Sulston Score

A classical physical-mapping tail score that estimates, under a uniform independent-band null model, the chance of observing at least as many fragment-size matches as two clone fingerprints actually share.

Version

v1 · 2026-08-30 · History

Domain-specific #

2893

Origin domain

physical genome mapping

Subdomain

restriction-fragment clone fingerprinting

Aliases

Sulston probability, Sulston coincidence score

Core Idea

The Sulston Score is a classical chance-overlap statistic for restriction-fragment clone fingerprints in physical genome mapping. It asks a tightly bounded question: if two unrelated clones produced fragment positions under a simple uniform random model, what is the probability that their fingerprints would show at least the observed number of size matches? A smaller score makes the observed agreement harder to explain as random coincidence and therefore supports—without proving—the hypothesis that the clones overlap in genomic sequence.^[1][2]

The score was introduced in the mapping software described by Sulston and colleagues and became operationally important through FPC, which compares fingerprinted clones and uses coincidence scores when clustering them into contigs.^[1][2] Its stable identity is neither “any measure of fingerprint similarity” nor the FPC program. It is a particular null-model tail calculation whose inputs are two band counts, a sizing resolution or match tolerance, and an observed match count.

That definition must retain a major qualification. The traditional derivation treats match indications for the smaller fingerprint as independent Bernoulli trials. Wendl showed that the physical fingerprint-matching problem contains conditional dependence, derived a combinatorially exact distribution using Bell exponential polynomials, and found that the classical score can overpredict the actual match probability by orders of magnitude.^[3] A later correction method retained the inexpensive Sulston calculation as a provisional value and transformed it to approximate the exact probability more closely.^[4] The Encyclopedia node therefore preserves the classical score as a reproducible, historically entrenched abstraction while refusing to redescribe it as the exact law of clone-fingerprint coincidence.

The candidate survives reference-grade adjudication. It has a named mathematical identity, primary-source origin, independent critical literature, repeatable inputs and output, and a well-documented role in physical-map assembly. Existing primes explain its probability, p-value, independence, similarity, and thresholding components, but they do not cover its clone-fingerprint null, band-resolution convention, or overlap-assessment role.

Structural Signature

A classical Sulston-score instance contains these roles:

• Two clone fingerprints. Clone \(A\) contains \(m\) scored fragment positions and clone \(B\) contains \(n\) positions, conventionally ordered so that \(m\ge n\).
• A comparison range and resolution. The usable gel or fragment-size range is divided conceptually into \(B\) distinguishable bins, or equivalently a match tolerance \(\pm r\) is declared over a range of length \(G\). The approximation \(1/B\approx 2r/G\) connects the two descriptions.
• A band-matching rule. A fragment in one fingerprint matches a fragment in the other when their measured positions or sizes fall within the declared tolerance. The implementation must fix how ambiguous and competing matches are counted.
• An observed match count. Let \(h\) be the number of accepted fragment matches between the two fingerprints.
• An unrelated-clone null. Fragment positions are treated as uniformly distributed over the usable range, and the classical derivation treats target-band match events as independent.
• A per-band chance-match probability. For a fixed band of the \(n\)-band clone, the probability that at least one of the \(m\) bands lands in its match bin is

\[ p=1-\left(1-\frac{1}{B}\right)^m, \]

or, under the continuous-window notation,

\[ p=1-\left(1-\frac{2r}{G}\right)^m. \]

• A right-tail calculation. The classical Sulston probability is

\[ P_S(h;m,n,B)=\sum_{j=h}^{n}{n\choose j}p^j(1-p)^{n-j}. \]

It is the binomial-model probability of at least \(h\) chance matches, not just the probability of exactly \(h\) matches. - An evidential use. Smaller \(P_S\) values rank a proposed overlap as less compatible with the unrelated-clone null. A project may compare the score with an empirically chosen cutoff when building or reviewing contigs. - A model-validity boundary. The score's classical independence and uniformity assumptions travel with the value. Fingerprint chemistry, band-calling, size range, tolerance, repeated fragments, contaminants, and library-specific behavior can change its practical calibration.

The recognition test is consequently strict: remove the physical-map fingerprint roles or replace the binomial chance tail with an arbitrary similarity score, and the object is no longer a Sulston Score.

What It Is Not

• Not the posterior probability that two clones overlap. The score conditions on an unrelated-clone null model; it does not combine prior odds and likelihoods for overlapping versus nonoverlapping clones.
• Not proof of overlap. Equal-sized fragments can arise from different sequences, while sizing error, repeats, contamination, and coincidental co-migration can create matches. A low value is evidence that must be integrated with map structure, markers, and laboratory review.
• Not the shared-band count. The integer \(h\) is an input. The score converts that count, together with band counts and resolution, into a tail probability.
• Not overlap length. A stronger fingerprint coincidence does not directly state how many bases the clones share.
• Not generic sequence alignment. No nucleotide-by-nucleotide alignment is computed. The compared objects are experimentally estimated fragment sizes or positions.
• Not FPC. FPC is an interactive contig-building system that uses a Sulston-style coincidence score among other operations. The statistic can be described independently of that software.
• Not the cutoff. A project threshold such as \(10^{-x}\) is a decision setting applied to the continuous score, not the score itself.
• Not automatically \(-\log_{10}(P_S)\). A logarithmic display or transformed rank can be useful, but must be named as a transform rather than silently substituted for the probability.
• Not Wendl's exact distribution or corrected probability. Those later constructions address the dependence defect. They share the comparison problem but are mathematically distinct outputs.^[3][4]
• Not Hardy–Weinberg equilibrium. The frozen catalog's nearest semantic hit is a lexical genetics-and-probability neighbor. Hardy–Weinberg predicts genotype frequencies under population-genetic assumptions; it has no clone-fingerprint comparison roles.

Scope of Application

The home setting is clone-based physical genome mapping: large-insert clones are restriction digested, their fragment patterns are measured, and pairwise similarities are used to propose overlaps and assemble contigs. The original 1988 software work arose in automated genome mapping by fingerprinting, and the 1997 FPC paper formalized a system for clustering restriction-fingerprinted clones into contigs and supporting a minimal tiling path.^[1][2]

The abstraction applies when fingerprints can be represented as comparable fragment-size lists and a project declares a resolution or tolerance model. It was especially natural for gel-based restriction fingerprints and was subsequently discussed for high-throughput and capillary-electrophoresis conditions. Wendl's correction paper explicitly warns that accuracy and extrapolation depend on project-specific parameter ranges, so a score calibrated for one technology or library must not be imported uncritically into another.^[4]

Modern sequence assemblies can infer overlap directly from sequence and may not need the score. That historical change does not erase the abstraction. The score remains a coherent method within its technology-bounded practice, much as an older but still precisely defined estimator remains an abstraction after more accurate alternatives appear.

Clarity

To identify a Sulston-score claim, ask four questions. First, are the inputs two clone fragment fingerprints rather than sequences or generic feature vectors? Second, is a fragment-size matching tolerance or bin model explicit? Third, is the output the probability of at least the observed match count under an unrelated-clone chance model? Fourth, is the classical independence approximation acknowledged?

These questions expose common category errors. “The clones share six bands” reports \(h\), not \(P_S\). “The score is \(10^{-12}\), so there is a one-in-a-trillion probability the clones do not overlap” reverses the conditional and invents a posterior. “Their alignment score is high” concerns a different representation and scoring law. “The FPC cutoff was \(10^{-12}\)” names an operational decision rule, while another library may require a different cutoff even if both use the same equation.

The term “p-value” is serviceable only with the null stated carefully: fingerprints from nonoverlapping or otherwise chance-related clones under the specified uniform-bin approximation. Because the true matching events are dependent, the nominal score is a model-based classical tail value, not a perfectly calibrated probability for every laboratory setting.

Manages Complexity

A physical map with \(N\) clones presents \(N(N-1)/2\) possible pairs. Raw inspection of every pair's fragment lists does not scale, and a match count is not comparable across clones with different numbers of bands or different measurement resolution. The Sulston Score compresses each pair into a common ordered quantity that incorporates those factors. Algorithms can rank candidate neighbors, discard implausible pairs, form tentative contigs, and send ambiguous cases to human review.^[2][4]

This compression is useful precisely because it is lossy. The score suppresses fragment identities, local map context, marker evidence, clone quality, and detailed dependence. It is therefore best treated as a screening and ordering interface: a standardized number that makes an enormous comparison space tractable, not a complete biological verdict.

Multiplicity is part of that complexity. In a large library, even rare chance agreements can appear among the enormous number of pairs. FPC's methodology discusses choosing cutoffs with the number of clone comparisons and library behavior in view.^[2] The score itself does not perform family-wise correction; project-level threshold design remains a separate layer.

Abstract Reasoning

The formula licenses several useful deductions. Holding \(m\), \(n\), and \(B\) fixed, increasing \(h\) can only decrease or leave unchanged the right-tail score, so more observed matches give no weaker classical evidence. Holding \(h\), \(n\), and \(B\) fixed, increasing \(m\) raises the chance that any target band finds a match and can make the same \(h\) less surprising. Finer resolution—larger \(B\), or smaller \(r/G\)—lowers the chance-match probability under the model, although real gains can be limited when laboratory error or dependence remains.

The model also makes assumptions testable. If fragment lengths cluster in favored regions rather than behaving uniformly, \(1/B\) is a poor pair-match probability. If one band can match several candidates or matching choices share fragments, Bernoulli trials are coupled. If fingerprint cleanup changes \(m\) or the usable range, two scores computed under different preprocessing regimes may not be comparable. These are not editorial caveats; they change the probability law.

The exact-distribution result yields a further inference: a conservative classical score can reject real overlaps too often when it overpredicts chance-match probabilities. A smaller exact probability may classify an observed match as compelling even when \(P_S\) remains above a project cutoff. Conversely, replacing the raw score with a correction requires recalibrating thresholds; old empirically tuned cutoffs do not inherit their behavior automatically.^[3][4]

Knowledge Transfer

Literal transfer occurs across gel fingerprinting, high-throughput clone fingerprinting, physical-map construction, contig review, and clone ordering so long as the band-list and resolution model remain meaningful. The same score can be computed in different software because its mathematical roles are separable from its implementation.

Beyond physical mapping, the skeleton—represent two objects, count matches, specify a chance model, compute a tail, and threshold or rank—resembles record linkage, forensic matching, and similarity search. Those are analogies, not additional Sulston-score instances. Their feature distributions, dependency structures, and null hypotheses differ. The transferable logic is already represented by Probability, Statistical Significance, Similarity Measure, Classification, and Threshold-Triggered Rule Activation. The fragment-fingerprint realization is what justifies the domain-specific node.

The history also transfers a methodological lesson: a useful operational score can be stable and reproducible even when later analysis exposes model misspecification. The right response is to preserve the score's identity, document its assumptions, and distinguish exact or corrected successors—not to retrofit the original name onto every improved probability.

Examples

A transparent calculation. Suppose the usable measurement range is represented by \(B=1000\) equally likely bins. Clone \(A\) has \(m=20\) fragments, clone \(B\) has \(n=18\), and the matching algorithm reports \(h=4\) matches. The classical per-band chance is

\[ p=1-(999/1000)^{20}\approx0.019811. \]

The resulting tail is

\[ P_S=\Pr\{X\ge4\},\quad X\sim\operatorname{Binomial}(18,0.019811) \approx 3.77\times10^{-4}. \]

This says the independence model assigns about 0.000377 probability to four or more matches for a chance pair with these parameters. It does not say there is a 0.000377 probability that the clones fail to overlap.

Contig construction. A physical-map project computes pairwise scores for thousands of restriction-fingerprinted clones. Pairs below a library-selected cutoff become candidate overlap links; FPC clusters compatible clones, displays fingerprints, and supports manual resolution. The score supplies evidence for the link, while contig consistency and other biological data remain separate checks.^[2]

A corrected reanalysis. A pair has a raw Sulston value above the laboratory's acceptance threshold. Exact or corrected analysis finds the true chance-match probability materially smaller because the independence approximation was conservative in that parameter region. The corrected value may change the decision, but it is not renamed “the classical Sulston Score.”^[4]

Non-example: sequence overlap. Two reads align over 800 bases with 99.5% identity. A sequence aligner's likelihood or alignment score may strongly support overlap, but no restriction-fragment bins or Sulston tail are involved.

Non-example: shared marker. Two clones carry the same sequence-tagged site. That biological marker can corroborate overlap, yet the marker observation alone is not a Sulston calculation.

Structural Tensions

• Speed versus calibration. The binomial tail is cheap enough for massive pairwise scans. Exact dependence-aware evaluation is substantially harder, while fast corrections trade mathematical completeness for operational feasibility.^[3][4]
• Sensitivity versus false links. A stringent cutoff suppresses accidental contig joins but can split true physical contigs. A permissive cutoff recovers weak overlaps while admitting more chance matches and repeat-driven links.
• Resolution versus model realism. Finer measurement nominally expands \(B\) and makes matches rarer. It cannot by itself remove nonuniform fragments, conditional dependence, band-calling error, or contaminants.
• Comparable scalar versus biological context. One score permits sorting millions of pairs. It hides which fragments matched, whether the pattern is locally consistent, and whether independent markers corroborate the edge.
• Historical continuity versus improved probability. Keeping the name tied to the classical formula preserves reproducibility across papers and software. Treating corrections as interchangeable can silently change thresholds and reported map behavior.

Structural–Framed Character

Assessment: strongly structural, with a conditioned experimental frame. The band counts, resolution, matching relation, null model, and binomial tail are formal and reproducible. Two implementations using identical preprocessing and matching conventions should obtain the same value. The framed component enters through the laboratory's usable size range, tolerance, band-calling rules, clone-quality filters, and cutoff. Those choices determine what counts as a match and how the score informs an overlap decision.

The abstraction is therefore not a mere institutional label. Its central relation is mathematical. Nor is it substrate-free: the parameters gain meaning from restriction-fragment fingerprints and physical-map practice. A reasonable framed-character aggregate is about 0.20, reflecting strong formal structure plus consequential protocol choices.

Structural Core vs. Domain Accent

The portable core is a null-calibrated similarity test: represent two objects, count agreements under a declared tolerance, derive their chance frequency, and map the observed count to a right-tail score. Probability, Statistical Significance, Statistical Independence, Similarity Measure, and Classification already capture those broad relations.

The domain accent is indispensable: the objects are cloned genomic inserts represented by restriction fragments; the comparison space is a gel or fragment-size range; resolution defines indistinguishable positions; and the decision concerns physical overlap and contig construction. Remove those roles and the remaining structure is a generic p-value or matching statistic. Retain them and the Sulston Score remains recognizable across physical-mapping projects and software implementations.

Instantiates / Related Primes

The Sulston Score strictly instantiates prime:statistical_significance_p_value: it is a tail probability of a match-count statistic under a stated chance null and is read as evidence against that null. It presupposes prime:probability and relies classically on prime:statistical_independence, whose failure explains the Wendl correction program. prime:hypothesis_testing_null_vs_alternative describes the null-versus-overlap decision frame, although the score alone does not fully specify an alternative model. prime:multiple_comparisons_correction is related to library-scale cutoff selection but is not built into the raw formula. prime:similarity_measure covers the broader pair-to-degree interface, and prime:classification covers the downstream overlap/nonoverlap use.

Only Statistical Significance is proposed as a DAG parent. The other primes are analytic relations, dependencies, or downstream uses rather than additional minimal subsumption parents.

Relationships to Other Abstractions

Current abstraction Sulston Score Domain-specific

Parents (1) — more general patterns this builds on

• Sulston Score is a kind of Statistical Significance (p-Value) Prime

  The Sulston Score strictly instantiates prime:statistical_significance_p_value: it is a tail probability of a match-count statistic under a stated chance null and is read as evidence against that null.

Hierarchy paths (11) — routes to 5 parentless roots

• Sulston Score → Statistical Significance (p-Value) → Statistical Inference → Inductive Reasoning

Show alternative paths (10)

Neighborhood in Abstraction Space

Sulston Score sits in a sparse region of the domain-specific corpus (99^th percentile for distinctiveness): few abstractions share its structure, so a faithful description tends to retrieve it precisely.

Family — Unclustered & Miscellaneous (1565 abstractions)

Nearest neighbors

• Hunt–Szymanski Algorithm — 0.75
• Gower's Distance — 0.74
• Gap Penalty — 0.74
• Wilf Equivalence — 0.74
• Normal Order of an Arithmetic Function — 0.74

Computed from structural-signature embeddings · 2026-09-08

Not to Be Confused With

domain_specific:hardy_weinberg_principle, the frozen top semantic neighbor, is not coverage. It establishes a population-genetic genotype-frequency baseline; the Sulston Score evaluates clone-fingerprint coincidences. domain_specific:bayes_factor compares marginal likelihoods of two models, whereas the Sulston Score computes a tail under one approximate null. domain_specific:benjamini_hochberg_procedure controls false discovery rate across sorted p-values; it could process a family of scores only after separate validity questions are addressed.

prime:statistical_significance_p_value is the containing abstraction, not a duplicate: it omits clone fragments, measurement bins, observed band matches, and physical-map assembly. prime:similarity_measure is broader still and does not require a probability model. prime:statistical_independence names the assumption that makes the binomial derivation possible, not the derived score. Wendl's exact Bell-polynomial distribution and algebraically corrected score solve or approximate the same physical-mapping problem but must remain distinguishable from the historical formula.

References

[1] Sulston, J.; Mallett, F.; Staden, R.; Durbin, R.; Horsnell, T.; and Coulson, A. “Software for genome mapping by fingerprinting techniques.” Computer Applications in the Biosciences 4(1), 1988, 125–132. https://doi.org/10.1093/bioinformatics/4.1.125 registry ↩a ↩b ↩c

[2] Soderlund, C.; Longden, I.; and Mott, R. “FPC: A system for building contigs from restriction fingerprinted clones.” Computer Applications in the Biosciences 13(5), 1997, 523–535. https://doi.org/10.1093/bioinformatics/13.5.523 registry ↩a ↩b ↩c ↩d ↩e ↩f

[3] Wendl, M. C. “Probabilistic assessment of clone overlaps in DNA fingerprint mapping via a priori models.” Journal of Computational Biology 12(3), 2005, 283–297. https://doi.org/10.1089/cmb.2005.12.283 registry ↩a ↩b ↩c ↩d

[4] Wendl, M. C. “Algebraic correction methods for computational assessment of clone overlaps in DNA fingerprint mapping.” BMC Bioinformatics 8, 127 (2007). https://doi.org/10.1186/1471-2105-8-127 registry ↩a ↩b ↩c ↩d ↩e ↩f ↩g