Zipf for fillers inside a frame, chemistry side: substituents, R-groups, and which positions get filled,
What’s claimed and how well supported
- Labels on this page
- SPECULATIVE ×4
- Strongest label used
- SPECULATIVE
- Main caveat (quoted)
- Setting them beside the English verb-argument-construction (VAC) results is a ◆ SPECULATIVE comparison inside Grant's frame.
Counted automatically from the labels printed on this page. Counts are a reading aid, not a grade of the page.
Terms on this page (1), plain-language definitions
- Dopant
- Intentional dilute impurity meant to tune electronic/optical/catalytic behavior without replacing the host as majority phase.
Definitions quoted from the Chem for linguists primer glossary.
Date: 2026-10-06 PT (Navajo lexicalization push, midday)
Status: Scout. Every chemistry fact below is same-domain description from the sources opened (cheminformatics and medicinal-chemistry database studies, plus one textbook-level secondary page). Setting them beside the English verb-argument-construction (VAC) results is a SPECULATIVE comparison inside Grant's frame. It is not a correlate and not a joint field.
Audience: Grant (linguistics; less chemistry). Plain wording; chemistry terms glossed once.
Does not: claim that a substituent is an argument, that a ring position is a theta slot, or that a directing group is a theta role; supply a stand-in for a theta role; invent Navajo forms, glosses, counts or slopes; fit any curve to data the papers did not fit; revise Bridge A.
Why this note. This morning's note (Zipf at the frame) found that the two literatures rank opposite halves. English VAC studies hold a role-ordered frame fixed and rank the fillers (the verbs that go in it). The CAS framework study ranks the frames (molecular skeletons) after deleting the fillers and the attachment points. Its open question 1 asked for the missing chemistry half: are the groups that fill a given position on a given framework power-law distributed, and selective to that frame, the way verbs in a VAC are?
Glossary (once). A substituent or R-group is a small group of atoms hanging off a larger molecule by one bond (methyl, –CH₃; hydroxy, –OH; chloro, –Cl). The scaffold or core is the part that stays fixed across a series. A substitution site or position is the place on the core where a substituent is attached. An analogue series is a set of molecules that share a core and differ in their substituents.
What is new (one sentence). Chemistry does count fillers, and the counts are long-tailed: 849,574 different substituents in 3 million commercial molecules, about 64% seen only once (Ertl 2003), and a "typical power law (or 'long tail')" in natural products (Ertl 2021 preprint). But no opened source prints a slope or exponent for substituents, and none counts fillers position by position on one framework, so the VAC-style test is still undone. The nearest object with slots is the benzene ring. There, which positions get filled, with the identity of the filler ignored, is top-heavy (four of twelve patterns make up 76% of benzene rings in approved drugs; Nilova & Stuart). The authors tie that skew to a textbook rule: the kind of group already on the ring (electron-donating or electron-withdrawing), not which atom it is, picks the slot for the next group.
- 1. Sources opened this run
- 2. Question (a): are substituent frequencies power-law, and with what slope?
- 3. Question (b): is the filler distribution measured per position on a given framework?
- 4. Question (c): does anything rank fillers by the slot they take, as distinct from atom identity?
- 5. Does this close, narrow, or leave open the "opposite halves" gap?
- 6. What this means for Grant's frame, and for the Navajo question
- 7. Separation
- 8. Open questions
- 9. Sources that failed this run
1. Sources opened this run
| Source | How read | Used for |
|---|---|---|
| Ertl, P. 2003, "Cheminformatics analysis of organic substituents…," J. Chem. Inf. Comput. Sci. 43: 374–380 | Author reprint PDF (peter-ertl.com), text layer pp. 374–380; page image p. 375 (Fig. 1, Fig. 2) checked | Substituent definition; 849,574 unique substituents; 50 in >1% of molecules, 438 in >0.1%, ~64% singletons; growth curve (Fig. 2) and its log–log fit r² = 0.998; drug vs non-drug overlap (243 of top 500); "simply the most common organic substituents" |
| Ertl, P. 2021 ChemRxiv preprint, "Substituents of life: the most common substituent patterns present in natural products" (published 2022, Bioorg. Med. Chem. 54: 116562) | Preprint text as rendered by exa.ai's library copy. ChemRxiv's own PDF link returned a bot-check page to curl and timed out in WebFetch. Journal version not opened. Figures not seen. Author's data list npsubstituents.txt downloaded |
222,000 natural products; 16,977 unique substituents; "typical power law (or 'long tail')"; 24 in >1%, 157 in >0.1%, 8,963 (52.8%) singletons; top percentages |
| Takeuchi, K., Kunimoto, R. & Bajorath, J. 2021, "R-group replacement database for medicinal chemistry," Future Sci. OA 7(8): FSO742 (PMC8288244) | PMC full text (HTML saved); Zenodo record 4741973 file list via API | 17,254 analogue series; 61,312 indexed substitution sites; 50,545 R-groups; site-specific sampling then "all site-specific replacements were combined"; ranking by node degree; what the deposition contains |
| O'Boyle, N. M., Boström, J., Sayle, R. A. & Gill, A. 2014, "Using matched molecular series as a predictive tool to optimize biological activity," J. Med. Chem. 57: 2704 (PMC3968889) | PMC full text (HTML saved) | Matched series = same scaffold, different R-groups at the same position; five most common R-groups = 32.5% of training data; preferred activity orders (Tables 1–2); sampling-bias caveat |
| O'Boyle, N. M. & Sayle, R. A. 2019, "A medicinal chemistry based measure of R group similarity," poster, NextMove Software (Sheffield 2019) | PDF text layer | R-group frequency used as a proxy for the order in which chemists make analogues |
| Nilova, A. & Stuart, D. R. 2019, "Analysis of benzenoid substitution patterns in small molecule active pharmaceutical ingredients," ChemRxiv preprint (journal version, with Campeau & Sherer, J. Med. Chem. 2020, 63: 13389 — as given by search extracts; not opened) | Preprint text as rendered by exa.ai's library copy. ChemRxiv not opened directly (same bot-check). Journal version not opened. Figures not seen | Twelve substitution patterns counted "irrespective of the substituents identity"; percentages; time comparison; explanation via electrophilic aromatic substitution |
| Wikipedia, "Electrophilic aromatic directing groups" (accessed 2026-10-06) | Web page (HTML saved). Secondary, textbook-level | The directing rule (Crum Brown–Gibson 1892, as cited there); ranking when two directors compete; ortho/meta/para nitration splits (as cited there to Sykes and to Clayden) |
Files under lit/_pdfs/zipf-substituents/ (Ertl 2003 PDF + text + p. 375 image; poster PDF + text; PMC and Wikipedia HTML; Zenodo record JSON; npsubstituents.txt; excerpts-web-rendered.md (not published here) with the passages used from the two exa.ai-rendered preprints).
Not reached (do not treat as read): Bemis & Murcko 1999, "Properties of known drugs. 2. Side chains," J. Med. Chem. 42: 5095 (ACS returned a crawl-prevention page to curl). Takeuchi, Kunimoto & Bajorath 2020, "Global assessment of substituents on the basis of analogue series," J. Med. Chem. 63: 15013 (ACS crawl-prevention page to curl; WebFetch timed out; abstract only via search extracts). Benz, Swamidass & Baldi 2008, J. Chem. Inf. Model. 48: 1138 (ACS; abstract only, as in the Zipf note). Ertl, Altmann & McKenna 2020, "The most common functional groups in bioactive molecules…," J. Med. Chem. 63: 8408 (the author's free "articles on request" link also landed on ACS crawl prevention). Not tried: Lewell et al. 1998 (RECAP), Hajduk & Sauer 2008, Brown, Gagnon & Boström 2015 (p-chlorophenyl), Takeuchi et al. 2021 Eur. J. Med. Chem. 225: 113771.
2. Question (a): are substituent frequencies power-law, and with what slope?
Long-tailed, yes. A measured slope, no.
- Ertl 2003. 3,043,941 commercially available molecules were processed. A "substituent" is any group of up to 12 non-hydrogen atoms joined to the rest of the molecule by one breakable non-ring single bond. That gives 849,574 unique substituents. Only 50 occur in more than 1% of the molecules and 438 in more than 0.1%. About 64% are singletons, found in one molecule only (p. 374). The paper gives no rank–frequency plot and no fitted exponent. It names Lewell et al. 1998 for "a similar frequency distribution" of fragments in the World Drug Index, without a number.
- Ertl 2003, Fig. 2 (page image checked). This plot shows growth, not frequency. The number of unique substituents found keeps rising as more molecules are processed, with no plateau. Ertl says that on log–log axes (plot "not shown") the dependence "is clearly linear" with r² = 0.998 over 304 points. No slope is printed. Analysis, not Ertl's wording: linguists know this curve shape as Heaps' (Herdan's) law for vocabulary growth. Benz et al. 2008 (abstract only) do use "Heaps'-like" for chemical features.
- Ertl 2021 preprint (natural products). About 222,000 natural products yield 16,977 unique substituents (here, groups of up to 12 non-hydrogen atoms on a ring atom). "The substituents show a typical power law (or 'long tail') distribution with a few very common substituents and a large number of infrequent substituents." 24 occur in more than 1% of molecules, 157 in more than 0.1%, and 8,963 (52.8%) are singletons. Hydroxy is in 59.3% of molecules, methyl 56.7%, methoxy 22.1%, acetoxy 8.7%, hydroxymethyl 8.5%, carboxyl 5.7%, isopropyl 3.7%. The text read has no exponent (figures not seen). Recorded as printed: the preprint's summary of the 2003 paper says "only 400 of them were common, i.e. present in more than 0.1%," while the 2003 paper says 438.
- O'Boyle et al. 2014. In their ChEMBL matched-series data, the five most common R-groups (hydrogen, methyl, phenyl, chloro, methoxy) "comprise 32.5% of the training data." That is top-heaviness stated as a share, not a fit.
- Takeuchi et al. 2021. The top 20 R-groups are ranked by network node degree (how many different replacements a group takes part in), with H, methyl and phenyl on top. These three are so dominant that the authors drop them from the final network. No distribution fit.
So the chemistry filler counts are skewed in the way a Zipf curve is skewed, and one author calls the shape a power law in words. But no opened source puts a number on it that could stand next to the VAC slopes (−1, −0.687) or the framework slope (−1.07). Benz et al. 2008 report exponents of 1.5–3 for ring systems, rigid segments, paths and circular substructures (abstract only), but not for substituents as such.
3. Question (b): is the filler distribution measured per position on a given framework?
Not in anything opened. The data to do it exist in one place, and an abstract claims site preference.
- Ertl 2003 and 2021: pooled. Substituents are counted across all molecules and all attachment points, with the core thrown away. This is the exact mirror of Lipkus et al., who keep the core and throw the substituents away.
- Takeuchi et al. 2021: sampled per site, then pooled. The core of each of the 17,254 analogue series is isolated (a core must contain at least half the heavy atoms of each analogue), giving 61,312 indexed substitution sites. "For each substitution site, all R-groups were sampled." The replacement analysis "was substitution site specific… within the same chemical context. Then, all site-specific replacements were combined." The Zenodo deposition lists three files:
edges_final_network.csv(source, target, weight),top500_R_replacements.xml, andreadme.txt. Per the paper, the edge weight is a replacement's frequency "over all substitution sites." So the per-site counts were made but are not reported as a distribution. Going by the paper and the file list (the 107 MB CSV was not opened), they do not appear to be in the deposition either. - Takeuchi et al. 2020 (abstract only, full text not reached). From about 24,000 analogue series, "more than 19,000 substituents were isolated that were differently distributed"; about 400 "occurred most frequently in different structural contexts"; and "substitution site-specific replacement and network analysis revealed that chemically similar substituents preferentially occurred at given sites." This is the one opened sentence that claims something like frame selectivity for chemical fillers. It is an abstract claim; the measure behind it was not read.
- O'Boyle et al. 2014: one position, many scaffolds. A matched series is "a set of two or more molecules with the same scaffold but different R groups at the same position." But their statistics pool series across scaffolds and protein targets, and they count the order of activity among fillers, not filler frequency.
The like-for-like test from this morning's note (Zipf fit of the fillers at one position of one framework, plus a selectivity test against the fillers' overall frequency, as Römer et al. do with χ² and 1−Kendall τ) has therefore not been found.
One sourced remark bears on what that test would face. Ertl 2003 compared drugs with average organic molecules (243 of each set's 500 most common substituents are shared) and a toxic-molecule set, and wrote that these "are simply the most common organic substituents and will be found at the top after analyzing any large collection of molecules." Analysis: this is the chemistry version of the VAC authors' warning that a bare Zipf curve is "potentially trivial." Pooled counts are dominated by the same generic groups everywhere. Only a frame-by-frame comparison against the pooled ranking could show selectivity. Note that the VAC lead verbs are generic and frame-specific (COME for V across n, TALK for V about n). Ertl's top substituents are generic and the same across sets. Whether they also stay the same across positions on one framework is exactly what has not been measured.
4. Question (c): does anything rank fillers by the slot they take, as distinct from atom identity?
Yes, in three different senses. None is a Zipf law, and none is offered as a theta correlate.
4.1 Which positions get filled on one fixed frame, with filler identity deleted (Nilova & Stuart). They took FDA-approved small-molecule drugs (DrugBank up to 2018; 1,793 hits, 879 or 49% containing a benzene-type ring) and classified 1,323 benzene rings by which of the ring's positions carry a substituent. That gives twelve substitution patterns, counted "irrespective of the substituents identity." Results: 1,4- (para) 23%; mono 23%; 1,2,4- 18%; 1,2- (ortho) 12%. These four make 76%, and the other eight patterns are 1–5% each. The rare ones are "contra-electronic" (1,3-; 1,3,5-) or crowded (1,2,3-; 1,2,3,5-; 1,2,3,4-). The ranking was broadly stable between drugs disclosed in 1958–88 and in 1998–2018 (top four: 75% and 62%). The authors also find an "inverse relationship between structural variation and number of substituents": highly substituted rings lean on natural-product skeletons (about half, 15 of 33, of the 1,2,3,4- rings are morphine-derived).
Analysis: this is the third object the morning note's table lacked. Lipkus ranks frames with fillers and slots deleted. Ertl ranks fillers with frames and slots deleted. Nilova & Stuart rank slot-occupancy patterns on one frame with fillers deleted. With twelve categories it is a top-heavy table, not a power law, and nobody fits one.
4.2 The rule that picks the slot (directing groups; Wikipedia, textbook-level). In electrophilic aromatic substitution (a reaction in which a new group replaces a hydrogen on a benzene ring), the group already on the ring decides where the new one goes. Electron-donating groups send it ortho/para; electron-withdrawing groups, except the halogens, send it meta (the Crum Brown–Gibson rule, 1892, as cited there). The rule is stated over a class property (donor vs withdrawer, by resonance and induction), not over atom identity. The halogens show the classes are not just atom labels: they slow the reaction like withdrawers but direct ortho/para like donors. With two groups already present: "the most-activating substituent usually controls over the less-activating one"; ortho/para directors control over meta ones; when the two are comparable, steric hindrance decides; and "the position between two substituents, each meta to the other, reacts last." A second factor, bulk, shifts the split. Nitration of toluene (methyl) gives ortho 58 / meta 5 / para 37; tert-butylbenzene gives ortho 16 / meta 11 / para 73 (as cited there to Sykes).
Nilova & Stuart tie their frequency table to this rule. The four common patterns are "those most easily accessed by SEAr… and are the most widely commercially available," and the bias persists because "synthetic methods to access contra-electronic or sterically congested benzenoid rings have not kept pace." So for benzene there is a sourced chain: a slot-picking rule stated over group class → which slot patterns are easy to make → which slot patterns are frequent in drugs.
4.3 Ranking fillers against each other at one slot (O'Boyle et al. 2014). For R-groups at the same position, some activity orders are far more common than chance. [H, F] shows almost none (max enrichment 1.06). [H, F, Cl] reaches 1.85 (Cl > F > H). [H, F, Cl, Br] reaches 5.62 (Br > Cl > F > H, 230 observations). The least frequent order, Br > H > F > Cl, sits at 0.22 (9 observations). The alkane series C, CC, CCC, CCCC behaves the same way (5.64 for CCCC > CCC > CC > C). The authors rule out a general link to simple property descriptors. They have "not investigated" the origin beyond that, and they flag sampling bias: a chemist makes the butyl analogue only if it is expected to help. The 2019 NextMove poster likewise reads R-group frequency as the order in which chemists make analogues ("synthesised before… and are therefore more frequent").
Analysis, SPECULATIVE comparison only: 4.2 is the closest thing in this run to "the thing that picks a slot." It is a rule over a property class, it applies inside an event (a substitution reaction), and it has a stated ranking when two directors compete. Those are features the TBP and spinel notes looked for. It has not been tested against the six-point ranking-edge spec here; that is open question 3. The directing group is a property of a chemical group. Calling it a theta role would be inventing a stand-in, and this note does not.
5. Does this close, narrow, or leave open the "opposite halves" gap?
Narrows it; does not close it.
- Narrowed. The chemistry filler half exists: pooled substituent counts are long-tailed, singleton-heavy, and grow without a plateau (Ertl 2003, 2021). Site-specific sampling exists (Takeuchi et al. 2021), and an abstract claims site preference (Takeuchi et al. 2020). A slot-level ranking on one fixed frame exists (Nilova & Stuart), and a sourced rule links slot choice to group class.
- Still open. No opened source gives (i) a fitted slope or exponent for substituent frequencies, or (ii) a filler distribution at one position of one framework, tested for selectivity against the pooled ranking. Those are the two numbers that would make the comparison with Römer et al. like for like.
- Not claimed. That chemical fillers are Zipfian at −1. That positions are theta slots. That the near-match of slopes from this morning (−1 vs −1.07) now extends to fillers.
6. What this means for Grant's frame, and for the Navajo question
For the SPECULATIVE phrase ↔ molecule frame. The power laws found so far in chemistry sit on things with identity: frameworks, and substituents as types. The VAC result that matters for Grant's brief is that the Zipfian frame is a role-order object. In this run, chemistry's slot-level objects (benzene position patterns, directing rules, preferred orders at a slot) are sourced, but none of them has a power law attached. So the Zipf route has not yet produced a chemical object that is both power-law distributed and defined by slot rather than identity. That is a gap in what has been measured, not a refutation.
For Navajo and "phrasisized." No Navajo counts exist in this corpus, and none are given here. A Navajo version of the VAC test (morning note §5 and OQ5) would need, from Grant's texts: 1. The postposition + verb phrasal verbs in Chee 2025's sense, each instance listed with its postposition frame. This is the frame. 2. For each frame, the verb themes or stems that fill it, with token counts. These are the fillers. 3. The overall count of each verb theme across all the texts, so frame-selectivity can be tested against it (Römer et al.'s χ² / 1−τ step). 4. For the slot-level question, a per-instance record of which argument positions are filled and how they are marked (e.g. direct yi- vs inverse bi-, fourth person). That would give a Navajo analogue of Nilova & Stuart's "which slots are filled" table alongside the "which fillers" table. 5. A note of which phrasal verbs are listed as wholes in a dictionary (Young & Morgan, as available). If "phrasisized" units exist, one checkable expectation is that their filler slot is closed or nearly fixed rather than open. That is a test design, not a prediction from any source.
How many tokens would be enough cannot be said from these sources. The VAC studies used a 100-million-word corpus, and a small text collection would show only the head of any curve.
7. Separation
SOURCED (do not inflate). Ertl 2003 counts, thresholds, singleton share, growth-curve r² 0.998 (no slope), and the "most common organic substituents… at the top" remark. Ertl 2021 preprint counts and its "typical power law (or 'long tail')" wording (no exponent seen). Takeuchi et al. 2021 site-specific sampling, then combining. Takeuchi et al. 2020 site-preference sentence (abstract only). O'Boyle et al. 2014 matched-series definition, top-five share, enrichment tables, sampling-bias caveat. Nilova & Stuart twelve-pattern percentages and their SEAr explanation (preprint as rendered). Directing-group rule and competition order (Wikipedia, secondary).
ANALYSIS (this note). Heaps'-shape label on Ertl Fig. 2. The "three objects" reading (frames / fillers / slot patterns). Ertl's "same top everywhere" remark as the chemistry counterpart of the "potentially trivial" bare-Zipf warning. The 4.2 chain read as a slot-picking rule over a class.
HIS SPECULATIVE (not findings). Phrase ↔ molecule, word ↔ atom. "Phrasisized" units. A Navajo phrasal verb as a prescription with dopantIntentional dilute impurity meant to tune electronic/optical/catalytic behavior without replacing the host as majority phase.-like places.
8. Open questions
- Get the per-site distribution. Open Takeuchi et al. 2020 in full (ACS blocked this run) for how "preferentially occurred at given sites" was measured. Or find any study that reports R-group counts per position for one named scaffold.
- Get a number. Open Benz et al. 2008 in full for whether any of their exponents is for substituent-like fragments. Open Bemis & Murcko 1999 for its side-chain frequency table.
- Test the directing-group rule (4.2) against the six-point ranking-edge spec (TBP apicophilicity ranking edge): fixed roles in the event, a separate scale, a marked outcome when the scale overrides the default, impossible arrangements, an exit category. Use opened primary or textbook sources, not the Wikipedia summary alone. Pointer (2026-10-06 evening): taken up in EAS directing ranking edge (textbook + primary sources: bystander groups rank positions for one entering group by class; activation scale decides opposed directors; bulk and route give near-bans; point 6 leans fail; candidate, not correlate).
- Is there a power law over substitution patterns on a frame with more positions than benzene's twelve patterns allow (e.g. Vitaku, Smith & Njardarson 2014, nitrogen-heterocycle substitution patterns in FDA drugs; not opened)?
- Navajo: the five-item counting spec in §6, once Grant's texts are in hand.
- Language frame-level half (morning OQ2) is untouched here.
9. Sources that failed this run
ACS Publications returned its crawl-prevention page to curl for Takeuchi et al. 2020, Benz et al. 2008, Ertl et al. 2020 (including the author's articles-on-request link), and Bemis & Murcko 1999. WebFetch of the Takeuchi 2020 ACS page timed out. ChemRxiv's PDF link for Ertl 2021 returned a bot-check HTML page to curl, and WebFetch of the ChemRxiv article page and PDF timed out. The ChemRxiv public API returned 404 for the item and DOI routes. The Ertl 2021 and Nilova & Stuart preprints were read through exa.ai's rendered library copies. That rendering is a third-party copy of the preprint text, not the publisher's version, and the figures were not visible.
One-sentence finding: Chemistry does count the fillers, and they are long-tailed (849,574 substituents with about 64% seen once; natural products "a typical power law"), but no opened source gives a slope or counts the fillers position by position on one framework, so the VAC-style comparison is still undone. The nearest slot-level object is the benzene ring, where which positions get filled is top-heavy (four of twelve patterns = 76% of drug rings) and is traced to a rule in which the kind of group already on the ring, not its atom identity, picks the next slot.
Corpus note: zipf-filler-within-frame-chemistry-v0.md. Rank counts stay in the field that measured them. Grant’s phrase↔molecule line, phrasal-verb question, and quark/anyon analogy are his. Not Bridge A. Not a joint field.