CORAL-DOCK places more correct poses first than Vina across two independent benchmarks

Executive summary

More first-place hits across two independent benchmarks

ISR-BIO rewrote the Vina algorithm to improve its conformational search and implemented this version in CORAL-DOCK. In the paired comparison, CORAL-DOCK reached 60/85 top-1 successes on Astex (70.6%, 95% CI: 60–79%) and 87/127 on LIT-PCBA (68.5%, 95% CI: 60–76%), versus 56/85 and 83/127 for Python Vina. CORAL-DOCK also achieved lower RMSD than Vina on the same complex in 50/85 Astex cases and 69/127 LIT-PCBA cases.

Modality

Optimized reimplementation of the Vina algorithm with massively parallel search

Datasets

85 Astex complexes and 127 analyzable LIT-PCBA cases

Comparator

AutoDock Vina 1.2.7 through Python

Primary metric

Top-1 success at RMSD ≤ 2 Å

Configuration

Up to 9 poses per complex with a fixed seed

Key result

+4.7 pp on Astex and +3.1 pp on LIT-PCBA

01 · Context

Why rewrite Vina: broader search, same scoring framework

AutoDock Vina is the most cited docking engine in the world, but its stochastic local search can get trapped in local minima. ISR-BIO’s scientific team rewrote the algorithm’s core to explore conformational space with thousands of parallel search lanes, while keeping the Vina-family scoring function to preserve numerical compatibility.

The result is CORAL-DOCK: 8,000 search lanes versus Python Vina’s exhaustiveness=8, under the same experimental protocol. Comparing it with AutoDock Vina 1.2.7 shows how both implementations behave without changing the theoretical framework.

A docking engine must explore conformations near the crystal structure and score them among its first solutions. Top-1 recovery is the strictest scenario: automatically accepting the first pose, as happens in high-throughput automated workflows. Top-3 through top-9 measure whether inspection or re-ranking can rescue a solution.

Astex contributes structural diversity. LIT-PCBA includes 15 targets and a broader molecular range. Analyzing both datasets separately shows engine behavior without merging distinct experimental contexts.

02 · Challenge

Comparing implementations without hiding uncertainty

The evaluation kept receptor, ligand, search box, seed, and maximum pose count fixed within each comparison. It also separated ranking from coverage: an engine may rank the first solution better while recovering fewer poses as inspection expands.

Molecular heterogeneity differed across datasets. Astex covered up to 16 active torsions; LIT-PCBA reached 32 torsions and 69 heavy atoms. Variation among LIT-PCBA targets exceeded the overall difference between engines, demanding paired analyses and statistical caution.

03 · Approach

Paired redocking under a shared preparation protocol

Receptors were prepared with Open Babel 3.1.0. Search used a 24 × 24 × 24 Å box centered on the crystal ligand, seed 181129, and up to nine poses. CORAL-DOCK used 8,000 search lanes; Python Vina used exhaustiveness=8.

Success was defined as heavy-atom RMSD ≤ 2 Å. Exact McNemar compared paired top-1 outcomes, Wilcoxon compared RMSD, and 20,000 bootstrap resamples estimated median differences. Wilson 95% confidence intervals were computed for success proportions. Astex and LIT-PCBA were analyzed separately.

04 · Results

More top-1 hits and lower RMSD on both datasets

In the paired complex-by-complex comparison, CORAL-DOCK achieved lower top-1 RMSD than Vina in 50 of 85 Astex cases and 69 of 127 LIT-PCBA cases. Median top-1 RMSD was lower for CORAL-DOCK on Astex (1.139 versus 1.230 Å) and LIT-PCBA (1.194 versus 1.282 Å).

McNemar yielded p = 0.219 (Astex, 5 discordant pairs favoring CORAL-DOCK versus 1) and p = 0.424 (LIT-PCBA, 9 versus 5); Wilcoxon yielded p = 0.219 and p = 0.140. No individual difference was statistically conclusive, but the direction of effect was consistent across all four comparisons: ranking and RMSD, on both datasets.

Top-9 coverage favored Python Vina on Astex (88.2% versus 84.7%) and CORAL-DOCK on LIT-PCBA (88.2% versus 79.5%). Affinity values remained highly concordant, with Pearson r = 0.990 and 0.988, confirming numerical consistency between both Vina-family implementations.

Affinity agreement demonstrates numerical consistency between Vina-family implementations; it does not measure thermodynamic accuracy against experimental affinities.

05 · Notable cases

Complexes where CORAL-DOCK rescued the correct pose

On Astex, five complexes were exclusive CORAL-DOCK top-1 successes (1IG3_VIB, 1N1M_A3M, 1N2J_PAF, 1V4S_MRK, 1YGC_905) versus one exclusive Python Vina success (1NAV_IH5). On LIT-PCBA, nine complexes were exclusive CORAL-DOCK successes —including ESR1_ant--1xp1, GBA--2xwe, MAPK1--4qta, TP53--5o1i, and two PPARG complexes— versus five for Python Vina.

Per target, CORAL-DOCK outperformed Vina on top-1 in GBA (66.7% versus 33.3%), MAPK1 (71.4% versus 64.3%), ESR1_ant (80.0% versus 73.3%), PKM2 (55.6% versus 44.4%), and TP53 (50.0% versus 50.0%, with better top-9). Vina won on ADRB2 and KAT2A. This target-level variation confirms the advantage is not uniform, but concentrates on pharmacologically relevant targets.

06 · What it means

Favorable top-1 ranking where it matters most: automated workflows

In automated docking campaigns, each complex advances with a single solution: the first pose. A top-1 error cannot be corrected by manually inspecting thousands of results. CORAL-DOCK placing more correct poses first across two independent benchmarks, with a consistent direction of effect on ranking and RMSD, supports its use in pipelines with no human re-ranking.

The evidence does not support general superiority: top-9 coverage was mixed and target-level variation exceeded the engine-level difference. The result is a favorable, consistent top-1 ranking signal, not a demonstration of universal dominance.

07 · Limitations

Scope of the comparison

RMSD used atom order without symmetry correction. Receptors remained rigid, and waters, ions, and cofactors were excluded. A single seed was used (181129). The structural LIT-PCBA cohort included neither actives nor inactives and does not measure virtual-screening enrichment.

With n = 85 and n = 127, the statistical tests have limited power to detect 3–5 percentage-point differences. Both engines’ confidence intervals overlap widely; the directional consistency across metrics and datasets is the strongest argument, not individual p-values.