GNINA: Deep Learning-Enhanced Docking

• GNINA is a deep learning–enhanced docking framework that augments Vina/Smina pose generation with CNN-based scoring on voxelized 3D protein–ligand complexes.
• It can be deployed as a self-contained docking engine or as a rescoring layer, demonstrating improved pose ranking and virtual screening performance over classical methods.
• Workflow effectiveness varies with configuration, showing strong results in antiviral studies, hybrid generative pipelines, and enrichment tasks while highlighting data dependence and throughput challenges.

GNINA, expanded in one benchmark as “Grid-based Neural Network Interaction Assessment,” is an open-source molecular docking program built on the 2013 Smina fork of Vina and designed to augment Vina-style search with deep-learning-based scoring. In current usage, it occupies an intermediate position between purely empirical docking and fully generative docking: it retains a conventional docking workflow for pose generation, but replaces or supplements classical scoring with a 3D convolutional neural network trained on protein–ligand complexes. Across recent benchmarks, GNINA appears both as a full docking engine and as a rescoring layer applied to poses generated by AutoDock Vina, AutoDock-GPU, DiffDock, or other pose generators (Elton et al., 5 May 2026, Jain et al., 2024, Furui et al., 24 Aug 2025).

1. Conceptual position in molecular docking

GNINA is consistently described as a Vina-derived method rather than as an entirely separate docking paradigm. One comparative study states that Smina and GNINA were both derived from AutoDock Vina, and further characterizes GNINA as using “a form of deep learning to improve Vina performance, both for pose prediction and for virtual screening applications.” In that taxonomy, GNINA is neither a pure “classical” scoring function like Vina nor a generative diffusion model like DiffDock; it is a conventional docking workflow that incorporates a CNN trained on protein–ligand complexes for rescoring or reranking poses (Jain et al., 2024).

This position is reinforced in later benchmarking work. In antiviral benchmarking, GNINA is introduced as a docking engine that “attempts to remedy” Vina’s deficiencies by replacing the standard energy function with a 3D convolutional neural network, while remaining rooted in Vina/Smina-style sampling. In virtual-screening studies, it is often treated as the canonical “ML on top of Vina” baseline against which newer affinity-prediction or docking-guided models are compared (Elton et al., 5 May 2026, Furui et al., 24 Aug 2025).

A recurrent misconception is that GNINA is a single fixed protocol. The benchmarking literature instead presents it as a family of closely related operational modes. It can be run as a self-contained docking engine that generates and ranks poses with its own CNN-enhanced pipeline, as a rescoring method operating on externally generated Vina or AutoDock poses, or as the scoring component in hybrid workflows. A plausible implication is that “GNINA” in the literature often denotes a scoring framework as much as a specific search procedure.

2. Representation, scoring, and learned signal

Methodologically, GNINA is defined by CNN-based scoring over voxelized three-dimensional protein–ligand representations. One benchmark describes it as placing the protein–ligand complex into a 3D voxel grid and applying CNN filters to learn spatial patterns associated with binding; another describes the input as a voxelized grid “analogous to a 3D image, where atoms are treated as channels.” This is contrasted with empirical or physics-based scoring by emphasizing data-driven recognition of stereochemical and spatial features (Abo-Dahab et al., 3 May 2026, Elton et al., 5 May 2026).

Within the virtual-screening literature, GNINA’s modern scoring usage is frequently expressed through CNN-derived quantities such as CNN affinity, CNN score, and their product:

$$\text{VSScore} = (\text{CNN affinity}) \times (\text{CNN score})$$

In one MF-PCBA study, GNINA ranked nine poses generated by AutoDock Vina using CNN VSScore and selected the highest-scoring pose for each ligand (Furui et al., 24 Aug 2025).

The same papers make clear that the learned component is not an auxiliary detail but the central determinant of GNINA’s utility. In the antiviral benchmark, the control condition “GNINA (Vina, no CNN)” showed essentially zero correlation with experimental pKd and near-random classification behavior, whereas “GNINA (CNN)” achieved $r = 0.302$, $\rho = 0.312$, RMSE $= 1.60$ pKd units, AUROC $= 0.652$, BEDROC $= 0.753$, and $0.0\%$ failures. This strongly suggests that GNINA’s empirical advantage derives primarily from its learned CNN scoring rather than from the underlying Vina-style search alone (Elton et al., 5 May 2026).

The training paradigm used in GNINA-centered retraining studies is pose-classification oriented. A synthetic-data study adopts GNINA’s original label scheme: poses with RMSD below $2 \text{ \AA}$ are labeled binders, poses between $2 \text{ \AA}$ and $4 \text{ \AA}$ are left unlabeled as ambiguous, and poses above $r = 0.302$0 are labeled non-binders. In that work, retraining started from the default2018 model and used 10,000 iterations with batch size 128 on a single GPU (Khiari et al., 16 Sep 2025).

3. Operating modes and workflow configurations

Recent studies evaluate GNINA under several distinct usage modes. In a fair-comparison study against DiffDock, GNINA 1.1 was run as a full docking engine rather than only as a rescorer. It generated poses, used CNN scoring for ranking, sampled 20 poses per ligand, used --exhaustiveness=32, and defined the binding site around the cognate ligand with --autobox_ligand gold-lig.mol2. Receptor preparation followed standard AutoDock tools, and evaluation used automorph-corrected heavy-atom RMSD for the top-ranked pose and the best among the top five poses (Jain et al., 2024).

In contrast, the MF-PCBA virtual-screening study treated GNINA v1.3.2 strictly as a rescoring layer on top of AutoDock Vina v1.2.7. Vina generated nine poses per ligand with exhaustiveness = 8, GNINA rescored those poses with CNN VSScore, and the ligand was ranked by the maximum VSScore over the nine poses. This protocol did not use GNINA’s internal search as the primary pose generator (Furui et al., 24 Aug 2025).

The LIT-PCBA benchmark used yet another configuration. GNINA functioned only as a rescoring method for single poses generated upstream, producing GNINA_AD for AutoDock poses and GNINA_DD for DiffDock poses. The resulting workflows were denoted AutoDock-GNINA and DiffDock-GNINA, respectively, and GNINA’s standard CNNaffinity score served as the ranking quantity (Abo-Dahab et al., 3 May 2026).

Antiviral benchmarking broadened the picture further by evaluating GNINA in three modes: autonomous docking with CNN scoring, autonomous docking with Vina-style scoring only, and a hybrid “Uni-Mol + GNINA-CNN” pipeline in which Uni-Mol generated poses and GNINA rescored them. In that benchmark GNINA was classified as “Blind,” meaning that the input was the full protein rather than a manually cropped pocket box (Elton et al., 5 May 2026).

Taken together, these studies show that GNINA is best understood as a reusable scoring-and-docking framework. It is not restricted to one canonical protocol, and performance comparisons are sensitive to whether GNINA is deployed as full docking, blind docking, cognate redocking, or rescoring.

4. Pose prediction and docking performance

In cognate redocking with a known binding site, GNINA is consistently stronger than DiffDock and generally stronger than Vina, although not usually the top performer among all docking systems. In the fair-comparison study on the PDBBind 2020 “Clean Test Set,” the authors report that at the $r = 0.302$1 RMSD threshold both Vina and GNINA exceeded DiffDock by roughly 20–25 percentage points, and that GNINA improved top-ranked pose performance over both DiffDock and Vina. On the near-neighbor subset of 191 test cases, GNINA achieved $r = 0.302$2 Top-1/Top-5 success at $r = 0.302$3, compared with DiffDock’s $r = 0.302$4 and Vina’s approximately $r = 0.302$5. The differences between DiffDock and GNINA were reported as practically and statistically highly significant, with paired $r = 0.302$6-test $r = 0.302$7 for both Top-1 and Top-5 comparisons (Jain et al., 2024).

The same study also places GNINA in a broader hierarchy. Surflex-Dock and Glide led the known-site benchmark at $r = 0.302$8 and $r = 0.302$9 Top-1/Top-5 success, respectively, while DiffDock reached $\rho = 0.312$0. GNINA was described as much better than DiffDock and better than Vina for the top-ranked pose, but still not approaching Surflex-Dock or Glide. The authors explicitly caution that GNINA’s results are “difficult to interpret” because its scoring function relies on extensive training data and may memorize binding motifs represented in the test set (Jain et al., 2024).

Blind-docking benchmarks reveal a similar pattern, but with a stronger computational trade-off. In DSDP’s 995-complex blind-docking benchmark, GNINA at exhaustiveness = 64 achieved a Top-1 success rate of $\rho = 0.312$1 at RMSD $\rho = 0.312$2 and $\rho = 0.312$3 at RMSD $\rho = 0.312$4, with runtime $\rho = 0.312$5 s per system. On the time-split PDBBind set, the corresponding values were $\rho = 0.312$6 and $\rho = 0.312$7 with runtime $\rho = 0.312$8 s, and on DUD-E they were $\rho = 0.312$9 and $= 1.60$0 with runtime $= 1.60$1 s. These were among the strongest accuracy figures in that study, but GNINA was orders of magnitude slower than DSDP’s GPU-accelerated blind docking, which operated near $= 1.60$2 s per system (Huang et al., 2023).

Recent generative docking benchmarks use GNINA as a reference point for re-docking quality as well. On the PoseBusters Benchmark Set, OMTRA reported GNINA Top-1 performance of $= 1.60$3 for poses within $= 1.60$4 and $= 1.60$5 for poses that were both within $= 1.60$6 and PoseBusters-valid; GNINA Top-5 values were $= 1.60$7 and $= 1.60$8, respectively. In that comparison GNINA was clearly surpassed by OMTRA, but it remained a substantive re-docking baseline rather than a trivial control (Dunn et al., 4 Dec 2025).

5. Virtual screening, enrichment, and application domains

GNINA’s role in virtual screening is more heterogeneous than its role in redocking. On the realistic LIT-PCBA benchmark, GNINA rescoring of AutoDock-GPU poses produced the best single-method workflow among the docking and rescoring configurations tested. AutoDock-GNINA achieved median EF1% $= 1.60$9, average EF1% $= 0.652$0, median EF10% $= 0.652$1, median ROC-AUC $= 0.652$2, and median BEDROC $= 0.652$3 $= 0.652$4–$= 0.652$5. AutoDock baseline scoring reached median EF1% $= 0.652$6, while DiffDock-GNINA reached only $= 0.652$7. The study nonetheless emphasized that even this best hybrid workflow delivered only modest early enrichment: precision at the top $= 0.652$8 was $= 0.652$9, recall was $= 0.753$0, and balanced accuracy was only slightly above random at approximately $= 0.753$1 (Abo-Dahab et al., 3 May 2026).

The same benchmark also illustrates GNINA’s target dependence. It rescued complete failures for several targets, including ESR1_ago, IDH1, and especially OPRK1, where AutoDock-GNINA reached EF1% $= 0.753$2 while the DiffDock pathway yielded EF1% $= 0.753$3. Yet GNINA also failed on some targets, notably FEN1, where GNINA EF1% was $= 0.753$4 and AutoDock-NMDN performed better. The paper’s conclusion was that no single docking method dominates across targets, even though GNINA rescoring of AutoDock poses was the most dependable single option in that benchmark (Abo-Dahab et al., 3 May 2026).

A more unfavorable picture emerged on MF-PCBA. When GNINA v1.3.2 rescored nine Vina poses per ligand using CNN VSScore, its mean AP was described as “extremely low,” only slightly better than Vina, whereas Boltzina reached mean AP $= 0.753$5 and Boltz-2 reached $= 0.753$6. GNINA did improve enrichment factors relative to Vina, but the improvement was characterized as slight and consistently inferior to the performance increase achieved by Boltzina (Furui et al., 24 Aug 2025).

In antiviral drug discovery, GNINA was the strongest docking-based method in a 15-tool benchmark on 853 compounds spanning 16 crystal structures from 10 virus species. Its CNN-enabled configuration achieved Pearson $= 0.753$7, coefficient of determination $= 0.753$8, Spearman $= 0.753$9, RMSE $0.0\%$0, AUROC $0.0\%$1, BEDROC $0.0\%$2, fail rate $0.0\%$3, and approximately $0.0\%$4 s/mol. The same codebase without CNN scoring collapsed to $0.0\%$5, $0.0\%$6, AUROC $0.0\%$7, and BEDROC $0.0\%$8. The authors therefore concluded that GNINA did best among docking approaches in that antiviral benchmark, while still trailing fine-tuned sequence-based models such as DrugFormDTA in absolute affinity prediction (Elton et al., 5 May 2026).

These application studies collectively indicate that GNINA is most reliable as a ranking and enrichment tool rather than as a calibrated affinity predictor. This is explicit in the antiviral paper, which recommends using GNINA to rank and enrich rather than to predict exact affinities, and implicit in the LIT-PCBA study, where enrichment gains were meaningful but still modest in absolute terms (Elton et al., 5 May 2026, Abo-Dahab et al., 3 May 2026).

6. Limitations, reinterpretations, and related developments

A central limitation in the recent literature is data dependence. The fair-comparison study against DiffDock explicitly notes that GNINA’s results are difficult to interpret because its CNN was trained on many protein–ligand complexes, and it is possible that its internal representation has memorized binding motifs directly represented in the test set. That concern parallels broader criticisms of train/test contamination in docking benchmarks and cautions against reading GNINA’s gains as pure out-of-distribution generalization (Jain et al., 2024).

A second limitation is throughput. The FFT-based scalar-fields paper takes GNINA as the canonical modern CNN docking tool but argues that its throughput is limited because each pose requires grid construction and a full CNN forward pass. Their alternative scoring function was designed specifically to overcome GNINA’s optimization bottlenecks, and they report similar but faster performance on crystal structures together with greater robustness on computationally predicted structures (Jing et al., 2023).

A third limitation is that GNINA’s success is strongly workflow-dependent. On MF-PCBA, GNINA as a Vina rescoring layer performed poorly relative to Boltzina and Boltz-2; on LIT-PCBA, GNINA became the strongest single method only when paired with AutoDock-GPU rather than DiffDock; and in antiviral screening it was strongest among docking methods but still substantially below the best task-specific ML affinity predictors (Furui et al., 24 Aug 2025, Abo-Dahab et al., 3 May 2026, Elton et al., 5 May 2026).

Work on retraining and extension further refines GNINA’s role. Synthetic-complex generation was used to retrain GNINA-like CNN models, and the resulting synthetic-trained models performed slightly below but closely aligned with experimental-trained models on PDBBind core and PoseBusters. The authors did not observe significant improvements in docking or scoring accuracy over experimental training data, suggesting that synthetic complexes are viable substitutes for scarce data but not yet superior augmentations (Khiari et al., 16 Sep 2025).

At the ecosystem level, GNINA is now also the reference point for complementary generative systems. OMTRA, released in the GNINA GitHub organization, is presented not as “GNINA 2.0” but as a generative model that can perform de novo ligand design and docking upstream of GNINA-style rescoring. In that framing, GNINA remains the mainstream docking and screening engine, while OMTRA adds generation of ligands and poses under pocket or pharmacophore conditioning (Dunn et al., 4 Dec 2025).

The cumulative picture is therefore specific rather than monolithic. GNINA is best characterized as a deep-learning-enhanced Vina/Smina framework whose defining contribution is CNN-based scoring on voxelized 3D protein–ligand complexes. It has repeatedly shown strong pose-ranking and enrichment performance, especially when paired with competent pose generators, and it remains a standard baseline for both conventional and AI-based docking research. At the same time, the recent literature treats it as a method whose effectiveness depends materially on benchmark construction, training-data overlap, and deployment mode, and whose current research frontier lies in hybridization with alternative search, rescoring, synthetic-data, and generative-model pipelines.