From Agentic Cheminformatics to Walking Inside the Binding Pocket

Google DeepMind recently published work on agentic cheminformatics with Google Antigravity - Cheminformatics Skills that accelerate drug discovery workflows, search compounds, and propose novel rigidified macrocycles.

Read it and one thing becomes obvious: the design layer of structural science is being automated at speed. Here's the quieter question. When an agent hands you a proposed macrocycle, what does it actually look like? Where does it bind? Which residues line the pocket? How far apart are the two atoms you're worried about?

Most of us answer that on a flat screen, rotating a structure with a mouse, squinting at depth cues that 2D rendering was never designed to give.

The gap between compute and comprehension

Structural biology has never had more data. RCSB PDB, UniProtKB, AlphaFold DB, SWISS-MODEL, PDBe - the repositories are deep, and the models are getting better. The bottleneck has quietly moved from generating structures to understanding them.

Understanding a binding pocket is a spatial task. It involves occlusion, depth, scale, and hand-eye reasoning - the exact things a monitor flattens. And it's a task where the tooling stack has barely changed in two decades: load a file, pick a representation, rotate with a mouse, take a screenshot.

I've watched this loop up close. A colleague spent an afternoon chasing a suspicious contact in a kinase model - rotating, resetting the camera, rotating again, trying to convince herself a side chain was actually reaching into the pocket and not just overlapping it in projection. The file was fine. The question was fine. The screen was the problem.

Depth was something she had to reconstruct in her head, every frame, from a flat image. That's a lot of cognitive overhead to pay before you've asked a single scientific question.

What NexoPeek does about it

NexoPeek is a Meta Quest VR molecular laboratory by Seven Sportz. It has two halves, and they map cleanly onto two different jobs.

The first half is for learning. 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 watch aspirin block COX enzymes. You see CRISPR-Cas9 edit DNA at a scale you can walk around. Every tour runs on real structures from the Protein Data Bank - no artist's rendering, no stand-in geometry.

The second half is for working. A professional PDB viewer where a smart-search bar resolves a gene name across RCSB PDB, UniProtKB, AlphaFold DB, SWISS-MODEL, and PDBe, 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. Local files - .pdb, .cif, .mol2, and more - load with electrostatic field-line maps and electron-density volumes.

Think of it as the difference between a thumbnail and a floor plan. Same data, different affordance.

The analysis bench

This is where it stops being a viewer.

• Bare hands (or VR controllers) grab, scale, and rotate the molecule.
• AI pocket prediction highlights a glowing druggable site, using PrankWeb / P2Rank cloud AI or an offline engine.
• Foldseek superimposes a homolog so you can compare folds directly.
• The Ligand Inspector traces hydrogen bonds and pulls a 3D ligand from PubChem.
• Trigger-picking two atoms reports the angstrom distance.
• Residue "pills" tapped in the sequence project a 3D halo onto the structure.

Then hit Room View, and the molecule floats into your actual room while the wall menu stays operable.

Notice what's happening at the interface level. The pocket prediction isn't a 2D heatmap you eyeball and then translate back into 3D in your head - it's a lit region of a physical object you're standing next to. The distance measurement isn't a number in a panel; it's two points you selected with your own hands, with the result tied to the geometry in front of you.

Every one of these tools collapses a translation step that used to live entirely in working memory. That's the engineering bet: spend the compute on making the spatial relationship legible, not on adding another panel of numbers.

Why this matters right now

If agents are going to propose novel chemistry, humans need faster, more intuitive ways to evaluate it. A predicted pocket is a hypothesis, not a verdict - and the fastest way to build intuition about a hypothesis is to stand inside it.

That's a pedagogy argument as much as a research one. Spatial reasoning is trainable, and it trains faster when the object is the size of a room instead of a thumbnail.

There's a nice symmetry here. The agentic layer is automating proposal. The VR layer is aimed at comprehension - the part where a human decides whether a proposed macrocycle is worth a wet-lab week. Those two halves of the pipeline are supposed to feed each other, and right now the second half is running on 1990s ergonomics.

Honest framing

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

Step Inside the Molecule. Explore the Unseen World. Download NexoPeek on Meta Quest: https://www.meta.com/experiences/nexopeek/1453797551140347/