Protein-Ligand Complex Generator & Drug Screening via Tiered Tensor Transform

Jonathan P. Mailoa, Zhaofeng Ye, Jiezhong Qiu, Chang-Yu Hsieh, Shengyu ZhangarXiv 2023 · DOI · arXiv

Abstract

Modeling the "induced fit" of protein-ligand interactions is a critical yet computationally expensive step in drug discovery. We present the Tiered Tensor Transform (3T), a physics-based algorithm that generates flexible protein-ligand conformations with an 80-fold speedup over traditional Molecular Dynamics (MD). By utilizing a hierarchical optimization framework implemented in PyTorch, 3T captures structural flexibility through differentiable physics and an "energetic kick" mechanism. Our results demonstrate that 3T-generated ensembles significantly outperform state-of-the-art docking methods in identifying active drug candidates across multiple benchmarks, including DUD-E and DEKOIS 2.0.

Introduction: The Challenge of Induced Fit

In drug discovery, "docking" predicts how a small molecule (ligand) fits into a protein's binding pocket. Most standard tools treat the protein as a rigid object to save time, but in reality, proteins are flexible; they shift and "breathe" to accommodate a ligand—a concept known as induced fit. While Molecular Dynamics (MD) simulations can model this flexibility with high accuracy, they are computationally expensive, often taking days or weeks to process a single complex. This creates a bottleneck where researchers must choose between the speed of rigid docking and the accuracy of MD. The Tiered Tensor Transform (3T) bridges this gap, providing a way to generate diverse, physically realistic, flexible conformations at a fraction of the computational cost.

Methods: The Tiered Tensor Transform (3T)

The core of the 3T algorithm is a hierarchical approach to structure optimization that leverages differentiable physics to explore the protein-ligand energy landscape efficiently.

The 3T workflow: (a) Defining the binding pocket and cutoff radius. (b) The optimization process on the atomistic force field energy landscape, showing the "energetic kick" and subsequent minimization. (c) The resulting ensemble of flexible structures (magenta) compared to the initial (grey) and experimental co-crystal (green) structures.

Hierarchical Segmentation and Tiered Transformations

Instead of moving every atom independently, 3T groups atoms into a hierarchy. Micro-groups consist of amino acid residues or rotatable ligand fragments, while macro-groups consist of larger structures like protein loops or the entire ligand. The algorithm applies rotations and translations at different scales. Moving a macro-group allows the model to make large structural shifts (like a loop opening) that would take MD a long time to simulate.

Differentiable Physics and the "Energetic Kick"

Uniquely, the authors implemented an atomistic force field (calculating Lennard-Jones and Coulombic energies) directly in PyTorch. This allows the use of backpropagation to optimize the physical structure of the protein-ligand complex. To explore different possible shapes and escape local minima, the algorithm applies a random structural distortion—an "energetic kick"—and then relaxes the system back into a low-energy state through 3T minimization.

Key Insights and Results

Superior Accuracy through Flexibility

The paper demonstrates that 3T-generated structures are truly "ligand-induced." The protein pocket changes shape based on the specific ligand being docked, which is why 3T outperforms "ensemble docking" (which uses a library of pre-existing protein shapes). For the protein CDK2, 3T improved the ligand's positional accuracy (RMSD) by 2.03 Å compared to standard rigid docking.

Efficiency and Screening Performance

  • Speedup: 3T achieves a >80x lower computational cost compared to comparable MD simulations on a GPU.
  • Active Ligand Classification: In drug screening tests, using just 10 3T-generated conformations proved more effective than using hundreds of rigid experimental X-ray structures or MD snapshots.
  • Energy Landscapes as Features: The authors found that the energy path taken during the 3T optimization process provides valuable data. Using these energy features in a Gradient Boosting Tree classifier significantly boosted the accuracy of identifying real drug candidates from decoys.

Limitations and Future Work

While much faster than MD, this prototype of 3T is still slower than extremely lightweight semi-flexible docking tools. Additionally, 3T currently requires the user to define the binding pocket area manually. Future iterations could benefit from automated pocket identification. Beyond drug discovery, the authors suggest the "tiered transform" framework could be adapted to other fields requiring differentiable cost functions, such as optimizing optical nanostructures or microfluidic devices.

Conclusion

The Tiered Tensor Transform offers a powerful new tool for drug discovery, combining the physical rigor of MD with the speed required for high-throughput screening. By modeling the induced fit through a differentiable, hierarchical framework, 3T provides a scalable solution for predicting complex protein-ligand interactions.

Summary updated

Abstract

The generation of small molecule candidate (ligand) binding poses in its target protein pocket is important for computer-aided drug discovery. Typical rigid-body docking methods ignore the pocket flexibility of protein, while the more accurate pose generation using molecular dynamics is hindered by slow protein dynamics. We develop a tiered tensor transform (3T) algorithm to rapidly generate diverse protein-ligand complex conformations for both pose and affinity estimation in drug screening, requiring neither machine learning training nor lengthy dynamics computation, while maintaining both coarse-grain-like coordinated protein dynamics and atomistic-level details of the complex pocket. The 3T conformation structures we generate achieve significantly higher accuracy in active ligand classification than traditional ensemble docking using hundreds of experimental protein conformations. Furthermore, we demonstrate that 3T can be used to explore distant protein-ligand binding poses within the protein pocket. 3T structure transformation is decoupled from the system physics, making future usage in other computational scientific domains possible.
BibTeX
@misc{mailoa2023protein,
  author = {Jonathan P. Mailoa and Zhaofeng Ye and Jiezhong Qiu and Chang-Yu Hsieh and Shengyu Zhang},
  title = {{Protein-Ligand Complex Generator \& Drug Screening via Tiered Tensor Transform}},
  year = {2023},
  eprint = {2301.00984},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2301.00984},
}
Figure