17.02.2025
In this article, we take a look at how Artificial Intelligence (AI) is being used to predict complex biomolecular interactions.
Thank you
Protein-molecule interactions lie at the heart of many biological processes. Some of these molecules, such as ligands or drugs typically in nanometre scale, bind to proteins to modulate their activity. These small molecules often interact with cavities or pockets surrounded by amino acids in the protein, or occasionally with flat “cryptic” sites.
Beyond conventional drugs, proteins interact with a diverse range of molecules, including nucleic acids, ions, and antibodies. These interactions are essential for achieving precise control over protein activity that may impact disease mechanism.
Traditionally, experimental methods such as X-ray crystallography and Nuclear Magnetic Resonance (NMR) have been used to provide atomic-resolution structures of proteins and a limited range of protein-bound molecules. Cryo-electron microscopy (cryo-EM) further advanced the study of protein-molecule structures by allowing researchers to visualise large and complex molecules in their native state. These experimental methods typically demand significant amounts of purified proteins requiring complex sample preparation, data collection and analysis.
Further, predicting protein-molecule interactions has been challenging due to the complexity and diversity of biomolecular structures, and the dynamic nature of protein conformations. Those conformational changes are challenging to predict from the native, non-ligand bound structure of the protein.
Advances in computational modelling have significantly reduced reliance on traditional experimental methods. Molecular Dynamics (MD) simulations identify stable positions of a protein-ligand complex by sampling multiple conformations of the complex that are at the local energy minima. Monte Carlo simulations and fragment-based docking techniques have also been widely used to predict binding modes of a ligand onto a protein. Other notable computational techniques include quantum mechanics/molecular mechanics (QM/MM) simulations for capturing electronic-level changes in binding.
These computational methods, combined with advancements in GPU computing and generative AI-driven ligand design, are overcoming challenges associated with protein-ligand interactions.
In a previous article, we discussed how using generative AI can be leveraged to predict and design novel protein structures. Building on that, we now explore how generative AI models can predict complex protein interactions with a wide range of molecular complexes, including ligands, nucleic acids, and antibodies.
Alphafold3 represents a significant advancement in biomolecular modelling by introducing a generative AI-based "diffusion" module that operates directly on 3D atomic coordinates. This approach contrasts with Alphafold2, which relied primarily on amino acid sequence inputs. A recently published patent application from DeepMind (WO2024240774) outlines the diffusion-based method, which begins by introducing Gaussian noise to atom coordinates, creating "noisy molecular structure data." The diffusion model processes this data to generate 3D molecular structures in the form of raw atomic coordinates. This approach significantly improves accuracy while eliminating the need for extensive parameterisation, by providing an output in a form of distribution of possible structures. Alphafold3 is reportedly capable of predicting protein interactions with a diverse range of molecules, delivering at least a 50% improvement over traditional docking tools and other conventional methods[1].
Another diffusion-based generative AI model often used to predict protein-molecule binding is DiffDock. DiffDock builds on diffusion models by incorporating physics-based constraints. For instance, the protein being modelled can be treated as relatively rigid, while the ligand is considered more flexible. DiffDock calculates translational, rotational, and torsional scores to accurately predict the ligand's position and its interaction with the protein.
Other computational strategies employ a new approach of initially designing the surface that lies between the protein-molecule, and optimising the molecule based on the predicted surface. One approach combines geometric features, such as distance-dependent curvature, with chemical features including hydrogen bond donor/acceptor propensity, to tailor the design of protein-molecule surface[2].
The generative AI approaches discussed above can be combined to optimise the target binding candidate molecules. In a situation where the location of the binding site is unknown, “blind docking” methods can be performed by computationally placing the molecule around the protein and iteratively translating and rotating it to explore potential binding conformations. DynamicBind serves as an example of a blind docking model.
If the identity of the binding molecule is uncertain, fragment-based docking techniques – long-established in computational biology – can be integrated with a generative AI approach to design a specific portion of the molecule. The remaining portion can be further optimised. Alternatively, the entire molecule can be generated using Alphafold3 or DiffDock.
The stability of the target binding candidates can be assessed using molecular dynamics (MD) tools. Focusing MD predictions on a limited number of candidates significantly reduces computational resource requirements while covering the appropriate timescales of conformational changes in proteins upon binding. In addition to stability analysis, other properties such as absorption, distribution, metabolism, excretion, and toxicity (ADMET) can also be predicted using various AI tools.
Collectively, the approaches outlined above have mapped complex molecular assemblies such as RNA-protein complexes, multi-protein systems, and drug-bound proteins1. By integrating generative AI with existing computational methods and experimental workflow, these approaches are expected to accelerate the discovery of novel therapeutic candidates.
Keltie attorneys have extensive experience in this rapidly evolving field. Our experts in chemistry, biology and AI combine industry knowledge with business insight to help clients navigate the complexities of IP protection. If you would like to discuss your innovation, please do not hesitate to get in touch with us.
[1] Abramson, J., Adler, J., Dunger, J. et al. Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature 630, 493–500 (2024).
[2] Marchand, A., Buckley, S., Schneuing, A. et al. Targeting protein–ligand neosurfaces with a generalizable deep learning tool. Nature (2025).
13.02.2025
Packaging innovations and IPAhead of the Packaging Innovations & Empack exhibition, Nathaniel Taylor takes a look at the forms of Intellectual Property (IP) typically arising in the packaging industry and the boundaries between the different forms of protection that might be available. In the packaging industry, companies typically seek various forms of IP protection for packaging innovations, including patents, registered designs, and trademarks.
05.02.2025
The role of patents in promoting AI investmentIn January 2025, UK Prime Minister Keir Starmer and Secretary of State for Science, Innovation and Technology Peter Kyle announced the AI Opportunities Action Plan. The Plan has three goals: (1) Invest in the foundations of AI; (2) Push hard on cross-economy AI adoption; and (3) Position the UK to be an AI maker, not an AI taker.
Thank you
