative multi-agent framework powered by a large language models (LLMs) to address these challenges, with particular emphasis on multi-objective molecular optimisation. The system applies instruction tuning to adapt the LLM backbone, enabling it to reconstruct the original SMILES string of a molecule when provided with its scaffold SMILES and specified target properties. This process is reinforced by the integration of domain knowledge to assess and prioritise candidate molecules.
Our solution demonstrates strong performance in a range of optimisation tasks, including the enhancement of drug properties and the improvement of target selectivity. It performs effectively in both single-property and multi-property optimisation scenarios, offering a scalable and systematic approach that reduces reliance on trial-and-error processes while improving the quality and relevance of optimised compounds.
Fragmentation-Driven Molecular Property Prediction
Molecular property prediction plays a central role in computer-aided drug discovery and design. Accurate characterisation of molecular properties can significantly shorten development cycles and reduce experimental costs, enabling faster delivery of high-efficacy drugs to market.
Intractābilis has developed a novel graph neural network architecture that not only delivers high prediction accuracy but also provides interpretable insights into the underlying chemical factors. By constructing a fragment-based molecular graph, the method enables message passing between fragments that are not directly linked through covalent bonds, capturing long-range chemical interactions.
The framework adopts a hierarchical learning strategy across multiple graph representations. Representations learned at lower structural levels are used to initialise feature sets at higher levels, enabling the model to build progressively richer chemical representations. This approach improves both the robustness and expressiveness of the learned features.
Importantly, the predictions produced by the model align closely with established chemical principles, reinforcing the reliability and trustworthiness of its outputs. By analysing the model’s prediction scores, we can identify the molecular constituents that have the most significant influence on property variations, providing valuable guidance for targeted drug design.
Quantum Circuit Synthesis for ADME Property Prediction
Accurate prediction of ADME (absorption, distribution, metabolism, and excretion) properties is a critical component in the assessment and optimisation of potential drug candidates. These properties directly influence a compound’s safety, efficacy, and overall suitability for further development, making them essential in early-stage drug discovery and evaluation.
Intractābilis has developed a novel Quantum Circuit Search (QCS) architecture for biomedical ADME prediction tasks. This approach leverages quantum computing to design circuits specifically optimised for the unique requirements of these applications. The framework incorporates an innovative scoring mechanism to evaluate circuit performance effectively, ensuring that the most promising quantum models are identified for further refinement.
Our methodology addresses both regression and classification challenges inherent in ADME prediction, including the difficulty of working with imbalanced datasets. In certain scenarios, the proposed solution demonstrates a measurable performance advantage over classical models, such as achieving a lower mean squared error in regression tasks. This opens the door to tangible improvements in predictive accuracy and robustness for pharmaceutical research, while also providing a foundation for continuous enhancement of quantum-based approaches in the life sciences domain.
Together, these capabilities represent a scalable, modular platform designed for modern drug discovery. We empower R&D teams to prioritise, design, and de-risk drugs compounds faster. By combining prediction, generation, and interpretation in a unified framework, we help organisations extract maximum value from minimal data and accelerate therapeutic development with clarity, confidence, and scientific rigour.
© 2025 Intractābilis