Multiple docking programs no longer supported outside decoy generation mode?

[prateekrai18]

Hi,

I recently updated to the newer version of DockM8 and noticed that specifying multiple docking programs now throws the following error:
dockm8.py: error: Multiple docking programs can only be specified in decoy generation mode. Use -generate_decoys=True to enable decoy mode.

I wanted to ask if there was a specific reason for restricting multiple docking programs to only decoy generation mode. In the older version, I was able to use multiple docking programs during standard virtual screening workflows (without generating decoys), and I found this feature extremely useful for comparing docking results and applying consensus approaches across different docking engines. This was one of the features I appreciated most in the older version, so I was wondering: Was this functionality intentionally removed due to technical/performance issues?
Is there an alternative workflow now recommended for running multiple docking programs in a regular screening pipeline? Are there plans to reintroduce this feature in future versions?

Thanks for developing and maintaining DockM8 — it has been very helpful for my research work.

[Tonylac77]

Hi,

Thanks for flagging this.

The change was intentional, based on internal benchmarking we did while building v1. When we compared single-engine screening against running multiple engines and combining them, the gain in enrichment and pose quality was pretty small relative to the extra compute cost (3 to 5x longer runs for marginal improvements on our test sets). Since most users are compute-limited, we figured picking the best engine for a given target gives a better tradeoff than running them all.

That's the role multiple docking programs now play in v1: in decoy generation mode, DockM8 docks actives and decoys with each requested engine and tells you which one discriminates best for that specific target. You then run the main screen with that engine. So the multi-engine capability is still there, it's just been repositioned from a "consensus" tool to a "selection" tool.

Your use case sounds different from the average user though, so a few questions :

• What library size are you typically working with?
• Did you actually validate that the consensus over poses from different sources was outperforming the best single engine on your targets? Or was it more about cross-checking poses qualitatively?
• What does "consensus" mean in your workflow?
  • Are you taking one pose from each engine and collapse/average that down into a single consensus for ranking?
  • Are you doing pose selection / rescoring with all engines combined to get one pose per compound?
• If it turns out consensus screening genuinely helps for your kind of targets, we'd consider reintroducing it as an opt-in advanced mode, but we want to base that on real evidence rather than restoring it by default.

Thanks again,
Tony