On November 3, 2026, Google DeepMind published a blog post announcing that Google Antigravity now ships Cheminformatics Skills, enabling agents to orchestrate drug discovery workflows, search compound libraries and propose novel rigidified macrocycle candidates. The move turns a general-purpose coding agent into a domain-equipped research executor, shifting cheminformatics from hand-written scripts to instructions plus agent-authored code that is read back and iterated. The central judgment
The timeline traces back to November 2025, when Google launched Antigravity alongside Gemini 3 as an agent-first development platform whose core interaction lets agents work across editor, terminal and browser and produce reviewable artifacts. Google has since extended that surface through a Skills mechanism: a skill is essentially a documented bundle of deterministic tools and script templates that an agent loads on demand rather than something baked into model weights.
The November 3, 2026 post introduces Cheminformatics Skills. Per the blog, the package covers three task families: orchestration of drug discovery workflows that chain target information, library screening and property calculation into an executable pipeline; compound search across substructure, similarity and constraint-based queries; and macrocycle design, where the agent proposes rigidified macrocycles, adding
Why now. First, since late 2025 skills and connectors have become the default route for wiring general models into vertical workflows: Anthropic shipped Agent Skills, OpenAI expanded its connector ecosystem, and vendor competition moved from whether a model can do something to whether it can call professional tools and hand back auditable results. Second, medicinal chemistry happens to have an unusually mature toolchain. RDKit and Open Babel supply open-source parsing, descriptors and substructure search, AlphaFold-lineage models solved structure prediction, yet much of the glue between steps is still hand-written scripts, which is exactly the seam an agent can occupy. Third, macrocycles
Liu Gong's take: within 12 months, no more than five major pharma companies will actually wire general-purpose agents like this into lead optimization. Within 18 months, however, the skill package format will likely degrade from natural language instructions into a set of version-controlled, unit-tested deterministic tool calls, because regulated R&D needs reproducibility, not conversation.
From a technical standpoint, the chemistry tasks agents handle well share one property: verifiable intermediate results. Whether a SMILES parses in RDKit, whether a substructure match fires, whether LogP lands in range, all have deterministic checks, so a model error surfaces immediately and the loop holds. Antigravity's real contribution is turning those functions into a service layer callable in plain language, sparing chemists from memorizing APIs. But developability judgments lean on negative data: which scaffolds cannot be made, which polymorphs are unstable, which compounds underperform in vivo. That data sits in pharma and CRO electronic lab notebooks and is nearly absent from public corpora. The skills therefore accelerate enumeration of the feasible space, not adjudication of it.
On the business side, the moat at the skill layer is far shallower than at the model layer. A cheminformatics skill set is essentially tool documentation plus script templates; copying it costs a chemist's writing time, not training compute. Google gains a doorway rather than a wall: keep a chemist's working day inside Antigravity, then bill Gemini inference behind it.
For the industry, pressure lands first on two groups. SaaS vendors selling screening and ADMET prediction will see interface value diluted by one-sentence invocation. CROs face a pricing problem: if an agent finishes two weeks of enumeration and triage in an hour, per-compound quotes need rewriting. An observation from Liu Gong's industrial inspection work applies directly: vision model accuracy stopped being the bottleneck long ago, while consistency of acceptance criteria still stalls production lines. In chemistry the analogue is whether developability standards can be written as executable, auditable rules. Rigidified macrocycles are the sharpest case, since conformational entropy gains must be booked against molecular weight, polar surface area and synthetic step count. A pretty structure is not a pass.
- Approved macrocyclic drugs number in the dozens, under 2% of all approved small molecules, while a single macrocyclization typically adds roughly 100–300 Da against an empirical oral small-molecule ceiling near 500 Da, so rigidification must subtract on both ledgers at once.
- Publicly enumerable chemical space spans billions to tens of billions of compounds (ZINC around 2 billion, Enamine REAL in the tens of billions), and exhaustive virtual screening costs thousands of dollars in compute; an agent's search strategy decides whether it screens ten thousand or a hundred million, a two-to-three order-of-magnitude cost gap.
- Oral cyclic peptides generally show membrane permeability in the
For the industry, the next three months will reshape the interface layer of cheminformatics: teams that write RDKit scripts fluently lose marginal advantage, teams that can define which questions are worth asking gain, and a crop of third-party skill packs will appear around Antigravity. Over a one-year horizon, pricing at CROs and SaaS vendors will be forced to adjust, with per-compound screening services hit first and per-validated-hit pricing a plausible successor.
For everyday users, the near-term benefit is concrete: bench medicinal chemists can run library searches, property calculations and macrocycle scaffold enumeration without writing code, cutting repetitive work. Yet proposed
Questions worth exploring:
- Will skill packs displace the conventional cheminformatics software interface, and what interface value survives
- The conformational entropy ledger for rigidified macrocycles: can proposed structures be made, and at what cost
- Reproducibility and audit trails once general-purpose agents enter regulated R&D workflows
Further reading: 2026 Frontier LLM Comparison: A Practical Guide — a durable, evergreen reference. Bookmark it.
This item comes from DeepMind Blog (published 2026-11-03 17:01:49). Wenfeng Radar provides a Chinese/English summary and commentary on overseas AI developments; the original content is owned by its author.
Daily aggregation of overseas AI developments with in-depth analysis. Bookmark us to never miss an important signal; back to homepage.