If you have ever set up a docking experiment or tuned a docking workflow for GNINA, you may have found yourself asking:

What are the “best” settings for running GNINA?

Unfortunately, there is no single objectively correct answer. The optimal settings will depend on the input data, the goal of the docking experiment, and the computational resources available. One reasonable approach is simply to use the standard GNINA settings, dock each molecule once using a single conformer, and leave it at that. The default settings already perform well in many cases. But that does not mean performance cannot be improved!

There are several ways to increase pose sampling in GNINA, often at the cost of additional compute. For example, we can increase GNINA’s exhaustiveness setting to run more independent searches, repeat runs with different random seeds, or start from multiple ligand conformations, similar to the multi-conformer docking strategy described in Inductive Bio’s strong docking baseline for AlphaFold 3. We can also retain more of the generated poses, giving downstream selection a larger candidate set to work with.

An interesting question is therefore not simply whether more sampling improves pose recovery, but which settings give the best improvement for the additional computational cost. In particular, can we improve on the standard setup while still keeping the protocol practical for larger-scale virtual-screening workflows?

During my work on the OpenBind project, I worked on the curation, benchmarking, and analysis of the first Enteroviral 2A Protease dataset release, which we recently described in a bioRxiv preprint. With a well-characterised dataset and evaluation pipeline already in place, this seemed like a good opportunity to investigate that question in a controlled setting: if we spend more compute on sampling, where and how do we actually get the most benefit? If you want to follow along or reproduce the analysis, the code and processed results are available in the accompanying GitHub repository, with the complete GNINA docking outputs archived on Zenodo.

Benchmark setup

To explore this question, I specifically used the 802 curated, non-fragment binding events from the OpenBind Enteroviral 2A Protease dataset. This provided a convenient test set for comparing different GNINA sampling strategies under the same evaluation framework, while also allowing the runtime trade-offs to be measured consistently.

For this analysis, I focused on redocking experiments, where each ligand is docked back into the receptor structure from its own experimentally determined complex. This is much more favourable than a prospective docking setting, where the appropriate receptor conformation is not known in advance. That is also what makes redocking useful for this experiment. By largely removing receptor conformational mismatch, we can focus more directly on ligand sampling and pose selection. Good redocking does not guarantee good prospective docking, but poor sampling here would not be encouraging for harder settings.

Success criteria

To determine whether a predicted pose was successful, I used the same three criteria as in the OpenBind benchmark:

• Ligand RMSD ≤ 2 Å, indicating that the predicted ligand pose is close to the experimentally observed pose.
• LDDT-PLI ≥ 0.8, requiring sufficiently good recovery of the native protein-ligand interaction geometry.
• PoseBusters-valid, requiring the predicted complex to pass physical and chemical validity checks using PoseBusters.

A pose was counted as successful only if it satisfied all three criteria.

Sampling the same ligand in different ways

For this analysis, I mainly varied two aspects of ligand sampling: the number of independent searches performed and the diversity of the ligand starting conformations.

GNINA uses Monte Carlo sampling to explore candidate ligand poses. One of its most commonly adjusted parameters is exhaustiveness, which determines the number of independent Monte Carlo chains run for each ligand. The default value is 8, so increasing exhaustiveness to 16 or 32 runs more independent searches and generally increases the computational cost accordingly.

I compared exhaustiveness values of 8, 16 and 32 across three sampling strategies:

• Single run: one starting conformer and one random seed.
• 5 seeds: five independent runs from the same starting conformer, using different random seeds.
• 5 conformers: five runs starting from different ligand conformations, using the same random seed.

For the two five-run strategies, I retained the top eight poses from each run and pooled them into a common set of 40 poses. These were then globally reranked using CNNscore, GNINA’s default score for sorting the final output poses.

This setup lets us separate two different ways of increasing sampling: repeating the stochastic search from the same starting geometry, or introducing diversity before the search even begins by using different ligand conformations.

How do the sampling strategies compare?

To start, I compared the three sampling strategies across all exhaustiveness settings and checked how often they recovered a successful pose within the Top-1 and Top-25 ranked poses (Figure 1). This gives a broad overview of how the different ways of spending additional sampling effort affect pose recovery before looking at the individual effects in more detail.

Several trends are immediately obvious. First, increasing exhaustiveness gives a modest but consistent improvement when the rest of the setup is kept the same. For example, for the Top-1 pose from a single run, the success rate increases from about 34% at exhaustiveness 8 to 37% at 16, and 40% at 32. The same overall trend is seen across the other sampling strategies.

Repeating the docking calculation with five different random seeds also consistently improves performance compared with a single run. For example at exhaustiveness 8, the Top-25 success rate increases from 45% to 56%. This pattern holds across the rank cutoffs considered: the five-seed protocol at exhaustiveness 8 matches or outperforms every single-run setting, even at exhaustiveness 32.

However, an even larger improvement comes from using five different starting conformers. At exhaustiveness 8, the five-conformer setup is the best-performing of the three strategies by a wide margin: Top-25 success reaches 85%, compared with 56% for five seeds and 45% for a single run. This suggests that introducing diversity in the ligand starting geometry contributes much more than simply repeating the stochastic search from the same structure.

Another interesting observation is the difference between the full success criterion and the less stringent RMSD + PoseBusters criterion (solid vs hatched bars). This gap is consistently smaller for the five-conformer protocol than for the other two strategies, suggesting that the poses recovered by the multi-conformer approach are not only close in RMSD, but also more often recover the native protein-ligand interaction geometry sufficiently well.

As expected, considering more ranked poses increases the chance of recovering a successful one. The jump from Top-1 to Top-25 is substantial. Extending the candidate set further to Top-40 provides almost no additional benefit across the settings tested, so I focus on Top-1 and Top-25 in this analysis. The full comparison including Top-40 is available in the accompanying repository.

How does performance improve from single runs to ensembles?

Figure 1 showed that pooling five conformer runs substantially improves pose recovery compared with either a single run or five random-seed repeats. To see how this gain develops, I next varied the number of runs included in the ensemble from one to five, for both random seeds and starting conformers.

Figure 2 shows this progression directly. Since the individual seeds and randomly generated conformers have no meaningful ordering, each ensemble size was evaluated using all possible subsets of the five available runs. Concretely, the two-run result averages over all ten possible pairs, the three-run result over all ten possible triplets, and the four-run result over all five possible combinations. The shaded region spans the lowest and highest success rates observed among those possible ensembles.

The two strategies diverge quickly as additional runs are added. At exhaustiveness 8, their starting points are very similar: Top-5 success is approximately 44% for a single seed run and 46% for a single conformer run. Adding a second conformer already raises this to around 63%, whereas two seed runs reach only about 47%. With all five runs pooled, the conformer ensemble reaches approximately 81% Top-5 success, while the five-seed ensemble remains at around 48%. The same overall behaviour is also observed for the Top-1 poses.

A likely explanation is that some starting conformations may provide a better starting point for finding the correct binding pose than others, but which conformer will be most useful is generally not known in advance and may differ between binding events. Pooling poses across several conformers therefore gives GNINA a more diverse set of starting geometries and increases the chance that at least one of them provides a favourable starting point for docking.

The gains become smaller as the ensemble grows, but the fourth and fifth conformers still provide additional pose recovery. In contrast, the seed ensembles plateau much earlier, with only modest gains beyond the first few runs. Overall, this suggests that the multi-conformer strategy benefits from complementary sampling across different starting geometries, and that varying the starting conformation provides more useful sampling diversity than simply repeating the search with different random seeds.

What does the extra sampling cost?

So far, we have looked mainly at pose-recovery performance. But for a protocol intended to scale to larger screening campaigns, another important aspect is the runtime. The different sampling strategies require substantially different amounts of compute, so I recorded the runtime for each docking calculation alongside its pose-recovery performance (Table 1).

For the five-seed and five-conformer strategies, the values in the table are the sum of the five individual docking runs. These calculations are independent and could be run in parallel if sufficient resources are available, so the values are better interpreted as total computational cost than as unavoidable wall-clock time.

The scaling is fairly intuitive. A single calculation with exhaustiveness 8 takes a median of 8.7 seconds, whereas performing five independent runs increases this to roughly 43–50 seconds. Increasing exhaustiveness further raises the cost, with the five-run protocols approaching 100 seconds per complex at exhaustiveness 32.

The important question is therefore not simply which protocol performs best, but whether the additional pose recovery is worth the additional runtime.

Where is the best performance-runtime trade-off?

While Table 1 shows approximately how long each protocol takes to run, it does not immediately tell us whether that additional cost is worthwhile. To look at both dimensions together, I plotted pose-recovery success against median compute time for each strategy and exhaustiveness setting (Figure 3).

The clearest result is that the five-conformer protocols consistently sit well above the five-seed alternatives: while their runtimes are very similar, there is a substantial increase in performance. For example at exhaustiveness 8, five seeds require around 43 seconds of compute per complex and reach 56% Top-25 success, whereas five conformers require only about 50 seconds and reach 85%. In other words, moving from five repeated seeds to five starting conformers adds relatively little computational cost while producing a much larger improvement in pose recovery.

Increasing exhaustiveness within the five-conformer strategy continues to improve performance, but with diminishing returns when considering the runtime. Moving from exhaustiveness 8 to 32 raises Top-25 success from 85% to 91%, while median compute time almost doubles from roughly 50 to 96 seconds.

This makes the trade-off particularly clear: starting-conformer diversity provides a much larger return on additional compute than simply repeating the same starting structure, while higher exhaustiveness offers smaller incremental gains at increasing cost.

Practical takeaways

For this target, the five-conformer protocol at exhaustiveness 8 looks like a particularly attractive operating point. It captures most of the available pose-recovery performance while avoiding much of the additional cost of the higher-exhaustiveness settings.

This analysis is based on a single target and a redocking benchmark, so the exact settings should not be interpreted as universally optimal. Within this benchmark, however, the trend is quite clear: varying the ligand starting conformation provides a much larger benefit than either increasing exhaustiveness or repeating the same starting structure with different random seeds. A similar benefit from multi-conformer docking was also observed in Inductive Bio’s strong docking baseline for AlphaFold 3, suggesting that the value of starting-conformer diversity is not unique to this dataset.

For larger-scale screening, where computational cost quickly becomes limiting, there is inevitably a trade-off between runtime and pose-recovery performance. But when improved performance is worth spending additional compute, the multi-conformer approach appears to be a particularly effective place to spend it. Among the settings tested here, five starting conformers at exhaustiveness 8 offered a strong balance between the additional computational cost and the resulting gain in pose recovery.

Relevant resources

• OpenBind Enteroviral 2A Protease Preprint
• Code and processed results for this analysis
• GNINA docking outputs on Zenodo
• GNINA
• Inductive Bio: A strong docking baseline for AlphaFold 3