Valve：仍相信 VR，但內容與用戶仍陷「先有雞還是先有蛋」的困局

Valve Says VR's Problem Is Content. We Shipped a Spatial Viewer for a 50-Year-Old 3D Database.

At a Steam Indie Connect event in Korea, Valve said it still believes in VR while naming the field's central obstacle: a chicken-and-egg dilemma between content and users (Road to VR). Platform holders rarely say this out loud, so it's worth taking at face value.

But there's a category of VR content that doesn't need to be invented: scientific data that is already natively three-dimensional. The PDB, AlphaFold DB, UniProtKB, PDBe, SWISS-MODEL — one of the largest collections of spatial data in existence, consumed almost entirely through a flat rectangle.

We built NexoPeek (Meta Quest, by Seven Sportz) to be a second viewing medium for that data. Here's the developer-facing breakdown.

Architecture at a Glance

Data layer — multi-source resolution
A single search input resolves a gene name across five remote resources:

Source	Role
RCSB PDB	Experimental structures
UniProtKB	Sequence → protein identity
AlphaFold DB	Predicted structures
SWISS-MODEL	Homology models
PDBe	European PDB mirror + annotations

Resolution means handling heterogeneous identifiers, differing coordinate conventions, and inconsistent metadata across five APIs. If you've built on any two of these, you know the normalization layer is the actual work.

Rendering layer — Mol* presets
One-tap representation cycling: cartoon, ball-and-stick, spacefill (CPK), molecular surface, B-factor heat, tube trace. Mol* handles the heavy lifting; the work here is preset state management and keeping representation switching fast enough to feel like a toggle rather than a load.

Local file path
.pdb, .cif, and .mol2 load directly, with electrostatic field-line maps and electron-density volumes rendered alongside the primary structure. Volumetric + mesh rendering in the same scene, on a mobile-class SoC, is where the perf budget actually goes.

The Interaction Model

Bare-hand grabbing, scaling, and rotating (or VR controllers as fallback). Three interaction patterns worth noting:

1. Atom-pair picking with trigger
Two atoms selected, distance reported in angstroms. This is a hit-testing problem where the target is sub-pixel at realistic zoom levels. We had to trade raw precision against reachability — you can't require a 1-pixel-accurate raycast on a 0.5nm atom.

2. Residue pills → 3D halo projection
Tapping a residue in the 2D sequence panel projects a 3D halo onto the corresponding position in the structure. This is a bidirectional binding problem: sequence index ↔ residue identifier ↔ spatial coordinates. Mapping is not always 1:1 (missing residues, insertion codes, multi-chain assemblies), so the projection layer carries the ambiguity handling.

3. AI pocket prediction
Two execution paths:

Cloud: PrankWeb / P2Rank endpoints
Offline: local inference engine

The offline path matters for headset ergonomics — network round-trips inside a VR session feel much worse than on a laptop, and cold-start behavior differs.

The Analysis Bench, Listed Honestly

AI pocket prediction — highlights candidate druggable sites (PrankWeb / P2Rank cloud, or offline engine)
Foldseek superposition — structural homolog alignment for comparison
Ligand Inspector — traces hydrogen bonds, pulls a 3D ligand from PubChem
Atom-pair trigger picking — angstrom distance readout
Residue pills — sequence-to-structure 3D halo projection

Room View

A toggle that floats the structure into the user's physical room while the wall menu stays operable. The interesting problems here aren't visual — they're passthrough compositing, depth sorting between real and virtual geometry, and input routing so the wall menu remains interactive while a large mesh occupies the same visual field.

What This Is Not

NexoPeek is educational and research-visualization software — not medical advice, diagnosis, or treatment. Pocket prediction and structural alignment are triage and intuition aids: they support and never replace validated scientific pipelines and professional judgment. It is not a PyMOL or ChimeraX replacement; it's a different viewing medium with different tradeoffs.

Try It

Meta Quest: https://www.meta.com/experiences/nexopeek/1453797551140347/

If you've built multi-source scientific data pipelines or volumetric rendering on mobile XR, I'd genuinely like to compare notes in the comments — particularly on the identifier normalization layer and the offline inference path.