Pearl: A Foundation Model for Placing Every Atom in the Right Location (2510.24670v1)

Abstract: Accurately predicting the three-dimensional structures of protein-ligand complexes remains a fundamental challenge in computational drug discovery that limits the pace and success of therapeutic design. Deep learning methods have recently shown strong potential as structural prediction tools, achieving promising accuracy across diverse biomolecular systems. However, their performance and utility are constrained by scarce experimental data, inefficient architectures, physically invalid poses, and the limited ability to exploit auxiliary information available at inference. To address these issues, we introduce Pearl (Placing Every Atom in the Right Location), a foundation model for protein-ligand cofolding at scale. Pearl addresses these challenges with three key innovations: (1) training recipes that include large-scale synthetic data to overcome data scarcity; (2) architectures that incorporate an SO(3)-equivariant diffusion module to inherently respect 3D rotational symmetries, improving generalization and sample efficiency, and (3) controllable inference, including a generalized multi-chain templating system supporting both protein and non-polymeric components as well as dual unconditional/conditional modes. Pearl establishes a new state-of-the-art performance in protein-ligand cofolding. On the key metric of generating accurate (RMSD < 2 \r{A}) and physically valid poses, Pearl surpasses AlphaFold 3 and other open source baselines on the public Runs N' Poses and PoseBusters benchmarks, delivering 14.5% and 14.2% improvements, respectively, over the next best model. In the pocket-conditional cofolding regime, Pearl delivers $3.6\times$ improvement on a proprietary set of challenging, real-world drug targets at the more rigorous RMSD < 1 \r{A} threshold. Finally, we demonstrate that model performance correlates directly with synthetic dataset size used in training.

• The paper introduces Pearl, a novel generative model for protein-ligand cofolding that achieves atomic precision in structure prediction.
• It leverages an SO(3)-equivariant diffusion module and multi-chain templating to enhance accuracy and controllability during inference.
• Empirical benchmarks demonstrate significant improvements in ligand RMSD and physical plausibility, alongside notable inference and training efficiency gains.

Pearl: A Foundation Model for Protein–Ligand Cofolding with Atomic Precision

Introduction and Motivation

Pearl is introduced as a generative foundation model for protein–ligand cofolding, targeting the persistent challenges in computational drug discovery: limited and biased experimental data, physically implausible predictions, and insufficient controllability at inference. The model is designed to predict the three-dimensional structures of protein–ligand complexes with high accuracy and physical validity, directly addressing the limitations of both classical docking and recent deep learning-based cofolding approaches.

Architectural Innovations

Pearl’s architecture is characterized by a modular design that integrates rotationally and translationally invariant representation learning with an SO(3)-equivariant diffusion module. The trunk module, incorporating a lightweight triangle multiplication (trimul) mechanism, efficiently encodes pairwise relationships and conditions the diffusion process. The core innovation is the SO(3)-equivariant diffusion module, constructed from equivariant transformer (EqT) blocks, which ensures that the model’s predictions are equivariant to 3D rotations, thereby improving generalization and sample efficiency.

Figure 1: Key components of the equivariant diffusion module, including the equivariant transformer architecture and the equivariant feed-forward layer with a gated nonlinearity for vector components.

The model further generalizes the templating mechanism to include non-polymeric components, enabling conditioning on both protein and ligand structural priors. This multi-chain templating system supports both unconditional and pocket-conditional cofolding modes, providing fine-grained control during inference and facilitating the incorporation of auxiliary structural information.

Training Regimen and Data Strategy

Pearl’s training protocol is distinguished by its use of a large-scale, synthetically generated dataset in addition to curated PDB structures and monomer distillation data. The synthetic data, generated via physics-based methods with diverse virtual ligands, introduces chemical diversity that is absent from experimental datasets, directly addressing the data scarcity and bias inherent in the PDB.

Training is organized as a five-stage curriculum, progressively increasing task complexity and data diversity. Early stages focus on simpler, non-templated data, while later stages introduce complex structural priors and templating information. The model employs a conservative bfloat16 mixed-precision strategy for computational efficiency, with numerically sensitive operations retained in fp32 to ensure stability.

Figure 2: Pearl's performance scales with the addition of synthetic data, showing monotonic improvement in success rate as the proportion of synthetic data increases.

Inference and Controllability

Pearl supports two primary inference modes:

• Unconditional cofolding: Predicts complex structures from sequence and ligand topology alone, suitable for novel targets without known binding pockets.
• Conditional cofolding: Incorporates structural priors (e.g., known apo/holo structures or hypothesized pockets), enabling guided generation for practical drug discovery scenarios.

The model also supports advanced guidance and steering techniques, allowing users to enforce specific constraints or physical priors during sampling, and to modulate pose diversity at inference time.

Benchmarking and Empirical Results

Pearl is evaluated against state-of-the-art cofolding models (AlphaFold #13, Boltz-#11x, Boltz-#12, Chai-#11, ProteniX) on public (Runs N' Poses, PoseBusters) and proprietary (InternalXtals) benchmarks. The primary metric is ligand RMSD, supplemented by PoseBusters physical plausibility checks (PB-valid).

Figure 3: Runs N' Poses benchmark.

Pearl achieves 85.2% and 84.7% success rates (RMSD and PB-valid, best@15) on Runs N' Poses and PoseBusters, respectively, representing a 14.5% and 14.2% improvement over the next best models. Notably, the drop in success rate when applying physical validity checks is negligible (0.7% and 0.4%), indicating that Pearl generates almost exclusively physically plausible poses.

Figure 4: Stratified analysis of Pearl's generalization on RnP, showing strong performance in low-similarity regimes.

On the challenging InternalXtals dataset, Pearl demonstrates a 67.8% relative improvement over the next best open-source model in unconditional cofolding, and a 43% relative improvement in the conditional regime. Stratified analyses confirm that Pearl generalizes well to novel protein pockets and ligands, outperforming baselines in the most challenging, low-similarity settings.

Practical Relevance and Case Studies

Pearl’s high-accuracy performance at strict RMSD thresholds is critical for medicinal chemistry applications, where subtle errors (e.g., ring flips, missed interactions) can render predictions unusable. Qualitative analyses highlight both successes—where Pearl correctly predicts challenging poses missed by other models—and common failure modes, such as memorization artifacts and difficulties with large induced-fit changes.

Figure 5: Qualitative analysis of Pearl's successes and common cofolding failure modes, illustrating both correct predictions and typical errors.

(Figure 6)

Figure 6: Importance of high-accuracy predictions; even sub-2Å RMSD poses can contain critical errors such as ring flips or missed interactions.

Efficiency and Scaling

Pearl’s architectural and hardware optimizations, including the use of NVIDIA’s cuEquivariance kernels and mixed-precision training, yield substantial efficiency gains: 10–80% speedup at inference and 15% at training relative to vanilla PyTorch baselines. These optimizations enable training with larger context windows and facilitate large-scale experimentation.

Figure 7: NVIDIA's optimized cuEquivariance kernels provide significant acceleration for Pearl relative to a vanilla PyTorch baseline.

Implications and Future Directions

Pearl’s results substantiate several key claims:

• Synthetic data scaling: Model performance improves monotonically with the addition of synthetic data, validating this as a strategy for overcoming data scarcity in scientific ML.
• Equivariant architectures: Enforcing SO(3) symmetry via architectural design enhances sample efficiency and generalization.
• Controllability: The multi-chain templating system enables practical, guided structure generation, directly supporting real-world drug discovery workflows.

Despite these advances, challenges remain. Pearl’s accuracy degrades on out-of-distribution data, and pose selection via confidence models remains unreliable. Memorization artifacts persist, and the model is not immune to long-tail failures involving large conformational changes. Addressing these limitations will require further advances in OOD generalization, pose selection, and integration of protein dynamics.

Conclusion

Pearl establishes a new state-of-the-art in protein–ligand cofolding, combining innovations in data augmentation, geometric deep learning, and inference-time controllability. Its high-fidelity, physically plausible predictions are directly relevant for structure-based drug design, and its architectural principles—particularly the use of synthetic data and equivariant design—are likely transferable to other generative modeling problems in scientific domains. Continued development along these lines promises further advances in AI-driven molecular modeling and drug discovery.