Search space
Candidate sets routinely run into the thousands. Full docking on all of them is rarely on the table.
Pre-docking Decision Support
Cut the docking queue before it costs you weeks.
VOX reads peptide structure the way docking eventually will: 3D, charge-aware, receptor-framed - and tells you which candidates are worth the compute. Triage in minutes, not days.
The bottleneck
Peptide libraries are big, docking budgets aren't. The decision of which candidates deserve a docking slot is usually made on intuition or a 1D score.
Cost asymmetry
A single docking pass can take hours of compute. Picking the wrong 200 wastes the whole week.
Opaque ranking
Sequence-similarity heuristics ignore 3D shape and electrostatics - exactly what binding depends on.
Inside a VOX run
Each candidate is voxelized in a receptor-aware frame and scored against a reference. Below: a live comparison between the top-ranked and bottom-ranked peptide from a recent 2104-candidate batch; 3D shape coloured by partial charge, with a difference map highlighting where the good binder packs density that the bad binder doesn't.
How it works
Four steps. No black box; every stage produces an artifact you can inspect.
1
Structure in
Drop in peptide structures (PDB / mol2). VOX handles batch ingestion and per-candidate validation.
2
Voxel encoding
Each structure is rasterized into a 3D grid with geometry and partial-charge channels, aligned to the receptor frame.
3
Receptor-aware scoring
Fast similarity scoring against reference binders, with subgroup-aware thresholds so one bad cohort doesn't poison the run.
4
Triage out
Three labels: pass, reject, uncertain. Uncertain cases are tagged, not silently dropped.
How we measure it
We don't report a single headline number. VOX is evaluated on a panel of subgroup gates - because a filter that works on average but fails on the cohort you actually care about is worse than no filter.
Separation
Score distributions for known positives vs negatives, per receptor subgroup.
Feasibility gates
High recall on positives and meaningful rejection of negatives; both required, not averaged.
Robustness reruns
Re-evaluation against larger and harder negative sets to check that gains hold under stress.
Underpowered-cohort flags
Min-sample diagnostics so we never claim performance on a subgroup too small to support it.