PILOT: Equivariant diffusion for pocket conditioned de novo ligand generation with multi-objective guidance via importance sampling

TL;DR

PILOT addresses de novo ligand generation conditioned on protein pockets while optimizing multiple properties relevant to drug design. It combines an equivariant diffusion backbone with pocket conditioning and trajectory-based importance sampling guided by property surrogates, enabling backpropagation-free multi-objective control. Pre-training on a large conformer dataset and fine-tuning on CrossDocked2020 yield superior structural validity and drug-likeness, and Kinodata-3D experiments show enhanced potency predictions with maintained docking performance. The approach demonstrates practical potential for structure-based drug design, while highlighting data quality and distribution challenges and pointing toward integration with synthesis pipelines and expansion to biologics.

Abstract

The generation of ligands that both are tailored to a given protein pocket and exhibit a range of desired chemical properties is a major challenge in structure-based drug design. Here, we propose an in-silico approach for the $\textit{de novo}$ generation of 3D ligand structures using the equivariant diffusion model PILOT, combining pocket conditioning with a large-scale pre-training and property guidance. Its multi-objective trajectory-based importance sampling strategy is designed to direct the model towards molecules that not only exhibit desired characteristics such as increased binding affinity for a given protein pocket but also maintains high synthetic accessibility. This ensures the practicality of sampled molecules, thus maximizing their potential for the drug discovery pipeline. PILOT significantly outperforms existing methods across various metrics on the common benchmark dataset CrossDocked2020. Moreover, we employ PILOT to generate novel ligands for unseen protein pockets from the Kinodata-3D dataset, which encompasses a substantial portion of the human kinome. The generated structures exhibit predicted $IC_{50}$ values indicative of potent biological activity, which highlights the potential of PILOT as a powerful tool for structure-based drug design.

Figures (15)

• Figure 1: Top: PILOT $$ is first pre-trained unconditionally on an Enamine Real subset from the ZINC database. irwin_zinc15 We employ OpenEye's Omega to create at most five conformers per molecule.hawkins_2011_oeomega Afterwards, we fine-tune the model on CrossDocked2020 conditioned on the atoms of the pocket.crossdocked. Middle: Given the binding pocket of a protein, a noisy state of a ligand is sampled from the diffusion forward trajectory (here, t=300) as input to the diffusion model during training. The model has to retrieve the ground truth ligand ($M_0$). For training, a composite loss ($l_d$) is used for continuous (mean squared error) and categorical features (cross-entropy loss), respectively, together with a timestep-dependent loss weighting ($w(t)$). Bottom: At inference, a point cloud is sampled from a Gaussian prior (t=500). Given a binding pocket, the model retrieves a fitting ligand by following the reverse diffusion trajectory. At pre-specified steps, a property surrogate model (green crosses) guides the diffusion process towards desired regions in chemical space using importance sampling.
• Figure 2: Schematical depiction of the PILOT $$ network. Given fixed pocket atoms (purple), ligand atom coordinates, types, and charges as well as the ligands' topology get noised (green) using forward diffusion. Afterwards, attention-weighted message-passing is done on the fully connected ligand atoms (here not shown for better visibility) and the ligand-pocket and pocket-pocket interactions, which each are obtained using a radius graph for computational feasibility. The task of the model is to retrieve the ground truth atom coordinates, types, charges, and the bond types (red).
• Figure 3: The impact of varying dataset cutoffs and employing different training approaches (training from scratch versus pre-training) on the performance of our model and TargetDiff is analyzed. Top: We compare the sample quality using the PoseCheck metrics, where all values are min-max normalized to better evaluate the difference in performance. Bottom: We present the average clash counts (left) and average strain energies (right). Models with lower clashes and strain energies are considered to perform better and are thus preferred.
• Figure 4: Schematical depiction of the importance sampling algorithm. The shape of the prior (left) and target (right) distribution, where ligands at the target distribution are highlighted in two different regions based on a property function, which is synthetic accessibility in this case. At $t=T$ (left), noisy samples are drawn from the prior, and during the reverse trajectory, stochastic paths that lead to promising candidates are selected and de-noised in state-space to converge to samples from the data distribution at $t=0$ (right). Ligands in the green box refer to molecules with high synthetic accessibility according to SA score, while molecules in the red box refer to rather inaccessible ones.
• Figure 5: Correlation matrix that includes the number of rings, number of atoms, docking scores, quantitative estimate of drug-likeness (QED), and synthetic accessibility (SA) scores for the CrossDocked2020 training set.