A published Kd is a property of five experimental choices, and this beat measured how much each one is worth
also called how to read an affinity number · construct, valency, ligand, salt, method
A binding constant sounds like a fact about two molecules. It is really a fact about which piece of the protein you used, whether it was stuck to a tag, which DNA you chose, how salty the buffer was and how you measured it. This project got a 1,300-fold error by treating one such number as the value.
The same thing, for a scientist
The CXXC-DNA affinity beat produced a rare thing: nine measurements of one interaction across which each of five experimental choices can be priced against the others. Construct span, tag-borne valency, ligand valency and identity, ionic strength, and equilibrium-versus-kinetic readout each move the answer, and none moves it by more than about an order of magnitude on its own, yet together they span 1,300-fold. The transferable lesson is not 'be careful with Kd' but a concrete reading protocol: before comparing two affinities, tabulate those five fields for both, and if any differs, say so in the claim. This graph currently records kd_nM in interface entries as a bare number with no accompanying conditions field, which makes every stored affinity uncomparable by construction.
The five choices, priced
Measured on one interaction, the KMT2A CXXC/MT domain with CpG DNA, from five primary papers read in full 123:
| choice | range seen in this literature | what it was worth here |
|---|---|---|
| construct span | 1136-1208, 1146-1214, 1147-1203, 1147-1244 | confounded, but the 41 C-terminal basic residues are part of a 19-fold step and have never been priced alone |
| tag-borne valency | untagged / chemically synthesised / TEV- or thrombin-cleaved / GST, uncleaved | together with a halving of salt, 5.5-fold (Erfurth 3.5 uM against Risner 0.64 uM, same construct and same oligonucleotide). Since the salt halving alone is worth about 8-fold, the GST tag's own contribution is small and may be nil - far less than the field assumed |
| ligand valency and identity | 12-bp 1 CpG / 12-bp 2 CpG / 16-bp 3 CpG / 52-bp 6 CpG | 9-fold from sequence alone at fixed everything else (rows 2 vs 7); 2-4 fold from extra CpG motifs; duplex length alone, not detectable |
| ionic strength | 150 mM NaCl, 150 mM KCl, 250 mM NaCl, 300 mM NaCl | about 8-fold for halving 300 to 150 mM, from a tagless-against-tagless pair on the same ligand class (Cierpicki 44 uM against Liang 5.7 uM), confounded with method and temperature. It is the largest identifiable single term, as the entropically driven thermodynamics predict. No deliberate salt series has ever been run. |
| readout | NMR titration / ITC / fluorescence polarisation / SPR with immobilised ligand, kinetic k_d/k_a | rows 1 and 2 differ 1.4-fold across NMR and ITC; the SPR geometry is the outlier and is confounded with three other choices |
Two things about this table are worth more than the numbers in it.
First, the explanation everyone reached for was the wrong size, and probably the wrong variable. GST dimerisation is real and was flagged in print in 2008, but the tagless-versus-tagged pair bounds tag-plus-salt at 5.5-fold while a tagless-versus-tagless pair puts about 8-fold on the salt alone - so the tag may be contributing nothing measurable, and the term the field never discussed (ionic strength) is the larger one. A field can hold a correct qualitative explanation for eighteen years without anyone checking whether it is quantitatively sufficient, because a correct-sounding mechanism closes the question as effectively as a measurement does.
Second, the choices are not independent of the biology. 300 mM NaCl is chosen because the domain is basic and the sample misbehaves otherwise; a GST tag is chosen because the domain is small and expresses poorly alone; a 52-bp six-CpG duplex is chosen because it resembles a CpG island. Each choice is made for a good reason and each one biases the answer in a direction the experimenter can predict. That is why the fields must be stored, not just remembered.
The reading protocol, and what this graph should change
Before comparing two affinities, or quoting a fold-difference, tabulate for both:
- the exact residue span of each partner, against a live accession fetch
- whether either partner carried an oligomerising tag, and whether it was cleaved
- the ligand's identity and valency - how many copies of the binding site does it present
- the monovalent salt concentration and species, and the divalent cations
- equilibrium or kinetic, and in-solution or surface-immobilised
If any of the five differs, the fold-difference is not a fact and should not be stated as one. Say instead which measurements exist and under what conditions, and let the reader see the spread.
What to change in this graph. partner_interface entries store kd_nM as a bare integer - see the SCHEMA example, "kd_nM": 53. A bare number is exactly the object this concept says does not exist. The two interface entries written by this beat store a kd_measurements array in which every row carries span, tag, method, DNA, salt and buffer, and ana-cxxc-dna-affinity-ladder stores all nine rows that way. Proposed as a schema convention: any kd_nM must be accompanied by a kd_conditions object or a kd_measurements row carrying at least construct, tag, ligand, salt and method. This is a coordinator decision and is proposed, not made, here.
The same shape appears elsewhere in this graph already. Brief rule 76 is this concept for exposure numbers - 1,500 uM-min per dose against 98.0 mg-h/L cumulative, apparently 65-fold apart and agreeing to 0.5% once the frame is converted. Rule 83 is it for DepMap denominators. Rule 74 is it for antibody readouts, where the clone rather than the label is the identifier. A number that is right in its own frame and wrong in yours is this project's most frequently recurring error class, and it has now produced a 1,300-fold instance.
| total span fold | 1333 |
|---|---|
| schema proposal | Any kd_nM stored on a partner_interface entry should be accompanied by a kd_conditions object or a kd_measurements row carrying at least construct span, tag state, ligand identity and valency, monovalent salt, and method. Proposed to the coordinator; SCHEMA.md is not edited by beats. |
| error class | a number that is right in its own frame and wrong in yours |
| n measurements | 10 |
the five choices
[ "construct span, against a live accession fetch", "tag-borne valency and whether the tag was cleaved", "ligand identity and valency (how many binding sites per ligand molecule)", "monovalent salt concentration and species, plus divalents", "equilibrium versus kinetic, and in-solution versus surface-immobilised" ]
priced in this literature
{
"gst_tag_plus_halving_salt": "5.5-fold (Erfurth CpG1-I 3.5 uM untagged 300 mM NaCl against Risner same construct same oligo GST-tagged 150 mM KCl 0.64 uM). Read against the salt term above, the GST tag's own contribution is small and may be nil.",
"dna_sequence_alone": "9.1-fold (Erfurth CpG3A-I 32 uM vs CpG1-I 3.5 uM, identical construct, salt, method, temperature)",
"extra_cpg_motifs": "2-4 fold (Cierpicki, stated in text from their Supplementary Table 1)",
"duplex_length_12_to_20_bp_at_one_CpG": "not detectable (Allen, data not shown)",
"nmr_titration_versus_itc": "1.4-fold across two papers at matched salt (44 uM vs 32 uM)",
"post_cxxc_basic_patch_1204_1244": "NEVER PRICED",
"ionic_strength_alone": "about 8-fold for 300 to 150 mM NaCl (Cierpicki 44 uM NMR against Liang 5.7 uM EMSA, both tagless, both a 12-bp single-CpG palindrome), confounded with method and temperature. No deliberate salt series exists for any CXXC domain."
}related rules in the brief
[ "rule 76 - per-dose versus cumulative exposure, 65-fold apparent difference that resolves to 0.5%", "rule 83 - a DepMap number needs its release and the line count is gene-specific", "rule 74 - for an antibody readout the clone is the identifier, not the label", "the PDB residue-numbering offset, where ASH2L papers number from a 534-residue isoform against UniProt's 628" ]
- Applied to drug discovery: a compound's reported IC50 against this interface would be equally a property of which construct, which DNA and which salt the assay used, and a 1,300-fold assay-format spread would swamp any real potency difference between two series.
- Applied to this graph: adopting a conditions field retroactively means re-reading the Methods of every paper behind every stored affinity, which is expensive. Adopting it going forward is cheap and leaves a corpus in two states - which is worse than either, unless the retrofit is scheduled.
- How many of the other affinity numbers in this graph are stored as bare kd_nM with no conditions? A sweep of every partner_interface entry would size the exposure and has not been run.
- Is there a second interaction in this graph with two published numbers an order of magnitude or more apart that nobody has compared? The menin-KMT2A pair itself has 53 nM, 90 nM, 9.8 nM and 6.8 nM in one paper depending on construct length.
- What is the salt dependence of the CXXC-DNA contact, measured deliberately rather than inferred from a cross-paper pair? It is the largest identifiable term in the table and the one with the weakest direct evidence.
- E1strong confidence 4 of 5: Replicated in peer-reviewed work by more than one group.paper
Across nine published measurements of the KMT2A CXXC/MT-DNA interaction, a same-laboratory same-construct same-oligonucleotide pair differing only in GST tag, ionic strength and readout differs 5.5-fold, while a same-laboratory same-construct same-salt same-method pair differing only in DNA sequence differs 9-fold.
MLL protects CpG clusters from methylation within the Hoxa9 gene, maintaining transcript expression. · Erfurth FE, Popovic R, Grembecka J, et al. · Proc Natl Acad Sci U S A 2008What this evidence cannot show
The pairings are assembled in ana-cxxc-dna-affinity-ladder from five papers read in full this cycle. Neither pair is a designed control experiment - they are opportunistic comparisons across papers, which means a difference in protein preparation, instrument calibration or fitting convention is folded into each figure and cannot be separated. The 5.5-fold and 9-fold should be read as order-of-magnitude bounds, not as measurements of the individual variables.
Evidence that argues against this (1)
- Opportunistic cross-paper comparison is exactly the practice this concept warns against, applied one level up. The honest statement is that these pairs bound how much the choices can be worth, not how much they ARE worth; the designed experiment - one protein preparation, one instrument, one variable at a time - has never been run for this domain and is what exp-tagless-side-by-side-cxxc-versus-mt-domain proposes. Risner LE et al. 2013
- E2moderate confidence 3 of 5: One strong study, or several smaller consistent ones.paper
Ionic strength across the published measurements ranges from 150 mM NaCl to 300 mM NaCl, and halving it from 300 to 150 mM on a tagless domain against a 12-bp single-CpG palindrome moves the measured Kd about 8-fold (44 uM by NMR against 5.7 uM by EMSA); the interaction's ITC thermodynamics are entropically driven with small favourable enthalpies, the signature of a counterion-release-dominated protein-DNA contact for which affinity is steeply salt-dependent.
MLL protects CpG clusters from methylation within the Hoxa9 gene, maintaining transcript expression. · Erfurth FE, Popovic R, Grembecka J, et al. · Proc Natl Acad Sci U S A 2008What this evidence cannot show
The thermodynamic values are measured and the 8-fold is a cross-paper comparison confounded with method (NMR titration against EMSA) and temperature (25 C against on ice), so it is an order-of-magnitude bound on the salt term rather than a measurement of it. No deliberate salt series exists for this or any CXXC domain, so the largest term in the table remains the least directly evidenced.
Evidence that argues against this (1)
- EMSA systematically reports tighter Kd values than solution-equilibrium methods because the gel matrix cages the complex, so part of the 8-fold may be the assay rather than the salt. A single NMR or ITC titration repeated at 150 and 300 mM would separate them and has not been run. Liang L et al. 2021
- E3moderate confidence 3 of 5: One strong study, or several smaller consistent ones.paper
Allen et al. report that DNA-binding affinity did not differ detectably between 12-, 16- and 20-bp CpG duplexes, so duplex length by itself is not a large term in the comparison, which isolates CpG count, construct and geometry as the terms that matter.
Solution structure of the nonmethyl-CpG-binding CXXC domain of the leukaemia-associated MLL histone methyltransferase. · Allen MD, Grummitt CG, Hilcenko C, et al. · EMBO J 2006What this evidence cannot show
Reported as 'data not shown', so the underlying titrations cannot be inspected and the claim rests on the authors' assertion. All three duplexes carried a single central CpG, so this bears on length and not on CpG count.
Evidence that argues against this (1)
- Erfurth's 16-mer CpG1-III binds 6.7-fold more weakly than the 12-mer CpG1-I from the same cluster with a stoichiometry of 1.93, so in a series where length and CpG count vary together the length effect is not obviously negligible. Allen's single-CpG comparison and Erfurth's multi-CpG one are not testing the same thing. Erfurth FE et al. 2008
The ladder is the dataset from which the five experimental choices are priced against one another.
A worked instance of the ionic-strength term in the five-choices concept.
The DNA-identity term of the concept, priced at 9-fold.
The construct-span term of the concept, which is the one term in this literature that has never been priced.