AuroBind: Unified Framework for Structure-Function Screening

• AuroBind is a unified deep-learning framework that simultaneously predicts atomic-resolution protein–ligand complex structures and binding fitness scores.
• It integrates sequence-based inputs with advanced diffusion methods and direct preference optimization for coupled structure–function inference.
• AuroFast, the distilled student model, delivers 10^5x speedups, enabling ultra-large screening with high atomic-level accuracy and hit efficiency.

AuroBind is a unified deep-learning framework for structure-based virtual screening that simultaneously predicts atomic-resolution protein–ligand complex structures and a scalar “fitness” score reflecting binding potency, while enabling ultra–high-throughput virtual screening across ultra-large compound libraries (Zhang et al., 4 Aug 2025). Introduced in 2025, it fine-tunes a custom atomic-level structural model on million-scale chemogenomic data and integrates direct preference optimization, self-distillation from high-confidence complexes, and a teacher-student acceleration strategy. It was proposed in the context of a drug-discovery landscape in which over 96% of human proteins remain unexploited by approved therapeutics, and in which existing structure-based virtual screening methods were characterized as lacking atomic-level precision and failing to predict binding fitness (Zhang et al., 4 Aug 2025).

1. Conceptual scope and problem formulation

AuroBind is designed to jointly model structure and function. Rather than treating pose prediction and activity prediction as separate tasks, it produces full-atom protein–ligand complex structures together with a global binding-fitness scalar in a single framework (Zhang et al., 4 Aug 2025). The stated objective is to bridge the gap between accurate complex-structure modeling and large-scale functional screening.

The framework targets a specific limitation of conventional structure-based virtual screening. Existing methods were described as promising for expanding the druggable proteome, yet limited by insufficient atomic-level precision and by the inability to predict binding fitness. AuroBind therefore formulates virtual screening as a coupled structure–function problem rather than as docking alone. This suggests that its main departure from purely geometric or confidence-based ranking schemes lies in making potency-linked prediction a first-class output rather than a secondary proxy.

AuroBind is also presented as requiring only protein sequence and ligand SMILES, with no holo structure or known binders. In the authors’ formulation, this is central to its applicability to orphan, flexible, or cryptic-pocket targets (Zhang et al., 4 Aug 2025).

2. Core architecture

AuroBind builds on the core of a structural foundation model inspired by AlphaFold 3 (Zhang et al., 4 Aug 2025). Its encoder, termed “PairFormer,” comprises 48 blocks and processes unified protein, at residue level, and ligand, at atom level, token representations. The structural branch is a diffusion-based structure module that iteratively denoises coordinates to generate full-atom complex structures and outputs confidence metrics including pLDDT, PAE, and PDE.

Attached to the PairFormer outputs is a chemogenomic fitness head. This head predicts both a per-token residue/atom fitness map and a global binding-fitness scalar. In this design, the structural representation is not merely an intermediate for coordinate generation; it also serves as the substrate for functional prediction. The architecture is therefore explicitly multi-output: geometry, confidence, and potency-linked fitness are all learned within one system (Zhang et al., 4 Aug 2025).

To support ultra-large library screening, the full model is distilled into a lightweight student model termed AuroFast. AuroFast retains only the components needed for fitness prediction and is intended to preserve the full model’s ranking utility while providing order-of-magnitude speedups (Zhang et al., 4 Aug 2025).

3. Training corpus and optimization strategy

AuroBind is trained in multiple stages spanning structural supervision, self-distillation, and chemogenomic fine-tuning (Zhang et al., 4 Aug 2025).

Stage	Source	Content
Stage I	PDB	~100 K high-quality protein–ligand complexes from the PDB (≤9 Å resolution, ≤300 chains), filtered for chemical completeness and low clash counts
Self-distillation augmentation	BindingDB predictions	~230 K additional complexes meeting ipTM > 0.8, ligand_pTM > 0.5, pLDDT > 70
Stage II	ChEMBL, PubChem, ExCAPE-DB procedures	~1.27 M chemogenomic pairs across 1,300+ protein domains with measured pXC₅₀ values

The optimization strategy combines regression and preference learning. For pairwise ranking, the framework uses direct preference optimization (DPO). Given two ligands $x_i, x_j$ for the same target, with true preference $y_{ij} = +1$ if $x_i$ is tighter than $x_j$ and $-1$ otherwise, the preference loss is

$$L_{\text{pref}} = \mathbb{E}_{i,j}\left[\ell\big(s(x_i)-s(x_j), y_{ij}\big)\right],$$

where $s(x)$ is the predicted fitness score and $\ell(\cdot,\cdot)$ is a pairwise logistic or hinge loss (Zhang et al., 4 Aug 2025).

In practice, the ranking objective is implemented with a Plackett–Luce style soft-ranking:

$$P(R_{\text{true}}|R_{\text{pred}}) \propto \prod_{k=1}^{K} \frac{\exp\left(s(x_{r_k})/\tau\right)}{\sum_{m=k}^{K}\exp\left(s(x_{r_m})/\tau\right)},$$

and training minimizes $-\log P(R_{\text{true}}|R_{\text{pred}})$ over confident samples with confidence $y_{ij} = +1$0, using temperature $y_{ij} = +1$1 (Zhang et al., 4 Aug 2025).

Self-distillation operates by having an initial teacher predict complexes on unlabeled BindingDB examples. Predictions exceeding the ipTM, ligand_pTM, and pLDDT thresholds are added to a self-distilled dataset. Subsequent student epochs mix equal batches of PDB and self-distilled samples. This procedure is intended to increase structural diversity while retaining confidence filtering (Zhang et al., 4 Aug 2025).

4. Joint structure–fitness inference and AuroFast distillation

AuroBind performs joint inference in a single forward pass: it outputs full-atom coordinates of the protein–ligand complex and a scalar fitness score (Zhang et al., 4 Aug 2025). Structural supervision retains standard diffusion reconstruction losses, specifically atom-pair distance $y_{ij} = +1$2 and frame-aligned coordinate losses. Fitness supervision combines mean-squared error on pXC₅₀,

$y_{ij} = +1$3

with the DPO ranking loss. During DPO, structural confidence scores weight each pairwise loss term, thereby focusing optimization on geometries the model trusts most (Zhang et al., 4 Aug 2025).

The acceleration mechanism used for AuroFast is teacher–student distillation. For each large cluster of screening compounds, formed by ECFP4-Tanimoto clustering, AuroBind computes a single structural embedding, or trunk representation, for the cluster centroid complex. AuroFast then learns the mapping

$y_{ij} = +1$4

by minimizing

$y_{ij} = +1$5

where $y_{ij} = +1$6 are student/teacher trunk embeddings and $y_{ij} = +1$7 are the fitness scalars (Zhang et al., 4 Aug 2025).

At inference time, each novel compound borrows its nearest-neighbor centroid’s structural embedding, enabling sub-millisecond fitness prediction without full complex modeling. This is the basis for the claimed screening scale of the student system (Zhang et al., 4 Aug 2025).

5. Benchmark performance

On structural benchmarks, AuroBind was evaluated on PoseBuster V1 and V2 with $y_{ij} = +1$8. It achieved success rates, defined as RMSD $y_{ij} = +1$9 Å, of 79.1% and 81.7%, compared with AlphaFold 3 at 78.0% and 81.0%, Protenix at approximately 73%, and classical docking methods, Vina and GOLD, at 51–61% (Zhang et al., 4 Aug 2025). A specific example reported for the 7OFF complex gave RMSD = 0.61 Å for AuroBind versus 6.77 Å for AlphaFold 3.

On functional prediction benchmarks using DAVIS and BindingDB with AUPR as the metric, zero-shot AuroBind-AZ achieved AUPR = 0.22/0.24, whereas an AF3 confidence proxy achieved 0.08/0.13 (Zhang et al., 4 Aug 2025). Fine-tuned AuroBind-FT achieved AUPR = 0.61 on DAVIS and 0.70 on BindingDB, compared with the supervised baseline ConPLex at 0.46 and 0.63.

For ultra-large library screening on LIT-PCBA with 2.6 M compounds, AuroFast reached $x_i$0, exceeding the best deep-learning and docking baselines by approximately 37.6% (Zhang et al., 4 Aug 2025). Runtime per sample was reported as approximately 0.0012 s for AuroFast, versus approximately 75 s for AlphaFold 3 and approximately 30 s for AutoDock Vina, corresponding to $x_i$1 and $x_i$2 speedups, respectively. A plausible implication is that AuroFast is intended not simply as a faster surrogate, but as the mechanism by which the structural model becomes operationally usable for libraries whose size would otherwise preclude full-complex inference.

6. Prospective validation and target coverage

Prospective screening was carried out across ten disease-relevant targets: receptor tyrosine kinases TrkB and HER3; serine/threonine kinases GSK3α, GSK3β, and CDK2; the epigenetic regulator HDAC3; GPCRs CCR4 and mGluR5; and orphan GPCRs GPR151 and GPR160 (Zhang et al., 4 Aug 2025). Hit rates ranged from 7% for HDAC3 to 69% for HER3, with eight targets exceeding 20% hits at 10 µM.

Reported top IC₅₀ values included sub-nanomolar activity for TrkB at 220 pM and GSK3α at 540 pM, less than 10 nM for five targets, and less than 1 µM for eight targets (Zhang et al., 4 Aug 2025). The campaign design was also notable for scale asymmetry: compared to literature campaigns screening $x_i$3–$x_i$4 compounds, AuroBind screens approximately 30 M but tests only 30–50 compounds per target, yielding 10–100× higher hit efficiencies.

The orphan GPCR case studies were emphasized separately. For GPR151, 7 of 42 compounds were active, a 16% success rate, in BRET2 β-arrestin assays; agonists C15 and C40 had EC₅₀ values of 1.1 µM and 4.2 µM (Zhang et al., 4 Aug 2025). For GPR160, 14 of 46 compounds were active, a 30% success rate; two agonists, C45 and C05, had EC₅₀ values of approximately 1.2 µM in cAMP GloSensor assays. Structural models were reported to reveal orthosteric binding modes despite no prior ligand or crystal structure, and functional assays confirmed GPR160 modulation in liver and prostate cancer models (Zhang et al., 4 Aug 2025).

7. Limitations, interpretation, and projected development

AuroBind is presented as a generalizable framework for structure-function learning and high-throughput molecular screening, but its scope is explicitly bounded by several unresolved areas (Zhang et al., 4 Aug 2025). Dynamic or intrinsically disordered targets, exemplified by KRAS and c-Myc, await evaluation. Fine-grained fitness prediction in sparse-data regimes was identified as a setting in which enhanced preference-sampling or semi-supervised objectives could improve performance.

Several extensions were proposed within the original presentation. These include incorporation of explicit water, membrane context for GPCRs, and transition-state modeling, with the stated aim of extending applicability to covalent inhibitors and allosteric modulation (Zhang et al., 4 Aug 2025). As structural and chemogenomic databases grow, iterative self-distillation and multi-task pre-training were suggested as means to further raise structural precision and fitness sensitivity.

Two points help delimit common overinterpretations. First, the framework’s requirement of only protein sequence and ligand SMILES should not be read as evidence that all target classes have already been validated experimentally; the paper separately notes target classes that remain unevaluated (Zhang et al., 4 Aug 2025). Second, the reported screening efficiency does not eliminate the role of follow-up assays; rather, the virtual screen is coupled to a small experimental validation set, typically 30–50 compounds per target. This suggests that AuroBind is best understood as a system for prioritization at extreme scale while preserving atomic-resolution structural modeling and potency-linked ranking within a single pipeline.