AI Is Finding Candidate Molecules Faster Than We Can Inspect Them

DeepMind's new blog post — "Accelerating the search for antimicrobial peptides with Google Antigravity" — is a good reminder that AI is now generating candidate molecules faster than any human team can inspect them. Antimicrobial resistance is one of the most serious long-term threats in medicine, and computational pipelines are finally producing peptide candidates worth looking at. The bottleneck has shifted downstream.

Because "worth looking at" is the operative phrase. A model can rank a thousand peptides by predicted affinity. It cannot tell you whether the pocket geometry actually supports the interaction, whether the hydrogen-bond network is plausible, or whether a homolog in the same family already does something similar. Those judgments still belong to a human with structural training, and right now that human is doing it on a flat screen.

Why flat screens are the wrong shape for this work

A protein is a three-dimensional object with depth, occlusion, and spatial relationships that carry meaning. A 2D projection flattens all of that. Rotating a structure with a mouse gives you the illusion of depth while removing your ability to perceive it. Crystallographers and drug discovery teams have adapted brilliantly to this limitation — but adapting to a limitation is not the same as not having one.

There is an engineering framing here too. Rendering a protein at human scale in VR means solving a different problem than drawing it in a viewport: stereoscopic depth cues, hand-tracked manipulation of a mesh with tens of thousands of atoms, and a UI that has to stay readable while a molecule floats around you. The interesting design decision is not "3D vs 2D" — it is where you put the controls when the data is the room.

Think about the render budget. A mid-sized protein runs tens of thousands of atoms; a large assembly runs into the hundreds of thousands. Stereo rendering means drawing every frame twice, once per eye, and holding 72–90 fps or you get judder that breaks the depth illusion entirely. That pushes you toward instanced geometry, level-of-detail swaps as atoms move into the foreground, and representation presets that change the vertex load by an order of magnitude — a tube trace is cheap, a spacefill CPK with per-atom radii is not. The molecule is the same; the cost of showing it is not.

Then there is the interaction model. On a desktop you own one pointer. In a headset you have two tracked hands, no keyboard, and no cursor. A tap target that worked at 12px on a monitor is now a physical gesture in space, and it has to resolve against a mesh that is itself moving. That is why a lot of "VR molecular viewer" demos look great for thirty seconds and fall apart the moment you try to do actual work.

What NexoPeek is

NexoPeek is a Meta Quest VR molecular laboratory by Seven Sportz. It has two halves that serve two different audiences.

The first is guided learning: 12 narrated tours across 6 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 stand in the middle of glycolysis as glucose breaks down. You watch aspirin block a COX enzyme. You see CRISPR-Cas9 edit DNA. For biotechnology students, educators, and curious newcomers, this is the difference between reading about a mechanism and standing inside it.

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

The multi-source search is worth a note on its own. Gene nomenclature is a mess — the same protein carries different identifiers across RCSB, UniProt, and AlphaFold, and synonyms drift between organisms. Resolving a query across all of them before loading means you spend less time guessing which entry is the right one and more time looking at it. Small thing, but it is the kind of small thing that eats an afternoon.

The analysis bench

This is where it connects to the antimicrobial peptide story. On the analysis bench you use bare hands (or VR controllers) to grab, scale, and rotate the molecule. AI pocket prediction — PrankWeb / P2Rank cloud AI or a built-in offline engine — highlights a druggable site. A Foldseek pipeline superimposes a homolog so you can reason about domain function conservation and drug-swapping hypotheses. The Ligand Inspector scans hetero-compounds, traces hydrogen-bond networks, and fetches a 3D ligand from PubChem. You pick two atoms with the trigger and read the distance in angstroms. Residue "pills" tapped in the sequence project a 3D halo onto the structure, and you can isolate a single chain when the assembly gets noisy.

That is an inspection workflow designed for the shape the data actually has. The two-atom distance readout is the one I keep coming back to: on a flat screen you are eyeballing a measurement through a projection, and in stereo you are placing two points in space and trusting the depth. Same number, different confidence.

Honest boundaries

Pocket prediction and structural alignment are aids to human judgment. They do not replace validated scientific pipelines, and NexoPeek is educational and research-visualization software — not medical advice, diagnosis, or treatment. Verify findings against primary literature and validated methods before acting on them, and keep sessions comfortable with breaks.

It is also worth being clear about what a headset is not. It is not a replacement for PyMOL in a scripting workflow, and it is not where you run a batch alignment over a thousand structures. It is a place to look at one structure carefully, which is exactly the step that is now the bottleneck.

The larger point

If AI is going to accelerate discovery, the human verification step needs to accelerate too — or at least stop being the slowest part of the loop. Spatial computing won't replace the wet lab or the validated pipeline. But it can make the inspection step faster, more intuitive, and more memorable for the students who will run those pipelines in ten years.

Step inside the molecule: https://www.meta.com/experiences/nexopeek/1453797551140347/