State-specific protein-ligand complex structure prediction with a multi-scale deep generative model

TL;DR

NeuralPLexer tackles the challenge of predicting protein–ligand complex structures and their conformational changes directly from protein sequences and ligand graphs. It combines a Molecular Heat Transformer-based molecular encoding, a contact-prediction module, and an SE(3)-equivariant diffusion-based denoiser to sample atomistic structures and ensembles at multi-scale resolution. The approach achieves state-of-the-art results on blind docking and binding-site structure recovery, and demonstrates competitive holo-state accuracy by modeling ligand-induced conformations. This end-to-end framework enables rapid, scalable prediction of protein–ligand complexes and holds promise for accelerating drug design and protein engineering.

Abstract

The binding complexes formed by proteins and small molecule ligands are ubiquitous and critical to life. Despite recent advancements in protein structure prediction, existing algorithms are so far unable to systematically predict the binding ligand structures along with their regulatory effects on protein folding. To address this discrepancy, we present NeuralPLexer, a computational approach that can directly predict protein-ligand complex structures solely using protein sequence and ligand molecular graph inputs. NeuralPLexer adopts a deep generative model to sample the 3D structures of the binding complex and their conformational changes at an atomistic resolution. The model is based on a diffusion process that incorporates essential biophysical constraints and a multi-scale geometric deep learning system to iteratively sample residue-level contact maps and all heavy-atom coordinates in a hierarchical manner. NeuralPLexer achieves state-of-the-art performance compared to all existing methods on benchmarks for both protein-ligand blind docking and flexible binding site structure recovery. Moreover, owing to its specificity in sampling both ligand-free-state and ligand-bound-state ensembles, NeuralPLexer consistently outperforms AlphaFold2 in terms of global protein structure accuracy on both representative structure pairs with large conformational changes (average TM-score=0.93) and recently determined ligand-binding proteins (average TM-score=0.89). Case studies reveal that the predicted conformational variations are consistent with structure determination experiments for important targets, including human KRAS$^\textrm{G12C}$, ketol-acid reductoisomerase, and purine GPCRs. Our study suggests that a data-driven approach can capture the structural cooperativity between proteins and small molecules, showing promise in accelerating the design of enzymes, drug molecules, and beyond.

Figures (6)

• Figure 1: NeuralPLexer enables accurate prediction of protein-ligand complex structure and conformational changes. (a) Method overview. To perform predictions, the input protein sequence is first used to retrieve protein language model (PLM) features and structure templates; NeuralPLexer then combines the set of PLM and template features with molecular graph representations of the input ligands to directly sample an ensemble of binding complex structures via a multi-scale generative model. The main network of NeuralPLexer is comprised of a coarse-grained, auto-regressive contact prediction module (CPM) and an atomistic, diffusion-based equivariant structure denoising module (ESDM). SDE: stochastic differential equation. (b-c) Prediction example on a target with large-scale domain motions upon ligand binding (UniProt:P38998). (b) The structure similarities against experimental apo (i.e., ligand-free protein, PDB:3UGK) and holo (i.e., ligand-bound protein, PDB:3UH1) structures measured by TM-score are plotted for AlphaFold2 predictions (grey), ligand-free NeuralPLexer predictions (blue), and ligand-bound NeuralPLexer predictions (red). (c) Visualizations of representative NeuralPLexer-predicted structures (blue for apo, red for holo) are overlaid with the experimental apo structure (grey) and the holo structure (light yellow). (d) Visualization of a prediction example (PDB:7CKI, UniProt:P00953) for which NeuralPLexer achieves high atomic accuracy for both the ATP (blue) and an inhibitor bound to the tryptophan site (magenta) upon an induced-fit structure rearrangement.
• Figure 2: Architecture details. (a) Ligand molecules and amino acids are encoded as the collection of atoms, local coordinate frames (depicted as semi-transparent triangles), and stereochemistry-specific pairwise embeddings (depicted as dashed lines) representing their interactions. (b) Information flow in the contact prediction module (CPM) network. The CPM network samples block-wise adjacency matrices among protein and ligand nodes using an autoregressive decoding scheme, where the block adjacency matrices $\mathbf{l}_k$ sampled from last step are passed to the network to update the predicted histograms of pairwise distances ("distograms") and contact maps $\hat{\mathbf{L}}$. (c) The forward-time SDE introduces structured drift and noise covariance terms among protein $\mathrm{C\alpha}$ atoms, non-$\mathrm{C\alpha}$ atoms, and ligand atoms. (d) Denoising diffusion process to generate the binding complex 3D atomistic structure. The protein (colored as red-blue from N- to C-terminus) and ligand (colored as grey) structures are jointly generated through a reverse-time, simulated annealing SDE starting from randomly initialized coordinate variables. (e) Information flow in the ESDM neural network. The ESDM network operates on a heterogeneous graph formed by protein atoms (P), ligand atoms (L), protein backbone frames for all residues (B), backbone frames of the selected patches (S), and ligand local frames (F) to predict the denoised atomic coordinates $\hat{\mathbf{x}}_0, \hat{\mathbf{y}}_0$ used in the reverse-time SDE. The heterogeneous graph comprises randomized local edges (orange arrows) and densely connected long-range edges (blue arrows), where the long-range edges and inter-residue local edges are initialized via the CPM embeddings.
• Figure 3: Model performance on benchmarking problems. (a-d) Rigid backbone blind protein-ligand docking. The cumulative fraction of predictions with ligand RMSD below (a) 2 and (b) 5 over the test dataset are plotted against the number of ligand poses sampled per protein-ligand pair. (c) Assessing the fidelity of model-assigned confidence estimations. The ligand RMSDs are plotted against the model-assigned pLDDT score averaged over all ligand atoms (ligand pLDDT). (d) The precision-recall curve is evaluated based on the sampled structures ranked by the ligand pLDDT score, with the binary value of ligand RMSD < 2 being treated as the class label. (e-g) Flexible binding site structure recovery. (e) Visualization of prediction results on a test set example (PDB:6P8Y) near the structure recovery accuracy cutoff considered in this work. NeuralPLexer generates the binding site protein-ligand structure consistent with experiments, while directly aligning the AF2 prediction to ground-truth complex results in steric clashes (dashed circle). The red arrow indicates the qualitative conformational change from AF2 to the experimental bound-state structure. (f) Summary of binding site accuracy (measured by the lDDT-BS score) and ligand clash rate over the test dataset. 32 conformations are sampled for each protein-ligand pair; dots indicate the median value and error bars indicate 25th and 75th percentiles. (g) The structure recovery accuracy is compared to baseline methods. Solid lines correspond to recovery rates evaluated based on strict cutoffs (lDDT-BS > 0.7, ligand RMSD < $2.0\angstrom$, clash rate = 0.0), while the dashed lines correspond to relaxed cutoffs (lDDT-BS > 0.5, ligand RMSD < $2.5\angstrom$, clash rate < 0.05) with clash rate cutoff matching the 95% percentile of experimental structure statistics.
• Figure 4: Model predictions for contrasting apo-holo pairs from the PocketMiner dataset. (a-c) The TM-score with respect to apo and holo experimental reference structures are plotted for (a) AlphaFold2 (AF2) structures (b) NeuralPLexer-sampled apo structures, and (c) NeuralPLexer-sampled holo structures. For NeuralPLexer-sampled structures, the model-assigned pLDDT scores averaged among all protein residues are indicated by the dot colors. (d-e) TM-scores for baseline methods and NeuralPLexer averaged against all samples and the subset of targets for which any NeuralPLexer-predicted structure is of protein pLDDT=0.8 or higher. Error bars indicate the standard error of the mean calculated from six sets of predictions using independent random seeds and different AF2 model checkpoints to obtain the initial template. (f) The weighted Q-factor metric for conformational change prediction accuracy (see Methods, Weighted Q-factor) is plotted between NeuralPLexer predictions and AF2 predictions on all ligand-bound holo structures. Grey dots correspond to NeuralPLexer predictions in the absence of ligand inputs. (g) The weighted Q-factors for predicted ligand-bound holo structures are plotted against the maximum sequence similarity of each target to samples in the training dataset. (h) Assessing the fidelity of NeuralPLexer internal confidence predictions. The ligand-binding-site accuracy as measured by lDDT-BS is plotted against the ligand RMSD for all NeuralPLexer predictions; the model-assigned pLDDT scores averaged among all ligand atoms are indicated by the dot colors.
• Figure 5: Model predictions for recently determined structures. (a-b) TM-scores for baseline methods and NeuralPLexer averaged against all samples and the subset of targets for which any NeuralPLexer-predicted structure is of protein pLDDT=0.8 or higher. Error bars indicate the standard error of the mean calculated from six sets of predictions using independent random seeds and different AF2 model checkpoints to obtain the initial template. (c) The weighted Q-factor metric for conformational change prediction accuracy (see Methods, Weighted Q-factor) is plotted between NeuralPLexer predictions and AF2 predictions on all ligand-bound holo structures. Grey dots correspond to NeuralPLexer predictions in the absence of ligand inputs. (d) The weighted Q-factors for predicted ligand-bound holo structures are plotted against the maximum sequence similarity of each target to samples in the training dataset. (e) Assessing the fidelity of NeuralPLexer internal confidence predictions. The ligand-binding-site accuracy as measured by lDDT-BS is plotted against the ligand RMSD for all NeuralPLexer predictions; the model-assigned pLDDT scores averaged among all ligand atoms are indicated by the dot colors. (f-g) Model predictions on representative targets (PDB:6KPE, PDB:6PKH, PDB:6WQA) suggest structural elements for protein self-assembly and plausible models for enzyme catalysis and the activation of membrane receptors.