TLDR: Sesame is a novel generative AI model that efficiently transforms ligand-free (apo) protein structures into ligand-bound (holo-like) conformations. This is crucial for drug discovery, as holo structures are ideal for molecular docking but often hard to obtain. Sesame uses a flow matching framework to predict these conformational changes at a fraction of the computational cost of traditional methods like molecular dynamics. It has shown improved performance in predicting both large and small pocket changes, enhances cryptic pocket identification, and provides better input for molecular docking, thereby accelerating virtual screening workflows.
Molecular docking is a fundamental process in drug discovery, essential for understanding how potential drug molecules interact with proteins. Traditionally, this process relies heavily on high-resolution protein structures that already have a ligand (a drug-like molecule) bound to them, known as ‘holo’ structures. These holo structures provide the most accurate representation of a protein’s binding site, making them ideal for predicting how new drugs might fit. However, obtaining these precise holo structures through experimental methods is often expensive and time-consuming, limiting their availability.
In contrast, ‘apo’ structures, which are ligand-free, are much more accessible. They can be derived experimentally or predicted using advanced models like AlphaFold2. The challenge with apo structures is that their binding pockets are not always perfectly shaped to accommodate a ligand, leading to less accurate predictions in virtual screening campaigns. Historically, scientists have used computationally intensive methods like molecular dynamics simulations to artificially induce these holo-like conformations from apo structures, a process that can take days of GPU computation.
Introducing Sesame: A Generative Solution
A new generative model called Sesame aims to overcome this limitation by efficiently predicting the conformational changes needed to transform apo structures into holo-like ones. Published at the GEM workshop, ICLR 2025, Sesame offers a scalable solution to improve virtual screening workflows by generating geometries better suited for ligand accommodation at a fraction of the computational cost. You can read the full paper here: SESAME : OPENING THE DOOR TO PROTEIN POCKETS.
Sesame leverages a cutting-edge technique called flow matching, a generative modeling paradigm that learns to map one data distribution (apo structures) to another (holo structures). This allows the model to learn the continuous changes a protein undergoes when a ligand binds, predicting the desired ‘open’ or ‘induced-fit’ conformations without the need for lengthy simulations. The model focuses on the protein’s backbone atoms, representing them as rigid transformations, and decomposes the learning process into rotational and translational components.
Key Achievements and Impact
The researchers evaluated Sesame across various datasets, demonstrating its superior performance compared to existing methods like SBAlign and EGNN. Sesame showed significant improvements in predicting both large conformational changes (where the protein undergoes substantial shifts) and smaller, more subtle pocket movements. For instance, in the D3PM-Large dataset, Sesame achieved 38% of predictions with an RMSD (a measure of structural similarity) less than 2.0 Å to the reference holo structure, outperforming baselines.
Beyond just structural prediction, Sesame’s generated structures proved beneficial for other critical drug discovery tasks. When integrated with cryptic pocket identification algorithms like PocketMiner, Sesame-generated structures led to increased performance in predicting residues belonging to cryptic binding sites. This indicates that Sesame produces conformations that are indeed more geometrically similar to holo structures, enhancing the reliability of pocket detection.
Furthermore, Sesame-generated structures were successfully used as input for standard molecular docking protocols. Despite only modeling backbone atoms (requiring side chains to be predicted separately), the model’s output led to improved docking performance compared to using original apo structures. This confirms that Sesame provides viable and more accurate starting points for docking, which is crucial for identifying potential drug candidates.
Also Read:
- Unlocking Protein Language Models: A New Approach to Interpretable Features
- CellPainTR: A Breakthrough in Generalizable Cell Painting Analysis for Large-Scale Biological Discovery
Future Directions
While Sesame marks a significant advancement, the authors acknowledge areas for future development. A primary goal is to extend the framework to incorporate side chain modeling, as these play a vital role in protein-ligand interactions and docking accuracy. Additionally, they plan to expand their data generation efforts using molecular dynamics simulations to capture a wider diversity of binding pockets and movements, aiming for an end-to-end model capable of accurately modeling both subtle and large-scale conformational changes.
In conclusion, Sesame represents a powerful new tool in the computational drug discovery arsenal. By efficiently generating accurate holo-like protein conformations, it promises to accelerate virtual screening workflows and streamline the entire drug discovery pipeline, making the search for new medicines more efficient and cost-effective.
