DeepMind Built an Agent for Drug Discovery. The Hard Part Is Still the Structure.

Google DeepMind published a blog post on agentic cheminformatics built on Google Antigravity. The short version: an AI agent that accelerates drug discovery workflows, searches compounds, and proposes novel rigidified macrocycles. If you work anywhere near structure-based design, that headline lands.

Here is what it does not do. It does not tell you whether the pocket those macrocycles are aimed at is real.

The gap between a proposal and a pose

An agentic workflow compresses the ideation loop. Compound search, scaffold rigidification, candidate proposal — work that used to take a medicinal chemist a week now happens in a session. That is a genuine shift, and it deserves the attention it is getting.

But every proposed molecule still has to be evaluated against an actual three-dimensional structure. And that evaluation is not a lookup. It is a set of judgments:

• Is this PDB entry at a resolution that supports the claim being made about it?
• Do the B-factors in the binding region suggest a rigid, well-ordered site — or a floppy loop that was modeled optimistically?
• Does the electron density actually support the ligand pose, or is the ligand placed by inference?
• Is this pocket druggable, or is it a crystal contact that happens to look inviting?
• Does a homolog with a bound ligand exist that would tell us more than the apo structure does?

None of those questions are answered by faster generation. They are answered by people who can read a structure — and by tools that let them read it well.

From an engineering standpoint, this is a tooling problem more than a modeling one. The pipeline that matters is the last mile: take an identifier, resolve it across heterogeneous repositories, fetch the coordinates, render them at a fidelity that supports a decision, and keep the human in the loop with the evidence visible. That last mile is where most structure tooling gets thin, because it is unglamorous — identifier resolution, format normalization, representation switching, and provenance.

Why we built a VR molecular laboratory

NexoPeek is a Meta Quest VR molecular laboratory from Seven Sportz. It has two halves, and they serve different parts of this problem.

The first half is 12 narrated tours across 6 pillars: Biotechnology Basics, Foundations of Life, Pathways & Interactions, Medicines in Action, Genetic Technologies, 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 watch CRISPR-Cas9 edit DNA. This is the structure literacy layer — the part that turns a 2D diagram into something a student has actually inhabited.

The second half is the professional PDB viewer. A smart-search bar resolves a gene name across RCSB PDB, UniProtKB, AlphaFold DB, SWISS-MODEL, and PDBe, and the structure loads at human scale around you. Under the hood that means smart keyword cascades and UniProt gene synonym resolution with automated model validation — the boring plumbing that decides whether "did I get the right structure?" is a five-minute question or a five-second one. One-tap Mol* presets cycle representations: cartoon, ball-and-stick, spacefill CPK, molecular surface, B-factor heat, tube trace — six presets derived from MolViewSpec concepts, which is the pragmatic call: don't make a researcher hand-tune a scene graph for a quick sanity check.

Local files work too: .pdb, .cif, .mol2, .gro, .xyz, .dx, .xtc, and .py load with electrostatic field-line maps, custom bond definitions, and electron-density volumes. Mixed-format ingestion is a small feature on paper and a large one in practice, because real labs do not have tidy homogeneous data.

Then there is the analysis bench. Bare hands — or VR controllers — grab, scale, and rotate the molecule. AI pocket prediction highlights a candidate druggable site using PrankWeb / P2Rank cloud AI or a built-in offline engine. Foldseek superimposes a homolog so you can compare binding regions side by side. The Ligand Inspector scans hetero-compounds, traces hydrogen bonding networks, and pulls a 3D ligand from PubChem. Trigger-pick two atoms and it reports the distance in angstroms. Residue "pills" tapped in the sequence project a 3D halo onto the structure, so sequence and shape stop being two separate documents.

There is also a Room View passthrough toggle that floats the molecule into your actual room while the wall menu stays operable — useful for teaching, useful for lab meetings, useful for the moment when you need to point at something and say "this loop, right here."

The handoff that matters

The most interesting question raised by the Antigravity work is not "will agents replace structural biologists." It is "what does the human do in the loop now that generation is cheap?"

The answer, we think, is interpretation and instruction. Somebody has to look at the proposed macrocycle in the context of the real pocket and decide. Somebody has to teach the next cohort what that decision feels like. Somebody has to hold the validated pipeline and the professional judgment that a prediction aid is not allowed to replace.

NexoPeek is built for that layer. Not to generate candidates — to make the structure they target legible, at a scale where you can walk around it.

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. Its analysis aids — pocket prediction, alignment, ligand inspection — support and never replace validated scientific pipelines and professional judgment.