Pearl: A Foundation Model for Placing Every Atom in the Right Location

TL;DR

Pearl introduces a foundation model for protein–ligand cofolding that addresses data scarcity, physical validity, and controllability through three innovations: large-scale synthetic data with a curriculum, an SO(3)-equivariant diffusion architecture, and a flexible multi-chain templating system. The model achieves state-of-the-art performance on public benchmarks and demonstrates strong generalization to novel targets, particularly at drug-discovery–relevant high-accuracy thresholds ($RMSD < 1~\mathrm{Å}$ with PB-valid). Its conditional, pocket-aware mode enables high-fidelity, structure-guided pose generation, while unconditional mode supports exploration of unknown pockets. The results suggest Pearl’s practical utility in accelerating structure-based drug design and offer a generalizable blueprint for scalable, physics-informed foundation models in science.

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 Å) 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 Å threshold. Finally, we demonstrate that model performance correlates directly with synthetic dataset size used in training.

Figures (11)

• Figure 1: (a) Pearl prediction (magenta) superimposed on the experimental ground truth (blue) for SARS-CoV-2 (PDB: 8S8X, in the Runs N' Poses, released March 2024); Pearl performance in the (b) unconditional (Runs N' Poses) and (c) pocket-conditional (InternalXtals) cofolding modes; (d) Pearl training and inference flow.
• Figure 2: Key components of the equivariant diffusion module, including the equivariant transformer architecture (left) and the equivariant feed-forward layer with a gated nonlinearity for vector components (right).
• Figure 3: Pearl demonstrates state-of-the-art performance in the unconditional cofolding mode across multiple benchmarks. Shown are best@5 success rates for generating accurate (RMSD < 1 Å and < 2 Å) and physically valid (PB-valid) poses. Dashed lines group models for fair comparison (e.g., based on the training cutoff dates). Note 1(*): AlphaFold 3 results for PoseBusters use officially released metrics for poses selected for max confidence out of 25 samples (not best@5). Note 2: AlphaFold 3 was excluded from the InternalXtals benchmark due to license restrictions. Note 3: Trained with additional public data up to 2023-06-01, Boltz-2 is not directly comparable to the other evaluated models, which use 2021-09-30 or earlier training cutoff. Note 4: Statistical significance (one-sided t-test) is shown as: $*: \mathrm{p\leq0.05}$, $**: \mathrm{p\leq0.01}$, $***: \mathrm{p\leq0.001}$.
• Figure 4: Stratified analysis of Pearl's generalization on the RnP in the unconditional cofolding mode. Shown are the best@5 success rate for generating valid poses (RMSD < 2 Å and PB-valid) when stratified by (left) overall similarity to the training set (product of binding pocket coverage and combined overlap score [SuCOS] of the ligand pose vskrinjar2025have), (middle) ligand frequency, and (right) Tanimoto similarity. Pearl exhibits strong generalization in the most challenging, low-similarity regimes, where it generally leads other models. The number of test examples in each slice is denoted by $n$. Boltz-2 is excluded as its expanded training set does not conform to the same similarity measures.
• Figure 5: Pearl demonstrates state-of-the-art performance in the conditional cofolding mode across multiple benchmarks. Metrics shown are best@5 (from 20 samples). Dashed lines group models for fair comparision based on the training cutoff dates. Note 1: AlphaFold 3 was excluded due to license restrictions. Note 2: Trained with data up to 2023-06-01, Boltz-2 is not directly comparable to models using earlier cutoffs.