Zero Rotatable Bonds, Published in 2014

On 3 November, Google DeepMind posted a walkthrough of what it calls Cheminformatics Skills for Google Antigravity. Zoi Tsangalidou, Ruijie (Kelly) Zhang, Sukhdeep Singh and Michael V. LeVine describe an agent that searches literature for a drug and its analogues, pulls thousands of compounds out of documents, filters the buffers and solvents, runs matched molecular pair analysis, trains potency models, and proposes new molecules with synthesis routes. The claim about speed is specific: a workflow that “typically takes multiple days for a single iteration” finishes “in just an hour.”

The chemistry it runs on is real and well documented. I spent the evening checking the parts that can be checked against public records. The plumbing holds up. The worked example is almost too clean, because its answer has been in print for over a decade.

The pair at the centre

The demo uses rigidification: a floppy molecule pays an entropic penalty when it locks into the shape that binds its target, so chemists add a bond to pay that cost in advance. The worked case is crizotinib, a first-generation ALK and c-MET inhibitor used in non-small-cell lung carcinoma, and lorlatinib, a third-generation inhibitor that cyclises crizotinib’s ends into a macrocycle. The post says crizotinib has five major rotatable bonds and lorlatinib has zero.

The public records agree. PubChem gives crizotinib (CID 11626560) five rotatable bonds and lorlatinib (CID 71731823) zero; running the same two SMILES strings through RDKit 2026.09.1 returns 5 and 0 under every rotatable-bond definition it offers, and puts lorlatinib’s largest ring at twelve atoms. Both molecules have thirty heavy atoms. The tying-off is not a bolt-on; it is the same atom count rearranged.

Crizotinib against lorlatinib on four computed descriptors Four paired comparisons from PubChem. Crizotinib has five rotatable bonds, lorlatinib zero. Topological polar surface area is 78 square angstroms for crizotinib and 110 for lorlatinib. XLogP is 3.7 for crizotinib and 1.5 for lorlatinib. Molecular weight is 450.3 for crizotinib and 406.4 for lorlatinib. Both molecules contain thirty heavy atoms. Bars are scaled within each descriptor, so lengths compare only inside a pair. COMPUTED DESCRIPTORS · PUBCHEM CID 11626560 AND 71731823 · RETRIEVED 3 NOV 2026 Rotatable bonds Polar surface area (Ų) XLogP Molecular weight crizotinib lorlatinib crizotinib lorlatinib crizotinib lorlatinib crizotinib lorlatinib 5 rotatable bonds — PubChem CID 11626560 0 rotatable bonds — PubChem CID 71731823 TPSA 78 Ų — PubChem CID 11626560 TPSA 110 Ų — PubChem CID 71731823 XLogP 3.7 — PubChem CID 11626560 XLogP 1.5 — PubChem CID 71731823 Molecular weight 450.3 — PubChem CID 11626560 Molecular weight 406.4 — PubChem CID 71731823 5 0 78 110 3.7 1.5 450.3 406.4 Bars are normalised within each descriptor; compare only the two bars inside one pair, never across pairs. Both molecules have 30 heavy atoms. RDKit 2026.09.1 reproduces the rotatable-bond counts, 5 and 0, from the PubChem SMILES.
The tie-off removes all five rotatable bonds and adds nothing to the atom count. It also drops XLogP by 2.2 and raises polar surface area by 32 Ų, which is the part the potency story leaves out.

The post reports the agent’s own finding this way: plotting potency against rotatable bonds, coloured by macrocycle status, “immediately reveals that macrocycles like lorlatinib tend to have fewer rotatable bonds and higher potency.” That is true of this dataset. It is also, for the headline pair, two integer fields in PubChem and a published potency table.

What the gain actually was

Lorlatinib is PF-06463922. Johnson and colleagues published its discovery in the Journal of Medicinal Chemistry in June 2014, and the paper’s own framing is not mainly about potency against wild-type ALK. It is about resistance and the brain: crizotinib works, then patients progress, either through kinase-domain point mutations such as the L1196M gatekeeper or through metastasis to the brain. The macrocycles were optimised for potency across those mutants plus central-nervous-system exposure, with “low propensity for p-glycoprotein 1-mediated efflux.”

Zou and colleagues reported the cell data in Cancer Cell a year later. In three patient-derived lines carrying mutations acquired on crizotinib, ceritinib or alectinib, PF-06463922’s growth IC50 values were 47, 63 and 23 nM, against 3,240, 1,046 and 271 nM for crizotinib.

Cell growth IC50 in three patient-derived resistant lines Cell growth IC50 values from Zou and colleagues, 2015, on a logarithmic axis. In SNU2535, carrying EML4-ALK G1269A, crizotinib is 3,240 nanomolar and PF-06463922 is 47. In MGH021-5, carrying SQSTM1-ALK G1202R, crizotinib is 1,046 nanomolar and PF-06463922 is 63. In MGH056-1, carrying EML4-ALK I1171T, crizotinib is 271 nanomolar and PF-06463922 is 23. Lower is more potent. CELL GROWTH IC50, LOG AXIS · THREE PATIENT-DERIVED LINES · ZOU ET AL 2015, FIG 2G–I SNU2535 · G1269A MGH021-5 · G1202R MGH056-1 · I1171T crizotinib PF-06463922 crizotinib PF-06463922 crizotinib PF-06463922 Crizotinib 3,240 nM — Zou et al 2015 PF-06463922 47 nM — Zou et al 2015 Crizotinib 1,046 nM — Zou et al 2015 PF-06463922 63 nM — Zou et al 2015 Crizotinib 271 nM — Zou et al 2015 PF-06463922 23 nM — Zou et al 2015 3,240 47 1,046 63 271 23 10 100 1,000 10,000 nM Lower is more potent. Axis is logarithmic, so equal bar lengths mean equal ratios, not equal differences. Lines were derived from patients who progressed on crizotinib, ceritinib or alectinib. PF-06463922 is lorlatinib.
The gap is 69-fold in the first line and 12-fold in the third. It is a resistance-mutation and brain-exposure result, which is a different claim from "rigidification raises potency."

Viarengo-Baker and Whitty went through matched acyclic and macrocyclic kinase-inhibitor pairs in the Journal of Medicinal Chemistry this July and found that the effect of macrocyclisation on potency is “variable,” while a “profound effect on selectivity is common,” with benefits also appearing in permeability, efflux ratio, blood-brain-barrier penetrance and metabolic stability. The agent’s scatter plot, drawn on a dataset assembled around lorlatinib’s own chemical series, recovers the correlation that series happens to show. Pointed at a series where rigidification cost potency and bought selectivity, the same plot would say something else, and nothing in the write-up tells me how the agent would handle that.

The part I would want measured

The division of labour is the part of the demo I’d copy. The language model orchestrates, navigates databases and reasons; deterministic tools grounded in RDKit, with connectors to ChEMBL and PubChem, do the molecular arithmetic. That split matters more in chemistry than in most fields, because, as the post says, altering one atom or bond describes a different compound altogether. Putting practice in the skills rather than the prompt is also right: the agent prefers chronological splits, training on older compounds to predict newer ones, and falls back to scaffold splits to gauge generalisability.

The caveats are stated plainly, and I’d rather read that than a press release: “Sparse data, inter-lab assay noise, and narrow dynamic ranges limit the predictive precision of even the most powerful models; computational designs are testable hypotheses, not finished drugs.”

Four stages of the demo and what independently checks each The demo's four stages run left to right: compound retrieval from literature, matched molecular pair analysis, potency models, and new macrocycle designs with synthesis routes. Retrieval can be checked against PubChem and ChEMBL records. The pair analysis reproduces a relationship published for this series in 2014 and 2015. The models are validated only on a held-out split of the same literature set. The designs carry no experimental result at all, which the post states. DEMO STAGES AND THEIR EVIDENCE · DEEPMIND POST, 3 NOV 2026 Retrieve compounds from the literature Matched molecular pair analysis Train potency models Propose macrocycles and routes CHECKABLE NOW ALREADY PUBLISHED HELD-OUT SPLIT NO RESULT YET Every structure has a PubChem or ChEMBL record to compare against, atom by atom. The cyclisation and potency relationship for this series appeared in 2014 and 2015. Validated on a split of the same retrieved set. No prospective assay is reported. Designs are hypotheses. The post says each one must be synthesised and validated in the lab. The bold stage is the one the speed claim is about, and the only one with nothing yet to check it against. Stage descriptions come from the DeepMind post; the evidence notes are mine.
Confidence drops left to right while the claims get larger. An hour saved on the first two stages is worth having; it says nothing about whether the fourth stage's molecules work.

Even so, this is a retrospective on a case whose answer sits in two PubChem records and in the training data of every current model. Reproducing a known structure-activity relationship quickly tells you the pipeline works, not that the agent can discover anything. The benchmark I want is prospective: designs synthesised, assayed, and set against what a chemist proposed from the same data in the same hour. Bender and colleagues made this point for machine-learning tools in chemistry generally in Nature Reviews Chemistry in 2022, arguing that heterogeneous evaluation makes it impossible to compare new methods. Mirza and colleagues found something adjacent in Nature Chemistry this May with ChemBench: across more than 2,700 question-answer pairs, the best models beat the best human chemists in their study on average, and still “struggle with some basic tasks and provide overconfident predictions.”

Overconfidence is the failure mode that worries me most here, because an agent that hands a chemist a synthesis route and a predicted potency has produced something that looks finished. The deterministic tools guard the arithmetic. They do not guard the judgement about which analysis to run, and that judgement is where the hour goes. I wrote last month about a grader that was also the model being graded; the structure here is gentler but related. The agent chose matched molecular pair analysis, the post notes it “had also independently suggested this analysis,” and the dataset it chose that analysis for was one it assembled.

What I would ask for next

Three things would turn this from a good demo into evidence. Run it on a series where rigidification cost potency, and publish what the agent concludes. Report the compound count and the rejection rate during retrieval, because filtering out buffers and solvents is where the silent errors live. Then synthesise a handful of the proposed macrocycles and assay them.

The offer attached to the post, a free top usage tier for 10,000 scientists over six months, should produce the raw material for exactly that. Those scientists know their own series, including the ones where the textbook move backfired. What comes back from them will tell us more than the lorlatinib walkthrough does, because for lorlatinib the answer has been known since 2014.

References

  1. Tsangalidou, Z., Zhang, R. (K.), Singh, S., LeVine, M. V., “Agentic cheminformatics with Google Antigravity,” Google DeepMind, 3 November 2026. https://deepmind.google/blog/agentic-cheminformatics-with-google-antigravity/
  2. PubChem, “Crizotinib,” CID 11626560, computed descriptors retrieved 3 November 2026. https://pubchem.ncbi.nlm.nih.gov/compound/11626560
  3. PubChem, “Lorlatinib,” CID 71731823, computed descriptors retrieved 3 November 2026. https://pubchem.ncbi.nlm.nih.gov/compound/71731823
  4. Johnson, T. W., et al., “Discovery of (10R)-7-amino-12-fluoro-2,10,16-trimethyl-15-oxo-10,15,16,17-tetrahydro-2H-8,4-(metheno)pyrazolo[4,3-h][2,5,11]-benzoxadiazacyclotetradecine-3-carbonitrile (PF-06463922), a macrocyclic inhibitor of anaplastic lymphoma kinase (ALK) and c-ros oncogene 1 (ROS1) with preclinical brain exposure and broad-spectrum potency against ALK-resistant mutations,” Journal of Medicinal Chemistry 57(11), 4720–4744, 3 June 2014. https://doi.org/10.1021/jm500261q
  5. Zou, H. Y., et al., “PF-06463922, an ALK/ROS1 inhibitor, overcomes resistance to first and second generation ALK inhibitors in preclinical models,” Cancer Cell 28(1), 70–81, 13 July 2015. https://doi.org/10.1016/j.ccell.2015.05.010
  6. Cui, J. J., et al., “Structure based drug design of crizotinib (PF-02341066), a potent and selective dual inhibitor of mesenchymal-epithelial transition factor (c-MET) kinase and anaplastic lymphoma kinase (ALK),” Journal of Medicinal Chemistry 54(18), 6342–6363, 18 August 2011. https://doi.org/10.1021/jm2007613
  7. Viarengo-Baker, L. A., Whitty, A., “Macrocycles for conventionally druggable targets: lessons from macrocyclic kinase inhibitors,” Journal of Medicinal Chemistry 68(15), 15260–15284, 31 July 2025. https://doi.org/10.1021/acs.jmedchem.4c02886
  8. Mirza, A., et al., “A framework for evaluating the chemical knowledge and reasoning abilities of large language models against the expertise of chemists,” Nature Chemistry 17, 1027–1034, 20 May 2025. https://doi.org/10.1038/s41557-025-01815-x
  9. Bender, A., et al., “Evaluation guidelines for machine learning tools in the chemical sciences,” Nature Reviews Chemistry 6(6), 428–442, 24 May 2022. https://doi.org/10.1038/s41570-022-00391-9