HelixVS

HelixVS is a structure-based virtual screening platform that inserts a deep learning scoring model between classical molecular docking and hit selection. Virtual screening is limited by a well-known tension: physics-based docking programs such as AutoDock Vina are fast enough to sweep multi-million-compound libraries but their empirical scoring functions frequently rank decoys above true binders, while deep learning drug–target affinity models score more accurately yet generalize unevenly to targets outside their training distribution. HelixVS resolves the tension by staging the two technologies rather than choosing between them — docking generates and prunes poses cheaply, and a learned scorer re-ranks only the survivors.

The platform comes from the PaddleHelix team at Baidu PaddleHelix, the group behind the structure prediction models HelixFold-Single and HelixFold3. Unlike those models, HelixVS is distributed as a hosted service: a free web version with limited compute, plus private deployment for organizations with data-security requirements. The same pretrained scorer is applied across roughly a hundred targets and four internal drug-development pipelines without per-target retraining, and a companion module, HelixVS-Syn, designs new molecules from reference scaffolds and feeds them into the same screening workflow.

#Key Features

• Three-stage screening pipeline: Docking, deep learning re-ranking, and optional binding-mode filtering are applied in sequence, with distributed sorting between stages so that expensive computation is spent only on promising molecules.
• Learned pose scoring: The stage-2 scorer re-ranks docked poses more accurately than a docking scoring function, considering multiple isomers and multiple retained conformations per compound simultaneously.
• Interaction-aware filtering: An optional stage restricts hits to molecules that contact specified residues with specified interactions, then clusters the remainder so the shortlist handed to chemists is structurally diverse.
• Scaffold-based molecular design: HelixVS-Syn generates novel compounds from a known scaffold using retrosynthetic knowledge and synthesis-guided modification, so designed molecules remain synthetically accessible.
• CPU-scale throughput: The pipeline runs on commodity CPU infrastructure rather than large-VRAM GPUs, exceeding 10 million molecules screened per day.

#Technical Details

Stage 1 docks compounds with AutoDock QuickVina 2 and deliberately retains multiple binding conformations rather than a single best pose. Stage 2 feeds the lowest-ΔG poses to an affinity scoring model built on RTMscore and strengthened by data augmentation with additional co-crystal structures covering nearly all ligands in the PDB. Stage 3 applies the optional interaction filter and clustering. On the DUD-E benchmark — 102 proteins from 8 families, 22,886 actives, and 50 property-matched ZINC decoys per active — HelixVS reaches an enrichment factor of 44.205 at 0.1% and 26.968 at 1%, against 17.065 and 10.022 for Vina, 37.842 and 24.346 for Glide SP, and 25.958 and 15.848 for KarmaDock. Throughput is roughly 4,000 molecules per day per CPU core versus about 300 for Vina, with screening costs as low as 1 RMB per thousand molecules. On COMT, a Parkinson's disease target, HelixVS was the only method tested that recovered actives at all, with an EF at 0.1% of 61.51. On a 10-target DUD-E subset chosen for tractable GPU cost, HelixVS scored 35.306 at EF 0.1% against 33.767 for Boltz-2 in binary mode and 19.500 in affinity mode, though Boltz-2 in binary mode led at EF 1% (27.753 versus 24.413). For molecular design, HelixVS-Syn compounds targeting CDK5 scored higher than screened compounds on synthetic accessibility (0.708 versus 0.658) and novelty (0.641 versus 0.615).

#Applications

The platform is aimed at medicinal chemists who need to triage large commercial libraries without running computational tooling themselves; automated protein preparation, built-in purchasable libraries, and SMILES upload are all exposed through a web interface. Four internal campaigns illustrate the scope: a 7.8-million-compound screen against a previously unreported pocket on the CDK4/6–CCND1 interface yielded 6 of 40 purchased compounds that disrupted complex formation in a bimolecular fluorescence complementation assay; a 200,000-compound screen against the TLR4/MD-2 interface produced 6 active candidates from 103 tested, 2 with nanomolar activity; a 30,000-compound screen of the cGAS ATP pocket produced 17 actives from 96 assayed, 10 below 10 µM and one nanomolar; and a 10-million-compound screen against NIK produced one micromolar-IC50 scaffold from 7 candidates.

#Impact

HelixVS is notable less for a new architecture than for demonstrating that a carefully staged hybrid pipeline, run on CPUs, can outperform commercial docking software and GPU-hungry co-folding models on screening enrichment at a fraction of the compute cost — the paper estimates its total cost at roughly 1/500 that of a Boltz-2 screen. The wet-lab results across four pipelines, including two protein-protein interaction interfaces that are conventionally considered hard to drug with small molecules, give the enrichment numbers experimental grounding that benchmark-only screening papers lack. Two limitations bound its reach: the work is a technical report rather than a peer-reviewed study, and neither code nor model weights have been released, so the platform can be used but not independently reproduced, audited, or extended.