AI Just Sped Up Antimicrobial Peptide Discovery. The Human Bottleneck Is Still Depth.

Google DeepMind's blog post on accelerating the search for novel antimicrobial peptides with Antigravity for Science is the kind of story that deserves the attention it's getting. Antimicrobial resistance is one of the genuinely hard problems in medicine, and using AI to search sequence space faster than any lab could is a real advance.

But watch what happens after the model proposes a candidate.

Someone has to look at it.

Not just the sequence — the structure. Where does the peptide sit? Which residues line the pocket? Is that cavity actually druggable, or is it a shallow groove that will never hold a small molecule? That judgment call is still made by a human being, and it is still made by rotating a 3D object on a 2D screen.

The depth problem nobody lists as a problem

Structural biologists have adapted so thoroughly to flat representations that we've stopped noticing the tax. You load a structure, spin it with a mouse, and mentally reconstruct the third dimension from shading and occlusion. For a quick sanity check, that's fine. For a real binding-pocket inspection — the kind where you need to feel how deep a cavity goes and how a ligand sits inside it — you are constantly translating.

Virtual reality removes the translation. That's the entire premise behind NexoPeek, a Meta Quest VR molecular laboratory by Seven Sportz.

What it actually does

There are two halves to the product, and both matter.

Guided tours. Twelve narrated tours across six pillars: Biotechnology Basics, Foundations of Life, Pathways & Interactions, Medicines in Action, Genetic Technologies, and Green Biotech. You step inside hemoglobin and watch oxygen bind. You follow glucose through glycolysis. You stand at human scale while aspirin blocks a COX enzyme. You watch CRISPR-Cas9 edit DNA, or inspect engineered plastic-degrading enzymes developed for environmental cleanup. For a biotechnology student meeting these concepts for the first time, this is a different order of understanding than a textbook figure.

A professional PDB viewer. A smart-search bar resolves a gene name across RCSB PDB, UniProtKB, AlphaFold DB, SWISS-MODEL, and PDBe — with UniProt gene synonym resolution and automated model validation — then loads the structure at human scale. One-tap Mol* presets cycle through cartoon, ball-and-stick, spacefill CPK, molecular surface, B-factor heat, and tube trace representations. Local files — .pdb, .cif, .mol2, .gro, .xyz, .dx, .xtc, .py — load directly, with electrostatic field-line maps, custom bond definitions, and electron-density volumes.

Then the analysis bench, which is where it stops being a demo and starts being a tool. You grab, scale, and rotate the molecule with bare hands or VR controllers. AI pocket prediction — via PrankWeb / P2Rank cloud AI or a built-in offline engine — highlights a druggable site. A Foldseek pipeline superimposes a homolog to test domain-function conservation or drug-swapping hypotheses. The Ligand Inspector traces hydrogen-bond networks and pulls a 3D ligand straight from PubChem. Trigger-picking two atoms reports the distance in angstroms. Scrolling the per-chain amino acid sequence, tapping residue "pills," projects a matching 3D halo onto the structure — a small detail that turns sequence-space browsing into spatial navigation.

A note on the architecture of the work

There's an engineering assumption worth stating plainly: this is not a headset-native replacement for a computational pipeline. It's an ingestion-and-interpretation layer. Structures come in from live repositories or local files; the heavy analytics (pocket detection, alignment, ligand fetch) route to cloud services or run on a local engine; the visualization layer sits on Mol* presets derived from MolViewSpec concepts. That separation is deliberate. It keeps the client light enough to run comfortably on Quest hardware while letting the analysis stack stay swappable as better models ship.

The trade-off is honest: latency for anything cloud-routed, and a rendering fidelity ceiling that won't match a workstation GPU. In exchange you get spatial reasoning you simply cannot get through a monitor, and a Room View passthrough mode that floats the molecule into your actual physical space while keeping the wall menu operable.

Why this matters right now

The AI-driven discovery wave — antimicrobial peptides, de novo binders, structure prediction at scale — is generating candidates faster than any human pipeline can interpret them. That's a wonderful problem to have, and it makes interpretation throughput the next real constraint.

You don't fix that by making people read faster. You fix it by removing the mental conversion step between the data and the thing the data describes.

From classroom to candidate

A biotechnology student exploring ATP. An educator teaching CRISPR without a textbook diagram. A structural biologist rotating a binding pocket between two fingers before deciding whether to chase a homolog. Same headset, same structure, three very different depths of question.

Curiosity to candidate therapeutics, in one headset.

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

---

NexoPeek is educational and research-visualization software. It is not medical advice, diagnosis, or treatment. Analysis aids such as pocket prediction and structural alignment support — and never replace — validated scientific pipelines and professional judgment. Verify findings against primary literature and validated methods before acting on them.