Switch on polar contacts in a structure viewer and dashed lines appear; run a script over a trajectory and it returns a hydrogen-bond count and a salt-bridge list. Neither is a property of the peptide alone: both are conditional on the cutoffs you chose, and neither measures what those contacts are worth. Conflating the comparability problem with the energetics problem is how a structure figure becomes a stability claim it cannot support.
Hydrogen bond criteria are conventions, and moving them moves the answer
Almost every hydrogen bond scored in a peptide structure is inferred from heavy-atom positions rather than observed. Hydrogen scatters X-rays weakly; a review of sub-atomic resolution methods noted that the number of X-ray structures determined to sub-atomic resolution (≤1 Å) had passed 600, and that neutron crystallography remains "the only approach to locate mobile or highly polarized H atoms and protons" [1]. At ordinary resolution, the hydrogens your script uses to compute a donor–hydrogen–acceptor angle were placed by a geometric rule, and the angle criterion inherits that rule, most heavily for donors whose hydrogen position depends on a rotatable torsion.
The thresholds themselves are conventions. McDonald and Thornton quantified how sensitive the outcome is: under their standard criteria, in high-resolution structures 9.5% and 5.1% of buried main-chain nitrogen and oxygen atoms respectively fail to hydrogen bond, representing 5.8% and 2.1% of all main-chain nitrogen and oxygen atoms. Relax the thresholds and the picture shifts: "if the hydrogen bond criteria are relaxed many of these unsatisfied atoms form weak hydrogen bonds", and so does improving the data, since "as the resolution of the data improves, the percentages fall" [2]. A residue flagged as an unsatisfied donor under one protocol is hydrogen bonded under another, on the same coordinates.
Geometry may not be sufficient on its own. Torshin and colleagues applied a geometrically consistent criterion set to a data set of 30 structures; doing so, together with excluding abnormal backbone conformations, also excluded a common interaction of Ser/Thr side chains with Asp/Glu side chains, which they termed bifurcated hydrogen bonds and which are "among the most common packing patterns" for those side-chain pairs. Their conclusion: "the identification of hydrogen bonds cannot be based on geometrical criteria only and requires introduction of some physico-chemical criteria" [3].
The practical consequence: a hydrogen-bond count without its criteria is not a number a reader can use, and two counts from different programs are not comparable.
Salt bridges: the definition decides which class of contact you can see
Salt-bridge definitions fall into two families, and the choice is not cosmetic. Donald, Kulp and DeGrado, surveying salt-bridge geometry across a large structure database, set the alternatives out explicitly: "Barlow and Thornton used a cutoff of 4 Å between N-O atom pairs as a definition of salt bridge formation... Kumar and Nussinov have also used a more restrictive cutoff of 4 Å between functional group centroids." They took the atom-pair definition, and gave the reason: "the centroid distance restriction removes almost all 'backside' Arg salt bridges, one of the major types of Arg based salt bridges, from the dataset" [4].
The two conventions do not merely disagree on a count; they disagree on which structural class is representable at all, so a centroid-based tally and an atom-pair tally are counting populations defined to contain different things. An arginine-rich peptide is where that bites. None of this makes the geometry uninformative: the same survey found that salt bridges between Asp or Glu and His, Arg or Lys "display extremely well-defined geometric preferences" [4]. Geometry carries real signal; it is the cross-paper arithmetic that fails.
Present in a model versus occupied in an ensemble
A contact drawn in a deposited structure is present in that model, at that temperature, in that lattice. Fraser and colleagues re-analysed available data for 30 different proteins with new computational tools and found that "crystal cryocooling remodels the conformational distributions of more than 35% of side chains and eliminates packing defects necessary for functional motions", exposing "a bias in structural databases toward smaller, overpacked, and unrealistically unique models" [5]. A single cryogenic model is a weak basis for asserting that a particular ion pair is formed.
Occupancy, the fraction of frames in which the geometric criterion is satisfied, is the right quantity to take from a trajectory. It does not escape the criterion problem, and it adds a second dependency: the force field. Ahmed, Papaleo and Lindorff-Larsen computed, by enhanced sampling, the free energy of forming three possible lysine–carboxylate interactions in the B1 domain of protein G, benchmarked against NMR. The NMR experiments showed those interactions "are either not formed, or only very weakly formed, in solution", yet "the stability of the salt bridges is overestimated, to different extents, in simulations of GB1 using seven out of eight commonly used combinations of fixed charge force fields and water models"; the Amber ff15ipq force field gave weaker salt bridges in good agreement with the NMR experiments [6]. An occupancy figure is therefore a joint statement about the peptide, the criterion and the electrostatics of the potential, and it needs the sampling evidence discussed in the guide on molecular dynamics for peptide conformational sampling.
Counting contacts is not measuring stability
A count cannot be converted into a stability number because forming a buried polar or charged contact costs desolvation. Bosshard, Marti and Jelesarov partition the net electrostatic free energy of a salt bridge into three terms, "charge-charge interactions, interactions of charges with permanent dipoles, and desolvation of charges", and note that "energetically favorable Coulombic charge-charge interaction is opposed by often unfavorable desolvation of interacting charges. As a consequence, salt bridges may destabilize the structure of the folded protein" [7]. The sign of that sum is not visible in the geometry.
How the sum comes out is genuinely contested. Hendsch and Tidor calculated the electrostatic contribution to the free energy of folding for 21 salt bridges in 9 protein X-ray crystal structures using a continuum approach: 17 were electrostatically destabilizing, and the average calculated free energy change, a quantity they describe as analogous to mutation of the salt-bridging side chains to hydrophobic isosteres, was 3.5 kcal/mol. They flagged the reference-state trap directly, noting that this is "fundamentally different from stability measurements using pKa shifts, which effectively measure the strength of a salt bridge relative to 1 or more charged hydrogen bonds" [8]. Kumar and Nussinov ran continuum calculations on a larger set, 222 non-equivalent salt bridges from 36 non-homologous high-resolution monomeric protein crystal structures, and reached the opposite majority verdict: most were stabilizing "regardless of whether they are buried or exposed, isolated or networked, hydrogen bonded or non-hydrogen bonded", because although burial imposes large desolvation penalties, the electrostatic interactions are also stronger in the interior "due to the absence of solvent screening" and frequently more than compensate. They identify geometry as a critical factor, favourably positioned pairs being likely to stabilise anywhere in the structure [9].
Both are continuum electrostatic calculations on different datasets, 21 salt bridges against 222. Hendsch and Tidor's reference-state caveat is aimed at pKa-shift measurements, not at the opposing calculation, so it does not explain the divergence. Take the disagreement at face value rather than picking a side: the sign and the size of a charge pair's contribution are not readable off a distance measurement.
The experimental version of the same lesson is blunter. In Arc repressor the side chains of Arg 31, Glu 36 and Arg 40 form a buried salt-bridge triad. Waldburger, Schildbach and Sauer replaced the entire network with hydrophobic residues in combinatorial randomization experiments and obtained active mutants "significantly more stable than wild type"; a crystal structure of one mutant showed the substituted side chains packing against each other in an otherwise wild-type fold, so "simple hydrophobic interactions provide more stabilizing energy than the buried salt bridge and confer comparable conformational specificity" [10]. Three contacts a script would have scored as favourable were, on balance, worse than packing.
What the interior costs a charge is site-specific and large. Isom and colleagues engineered 25 variants of staphylococcal nuclease carrying lysine at internal positions; 19 had depressed pKa values, "some as low as 5.3", and 20 titrated without triggering any detectable conformational reorganization. The apparent polarizability reported by those pKa values "varied significantly from location to location inside the protein", and the pKa values of some internal lysines were affected by interactions with surface carboxylic groups [11]. Burial penalises the charged form by an amount that depends on where the burial happens and what else is nearby.
Hydrogen bonds carry the same context dependence, with an extra trap for anyone counting. Pace and co-workers measured the change in stability for a series of hydrogen-bonding mutants across four proteins and summarised the outcome as: hydrogen bonds contribute favourably to protein stability; that contribution "is strongly context dependent"; and "polar group burial can make a favorable contribution to protein stability even if the polar groups are not hydrogen bonded". Crystal structures of three of the ribonuclease Sa mutants showed the orphaned hydrogen bonding partners forming intermolecular hydrogen bonds to water [12]. A polar atom your script scores as unsatisfied may still be contributing, and deleting a bond does not delete the interaction; solvent takes it over.
At peptide scale, contacts are not additive
Isolated helical peptides give the cleanest test of whether counting works, because the contacts can be placed deliberately. Iqbalsyah and Doig studied Glu-Lys and Lys-Glu salt bridges formed when the residues are spaced i, i+4 on the surface of an isolated α-helix in aqueous solution, and found both stabilising to the helix, by -0.60 and -1.02 kcal/mol respectively, when the interacting residues are fully charged. When the side chains were spaced i, i+4, i+8 to form a Glu-Lys-Glu triplet, "the second salt bridge provides no additional stabilization to the helix", which they attribute to the inability of the central Lys to form two salt bridges simultaneously [13]. Two salt bridges drawn in a model can be worth one measured one, and the two orientations of the same residue pair are not interchangeable.
Experimental probes, and what each licenses
Amide proton temperature coefficients
Cierpicki and Otlewski analysed 793 amides from 14 proteins and found that amide protons with temperature gradients more positive than -4.6 ppb/K carry a hydrogen bond predictivity above 85%, rising above 93% for gradients between -4 and -1 ppb/K, with an inverse proportionality between the coefficient and hydrogen bond length [14]. The qualifications matter more than the threshold. For hydrogen bonds of similar length, gradients in α-helices are on average 1 ppb/K more negative than in β-sheets, so a number of helical amides in bonds shorter than 2 Å fall below -4.6 ppb/K. Conversely, because their hydrogen bonds are longer, 90% of amides in 310 helices and 98% in β-turns sit more positive than -4.6 ppb/K. Ring currents shift things too: in seven of eight cases, non-hydrogen-bonded amides strongly deshielded by neighbouring aromatic rings showed coefficients more positive than -2 ppb/K [14]. In a helix-rich or aromatic-rich peptide the threshold misreports in both directions.
Baxter and Williamson measured coefficients by 2D NMR in bovine pancreatic trypsin inhibitor and hen egg-white lysozyme, obtaining a range of roughly -16 to +2 ppb/K for amide protons, with the temperature dependence generally linear up to about 15 degrees below the global denaturation temperature. They concluded that "a combined use of both temperature coefficients and exchange rates produces a far more reliable indicator of hydrogen bonding than either alone": slow exchange plus a coefficient more positive than -4.5 ppb/K indicates hydrogen bonding, rapid exchange plus a coefficient more negative than -4.5 ppb/K indicates none [15]. Their threshold and Cierpicki and Otlewski's differ by 0.1 ppb/K, a reminder that a cutoff drawn from one dataset is not a constant of nature. The caveat that matters most for peptide work is also theirs: "the previously observed unreliability of temperature coefficients as measures of hydrogen bonding in peptides may arise from losses of peptide secondary structure on heating" [15]. On a marginally folded peptide the temperature ramp partly melts the structure you are trying to detect, and the coefficient reports the melting: the same problem seen from the other side in the guides on cyclic peptide conformational analysis by NMR and on differential scanning calorimetry for peptide thermal stability.
Hydrogen exchange
Protection is a ratio, and the denominator is sequence-specific. Bai, Milne, Mayne and Englander measured peptide group NH exchange for all 20 naturally occurring amino acids using dipeptide models, found both inductive and steric blocking effects, confirmed the additivity of nearest-neighbour blocking and inductive effects in oligo- and polypeptides, and determined reference rates for alanine-containing peptides together with the effects of temperature. Those results "provide the information necessary to evaluate measured protein NH to ND exchange rates by comparing them with rates to be expected for the same amino acid sequence" in unstructured oligo- and polypeptides [16]. A protection factor computed against a generic intrinsic rate rather than a sequence-matched one is wrong by whatever the neighbouring side chains contribute.
What protection licenses is also narrower than it looks. A slow-exchanging amide tells you that proton was shielded from solvent-mediated exchange over the observation window. It does not identify the acceptor.
Double-mutant cycles
The cycle is the only experiment on this list that isolates a pairwise term. Serrano and colleagues estimated the interaction energy between Asp-12 and Arg-16 in an α-helix on the surface of a barnase mutant by a cycle over wild type, the two single mutants and the double mutant. The changes in free energy of unfolding of the single mutants "are not additive because of the coulombic interaction energy", and "additivity is restored at high concentrations of salt that shield electrostatic interactions", an internal control worth copying. The mutant containing the ion pair was more stable than wild type by 0.5 kcal/mol, "but only a part of the increased stability is attributable to the electrostatic interaction", and the pair does not form a hydrogen-bonded salt bridge at all: the charges are separated by 5-6 Å. The geometry of that ion pair was assumed to be that of a highly homologous ribonuclease rather than determined [17], so even a careful pairwise measurement can rest on a modelled structure.
The cycle still does not deliver what a reader usually wants. Bosshard and colleagues are explicit that the mutation approach yields a coupling free energy which is "an indirect and approximate measure of the free energy of charge-charge interaction", that the pKa approach yields "the relative free energy gained or lost when ionizable groups are being charged", and that "neither the pK(a) approach nor the mutation approach can provide the net free energy of a salt bridge", which currently comes only from computational methods, themselves "often prone to large uncertainties due to simplifying assumptions and insufficient structural information" [7]. No bench experiment hands you the net energy of one salt bridge.
What to report
For any contact count: the program and version, the distance cutoff, the angle cutoff, whether the angle was measured at the hydrogen or between heavy atoms, and whether hydrogens were modelled or observed. For a static structure: resolution, data collection temperature, and whether alternate conformations were modelled. For a trajectory: force field, water model, criterion, sampling, and occupancy rather than presence. For temperature coefficients: the temperature range used, and independent evidence that the peptide's structure survives it. For exchange: the sequence-matched intrinsic rates used as the denominator. And keep the count and the energetics in separate sentences: the first describes a model, the second requires a mutation, a titration or a calculation, each carrying its own reference state.