Tools — PCR

PCR simulator

Paste a template and a primer pair, set the conditions, and see what the reaction would produce: primer thermodynamics, every product the primers can make, amplification curves and a gel. Everything runs in your browser. It is a teaching and planning model, not a substitute for the bench — see how it works and where it is wrong.

Mode

Template & primers

Protocol variants

Multiplex: extra primer pairs

Template amount & quality

≈ 1.0×10⁴ template copies in the reaction.
Fragment length: for degraded or FFPE DNA (blank = intact). Activity left: inhibitors carried over with the sample (100 % = none).

Cycling

Reaction

Resultup to date
Target250 bp236 ng
Ct25.1total signal
Efficiency97 %early cycles, target
Other products0off-target + dimers
Error-free55 %of target molecules, 26 doublings
Free Mg²⁺0.73 mMat the end (start 0.73)

What to expect

  • NoteTaq: ~45 % of target molecules carry at least one polymerase error after ~26 doublings (0.59 per molecule on average). For cloning or sequencing single molecules, a proofreading enzyme gives far more error-free product.
  • NoteForward primer has 4 G/C in its last five 3' bases; a very stable 3' end can prime at partly matching sites.
  • NoteTaq adds an untemplated 3' A to most products (251 bp strands): fine for TA cloning; blunt-end cloning needs the ends polished first.

Primer report

PrimerLengthGCTmHairpin ΔGSelf-dimer ΔG3' self-dimer ΔG
Forward21 nt52 %61.2 °C none-2.8-1.5
Reverse21 nt48 %60.2 °C none-1.7> 0

Tm: SantaLucia nearest-neighbour, Owczarzy 2008 Mg²⁺ correction, at 0.5 µM primer. ΔG: minimum-free-energy structures in kcal/mol at 37 °C; more negative is more stable. Tm difference 1.0 °C; cross-dimer -2.5 (3' end -1.3).

Products after 30 cycles

ProductSizeAmountCopiesError-freeWhere
Target250 bp236 ng (4.7 ng/µL)8.8×10¹¹55 %F → R · 477–726 · F exact, R exact
Target amplicon sequence (250 bp)
CATGTACGTTGCTATCCAGGCTGTGCTATCCCTGTACGCCTCTGGCCGTACCACTGGCATCGTGATGGACTCCGGTGACGGGGTCACCCACACTGTGCCCATCTACGAGGGGTATGCCCTCCCCCATGCCATCCTGCGTCTGGACCTGGCTGGCCGGGACCTGACTGACTACCTCATGAAGATCCTCACCGAGCGCGGCTACAGCTTCACCACCACGGCCGAGCGGGAAATCGTGCGTGACATTAAGGAG

Why the target behaves this way

Denaturation100 % meltsAt plateau concentration; strands separate at ≈ 90.6 °C (melt model) vs 95 °C, more easily early in the run
Primer sites−2 % primingNo significant template structure at the primer sites
Structure stallsnoneNo stable hairpin at the extension temperature
Extension100 % finishTaq, 30 s per cycle
Polymerase at the end74 % activeHeat inactivation; capacity factor 0.26 in the last cycle
Reagents leftprimers 94 %, dNTPs 98 %Pyrophosphate 15 µM; free Mg²⁺ 0.73 → 0.73 mM

Amplification (SYBR-style signal)

  • Target
10⁻⁴10⁻³10⁻²10⁻¹110¹051015202530 Cycle dsDNA product, ng/µL threshold Ct 25.1

Signal = double-stranded product (template excluded, like a baseline-subtracted qPCR trace). Ct is where the total crosses an assumed detection threshold of 1 ng/µL; real thresholds are instrument-specific, so compare Ct values between runs here rather than with your cycler.

Agarose gel

Lanes load 5 µL of the 50 µL reaction. NTC = the same reaction without template. Faint smear at the bottom: unused primers. Band brightness follows mass and saturates above ~100 ng; positions follow log(size) within the gel's resolving range.

How the model works

Each step is a published model where one exists and a labelled assumption where it does not. Assumed constants are marked assumed here and ASSUMED in the source. Absolute yields and Ct values are illustrative; trends (what happens when you raise the annealing temperature, shorten the extension, add template) are what the model is for.

1 · Primer melting temperature

Nearest-neighbour model with the SantaLucia (1998) unified parameters: Tm = ΔH / (ΔS + R ln(CT/4)) at 1 M Na⁺, then corrected for monovalent ions and Mg²⁺ with Owczarzy et al. (2004, 2008), including Mg²⁺ bound by dNTPs (Ka = 3·10⁴ M⁻¹). As published, the 2008 correction switches regime with the ratio R = √[Mg²⁺]/[Mon⁺]: below 0.22 the monovalent formula alone is used, so Tm can step slightly as Mg²⁺ crosses that ratio. DMSO lowers Tm by 0.75 °C per % (von Ahsen et al. 2001). Betaine is not modelled: it makes AT and GC pairs equally stable at 5.2 M (Rees et al. 1993), but no validated correction for oligonucleotide Tm exists, and we do not invent one.

2 · Hairpins and dimers

Minimum-free-energy folding by dynamic programming (Zuker & Stiegler 1981) with the DNA energy model of SantaLucia & Hicks (2004): Watson–Crick stacks, single mismatches, terminal mismatches, dangling ends, bulges, internal loops (with the asymmetry penalty) and hairpin loops, plus duplex initiation and terminal AT penalties. Entropies are salt-corrected per base-pair step with the SantaLucia salt term at the von Ahsen Na⁺ equivalent, as Primer3's thal does, so structure Tm values can differ by about a degree from primer Tm, which uses Owczarzy 2008. Dimer Tm follows Primer3's convention of treating the primer concentration as CT (CT/4 for two different primers); with each primer at that concentration the strict value is about 1 °C higher. Self-dimers use CT, as two copies of one strand should, plus the symmetry term when the pairing is symmetric. The parameter tables were extracted from Primer3's machine-readable copy of the 2004 set and checked for internal consistency. Where the overhanging bases at a duplex end could pair, the end is scored as a dangle (the helix would simply continue if they paired); this is our assumption, not part of the published rules, and Primer3 uses its Watson–Crick terminal-mismatch entries there instead. Adopting it improved agreement for mismatched duplexes and worsened one hairpin (the largest hairpin difference rose from 1.7 to 2.4 kcal/mol). Not included: multibranch loops (irrelevant at primer length), tetraloop/triloop bonuses (their units in the source files are inconsistent, so we leave them out rather than guess) and coaxial stacking. The primer report shows the most stable structure and the most stable one that pairs a 3' end, since only those can be extended. Primers folded into a hairpin at the annealing temperature are unavailable for priming in the kinetic model.

3 · Binding sites and products

Both primers are compared with both strands at every position, allowing up to three mismatches (no gaps). Sites are anchored at the primer's 3' end and extended towards the 5' end only while that pays (+1 per paired base, −3 per mismatch, at least 12 nt), so a 5' tail such as a restriction site, adapter or T7 promoter is recognised: it stays unpaired on the original template, the first cycles prime at the anchored segment's Tm, and the product carries the tails. Each candidate site is scored as a duplex with the same 2004 mismatch, terminal-mismatch and dangling-end parameters. A mismatch's effect on extension depends on its distance from the 3' end (×0.06 at the terminal base, ×0.25 one base in, ×0.5 two in, ×0.75 three or four in), a coarse summary of Kwok et al. (1990) and Stadhouders et al. (2010); the mismatch type is ignored here. Every forward-facing site paired with every downstream reverse-facing site within 15 kb is a candidate product, including single-primer products. If an intended pair has no exact product (a variant strain or allele), its most extendable mismatched product (up to four mismatches in total) is still reported as the target, flagged with where the mismatches sit.

Thermodynamics benchmarks

Errors of this implementation against reference implementations of the same published models. Every number in this table is recomputed by the test suite from reference data generated with Primer3 and Biopython, so the page cannot claim more than the code delivers. Agreement with Primer3 shows the parameters and algorithm are implemented correctly. It is not agreement with experiment; for that, see the comparison with 443 measured melting temperatures under Validation against published data.

QuantityReferencenMean abs. errorMedian90th pct.Max
Primer Tm (SantaLucia salt mode) (°C)Primer3 2.4300< 0.01< 0.01< 0.01< 0.01
Primer Tm (Owczarzy 2008 Mg²⁺) (°C)Biopython Tm_NN300< 0.01< 0.01< 0.01< 0.01
Mismatched duplex ΔG (kcal/mol)Primer3 thal400.10< 0.010.310.75
Mismatched duplex Tm (°C)Primer3 thal400.28< 0.010.992.09
Hairpin ΔG (kcal/mol)Primer3 thal320.240.080.952.41
Self-dimer ΔG (kcal/mol)Primer3 thal590.800.491.612.16
Cross-dimer ΔG (kcal/mol)Primer3 thal290.440.361.061.19

Hairpin ΔG is compared on the 32 of 60 primers where both tools find a hairpin. Primer3 also reports 8 structures with ΔG > 0, which we count as no hairpin, and we report 1 weak one (ΔG ≈ -0.1 kcal/mol) that Primer3 does not, so the two agree on which primers form a stable hairpin (their ΔG still differs by up to 2.4 kcal/mol). Of the 32 shared hairpins, 15 have identical ΔG and 4 differ by exactly one salt step (when the 3'-terminal base pairs, thal applies one salt correction fewer than there are stacks). The other 13 differ for other reasons; in the cases we inspected, thal had reported a different fold (it chooses the structure with the highest Tm, we the lowest ΔG), and the end-scoring difference described above also contributes. Self- and cross-dimers are on average 0.76 and 0.40 kcal/mol less stable than Primer3's (median 0.49 and 0.36), for two reasons. Where both strands overhang a duplex end, SantaLucia & Hicks prescribe one terminal mismatch, while Primer3's thal also allows the two dangling ends summed, which is usually more stable. And two copies of one primer pairing over a self-complementary core form a symmetric complex, which costs R·T·ln 2 = 0.43 kcal/mol (SantaLucia 1998); thal leaves that out. Scoring ends Primer3's way, the cross-dimer median difference drops to 0.00 kcal/mol and the self-dimer median to 0.43, the symmetry term (33 of 59 self-dimers differ by exactly that). The small negative cross-dimer bias that remains (-0.19) is probably Primer3 reporting the structure with the highest Tm where we report the lowest ΔG; we have not tested that. We keep the published rules (plus the dangle assumption for pairable ends above).

4 · Amplification, cycle by cycle

Each cycle is simulated as three steps: denaturation, annealing and extension. Strands are tracked per product: template → long product (one defined end) → defined product, which gives the usual two-cycle lag. In each cycle a denatured strand is copied with probability

p = D · capture(θ) · fext · fcomp · fcap · fMg · fE · q

  • D: fraction of product strands that separate at the denaturation temperature, from the same melting model as the melt curves (section 7), so the two agree. Two strands must find each other to stay paired, so separation is evaluated each cycle at the product's current concentration: easy while product is scarce, harder as it accumulates. A product whose GC-rich domain is near the denaturation temperature therefore amplifies late and plateaus low rather than failing outright.
  • θ: fraction of priming sites occupied at the annealing temperature, from competing equilibria on the site: θ = x/(1 + Ks + x), with x = K·[free primer] from the same thermodynamics as the Tm, and Ks for template secondary structure covering the site (folded at the annealing temperature). Primer folded into a hairpin is not free. capture = 1 − e−k·θ·t: bound primers are extended and trapped at rate k = 0.5 s⁻¹ (assumed), which is why annealing a few degrees above Tm still works.
  • fext: chance that the polymerase finishes the amplicon in the extension time, plus the ramp up from the annealing temperature (3 °C/s, assumed). Time needed = L/v + W + stalls: rate v (Taq ≈ 60 nt/s, Innis et al. 1988; Pfu ≈ 25 nt/s with processivity under 20 nt, Takagi et al. 1997), W the waits for the polymerase to rebind after each of the L/P dissociations (exponential, mean 0.25 s, assumed). 95 % of 1 kb strands then finish in about 40 s with Taq, 1.5 min with Pfu and 25 s with the Sso7d-fusion enzyme (rate assumed 80 nt/s), in line with common protocol guidance (30 s–1 min, 1–2 min and 15–30 s per kb; for the fusion enzyme by construction, since its rate was chosen within that guidance), and completion falls off gradually with length rather than at a cliff. Stalls at hairpins inside the amplicon use a passive-unwinding picture (Betterton & Jülicher 2005): the polymerase steps only when the base pair ahead is open, so it waits at each pair of a stem for that pair to fray, about e−ΔGstack/RT/v seconds (or for the rest of the hairpin to melt, if that is likelier), weighted by the chance the hairpin is folded. A very stable 16-bp GC stem costs just under a second at 72 °C, so stalls rarely limit a 30–60 s extension; GC-rich structure matters more through denaturation and primer-site access. This is an upper bound since real polymerases help open the fork. Per-molecule rates spread ±30 % (assumed); low dNTP (Km 15 µM, assumed) slows v.
  • fcomp: primer vs. complementary product strand competing for the same site, with hybridisation rates ∝ √length (Wetmur & Davidson 1968). Matters late, when product is abundant.
  • fcap: polymerase capacity from the unit definition (1 U = 10 nmol dNTP incorporated in 30 min at 74 °C). Capacity falls with heat inactivation (Taq half-life 40 min at 95 °C, 9 min at 97.5 °C; Lawyer et al. 1993), with enzyme bound to product DNA (Kainz 2000; half-inhibition at 20 ng/µL is assumed) and with inhibitors (the "activity left" setting). Inhibitors such as heparin or humic acid mostly remove active enzyme rather than slow each molecule, and enzyme is normally in excess, so they also lower per-cycle success through a saturating availability term fE = (1 + K)·a/(a + K) with K = 0.05 (assumed); fewer active enzymes also lower the capacity in proportion. Together, for the ACTB example in the qPCR preset (20 µL, 0.5 U), Cq shifts by +1.2 at 50 %, +5.2 at 20 %, +12.5 at 10 % and no Cq at 5 % activity; with more enzyme per µL the capacity part matters less.
  • fMg: free Mg²⁺, recomputed every cycle from total Mg²⁺, the remaining dNTPs (Ka 3·10⁴ M⁻¹, Owczarzy 2008) and the pyrophosphate released by synthesis (one per nucleotide; Ka 10⁵ M⁻¹, assumed tighter than dNTP). Free Mg²⁺ also feeds back into primer Tm. It scales per-cycle success once, with a saturating dependence (K = 0.25 mM, assumed), and is not applied again to the extension rate or capacity. q = 0.97: imperfect cycles (assumed).

Primers and dNTPs are used up as strands are made. Under standard conditions pyrophosphate reaches only tens of µM, so in this model it contributes little to the plateau, which comes mostly from polymerase capacity, product binding and re-annealing. Primer dimers are seeded from the continuous helix that ends at a primer's 3' end (at least 3 bp and ΔG37 below −3 kcal/mol; bulged extensions of it are not counted, since the polymerase needs the terminus paired), at a rate proportional to the fraction of primer held in that helix (ε = 10⁻⁹ per cycle, plus a set-up seed of 10⁻⁸ at 25 °C for non-hot-start enzymes; both assumed, chosen so that a strongly dimer-prone pair, ACGTACGTAAACCCGGGTTT / TTTAAACCCGGGTTTACGTA, gives a no-template Cq of 25.2 without hot-start and 31.1 with it in the qPCR preset at 3 mM Mg²⁺; other conditions give other values). Then they amplify like any other product.

Template quality. With random breaks or blocking lesions every S bp on average, a fraction e−L/S of template molecules is intact over an amplicon of length L (the basis of long-amplicon qPCR damage assays; Ayala-Torres et al. 2000). Inhibitors are summarised as the fraction of polymerase activity left.

Checks (unit tests): 10× template dilutions shift Ct by 3.2–3.7 cycles; 100× more template gives < 2× more final product (plateau); yield falls monotonically as annealing temperature rises above Tm; 3'-complementary primers form dimers in a no-template control while the example primers do not; a 78 % GC target fails at 95 °C and recovers with 98 °C and DMSO; degraded template and inhibitors delay Ct. These are behavioural checks. For comparisons with measured amplification curves see Validation against published data.

5 · Polymerase errors and chimeras

Error rates per base per doubling from single-molecule sequencing (Potapov & Ong 2017): Taq 1.8·10⁻⁴ counting all errors (1.5·10⁻⁴ substitutions only), Pfu 5.1·10⁻⁶ and Pfu–Sso7d fusion (Phusion) 3.9·10⁻⁶, substitutions only, so the Taq figure is not like-for-like: compare 1.5·10⁻⁴ with the others. With their definition, a product of length L after n = log₂(final/input) doublings carries L·n·e/2 errors on average, so the error-free fraction is e−L·n·e/2. Fidelity also depends on Mg²⁺ and dNTP balance (Eckert & Kunkel 1990); the rates are for each vendor's recommended buffer.

Chimeras need related templates: paste several FASTA records to model a mixture (alleles, related species). For targets of equal length, the chimera fraction is (n·r·D/2)·4x(1−x), where D is the span between the outermost differing sites, x the share of one template, and r = 1.1·10⁻⁴ per base per doubling, the recombination rate Potapov & Ong measured for Taq. It is applied to all enzymes because it has not been measured for the others; Judo et al. (1998) found a proofreading enzyme recombined several times more than Taq. Recombination also rises with incomplete extension, which is not modelled, so read these as order-of-magnitude estimates.

6 · Replicates at low copy number

Monte Carlo mode samples each replicate's starting copies from a Poisson distribution and amplifies them as a branching process (each molecule copied with the model's per-cycle efficiency) until about 10⁴ copies, then follows the deterministic curve, shifted by how far the replicate is ahead of or behind the mean (Peccoud & Jacob 1996). That captures dropouts (P = e−λ) and the Ct spread at low input. Pipetting and instrument noise (typically 0.1–0.2 cycles, see the low-copy benchmark below), plateau variation and stochastic primer dimers in negative replicates are not sampled, so above ~100 copies real replicates spread more than these.

7 · Real-time PCR and melt curves

SYBR Green reports all double-stranded DNA, primer dimers included. A hydrolysis probe binds inside the amplicon and is cleaved by the polymerase's 5'→3' exonuclease as it extends through (Holland et al. 1991). Each new extension along a probe-bound strand cleaves one probe with probability equal to the probe's occupancy at the extension temperature, from the same mismatch-aware thermodynamics as the primers. Only products carrying the probe site give signal, and proofreading enzymes without 5'→3' exonuclease (Pfu, Sso7d fusions) give none. Thresholds are 1 ng/µL dsDNA (SYBR) and 1 % of probe cleaved (assumed; instruments set their own). Extension speed falls at lower temperature, following Taq measurements of Innis et al. (1988): about 60 nt/s at 70 °C, 24 at 55 °C, 1.5 at 37 °C.

Efficiency is reported two ways: from the slope of log(signal) over the log-linear window of one curve (as in LinRegPCR, Ramakers et al. 2003) and from a standard curve, where the same reaction is run over a dilution series and E = 10−1/slope − 1. Inhibitors carried with the template are diluted with it, so a contaminated sample bends the top of the curve; the most concentrated points are dropped while R² < 0.98 or they come up later than the next dilution.

Melt curves come from an exact Ising/zipper model of each product (the Zimm–Bragg/Poland family reviewed by Wartell & Benight 1985): each base pair is open or closed, neighbouring closed pairs gain the nearest-neighbour stacking free energy, every internal loop costs a cooperativity factor σ = 10⁻⁵ (assumed), and separated strands carry the bimolecular initiation term. Oligonucleotide stacks overestimate polymer stability, so each class of stack is re-anchored to the polymer melting relation of Frank-Kamenetskii (1971) for the buffer's ions, with Mg²⁺ counted as a Na⁺ equivalent, [Na⁺]eq = [Mon⁺] + 120·√[free Mg²⁺] mM (von Ahsen et al. 2001), with free Mg²⁺ from the same dNTP binding equilibrium the kinetics use, so melting rises steadily with both Mg²⁺ and K⁺. (Oligo Tm in SantaLucia mode and structure energies keep Primer3's simpler Mg²⁺ − dNTP, so they stay comparable with Primer3; that rule has a kink where Mg²⁺ equals the dNTP concentration.) With that anchoring, the model's absorbance midpoint reproduces Marmur & Doty (1962) to within 0.41 °C on average (max 0.68 °C) at five GC levels from 30 to 70 %, a consistency check, since that relation was used to choose the anchoring.

Limits: the Na⁺-equivalent rule for Mg²⁺ was derived for oligonucleotides and none of the long-DNA benchmarks contain Mg²⁺, so absolute melting temperatures in PCR buffers are the least certain part of the model (an earlier version extrapolated the Owczarzy 2008 oligo correction instead; it lost almost all Mg²⁺ dependence and was replaced after independent review). SYBR binding (which stabilises the duplex), dye fluorescence changing with temperature, heteroduplexes and the heating rate are not modelled. Even uMelt, a dedicated melting predictor, notes that absolute melting temperatures depend on the chosen parameters, so read peak positions relative to each other. The probe suggestion is model-based, not a validated design.

8 · Protocol variants

  • Touchdown (Don et al. 1991): the annealing temperature steps down each cycle. Nothing special is coded for it: because mismatched sites hardly prime in the hot early cycles, the specific product gets a head start, and in the tests the target: off-target ratio improves more than 3× over constant low annealing.
  • Gradient: the same reaction at evenly spaced annealing temperatures, one gel lane each. Template-structure effects are evaluated at the main annealing temperature for every lane (an approximation).
  • Multiplex: every primer pair in one tube. Products from a forward primer of one pair and a reverse primer of another are off-target; dimers are checked for every primer combination. Pairs share polymerase, dNTPs and Mg²⁺, so a dominant target suppresses the others late in the run; a per-pair primer concentration lets you try primer-limited multiplexing.
  • RT-PCR: RT yield depends on the gene, priming strategy and RNA amount and cannot be predicted from sequence (Ståhlberg et al. 2004), so its efficiency is an input. The model adds how much cDNA covers the amplicon: e−L/λ from a gene-specific primer, e−d/λ from oligo(dT) with d the distance to the RNA's 3' end, e−L/F for random hexamers (λ = 2 kb and F = 800 nt assumed). First-strand cDNA is single-stranded, so only the forward primer primes it in the first cycle. Genomic DNA contamination is not modelled.
  • Nested PCR: round 1 runs with the outer primers; the transferred volume carries all round-1 products and leftover template into round 2 with the inner primers, where only products containing both inner sites can amplify. Carried-over outer primers are ignored (diluted ~50× by a 1 µL transfer).

9 · Gel

Distance travelled is a logistic function of log₁₀(bp), linear across the resolving range for the agarose percentage (Helling et al. 1974; ranges from Sambrook & Russell 2001) and compressed outside it. Brightness is proportional to mass and saturates (1 − e−ng/40, assumed). Blur, smile and grain are cosmetic.

Validation against published data

The model was run against published measurements. The qPCR data come from the datasets compiled in the R package qpcR (Ritz & Spiess 2008). Their primer and template sequences are not part of the data, so these tests check behaviour that should not depend much on sequence. The model needs some assay to run, so every series uses the built-in ACTB example as a stand-in scaffold; a different stand-in would shift the numbers somewhat. All curves are processed with one fixed recipe: baseline = median of cycles 2–6, threshold = 10 % of the median plateau of the most concentrated dilution. Every number below is recomputed by the test suite, which runs before each build; the whole-genome rows need ~75 MB of sequences, so the build checks the same solver on synthetic genomes and the full table is rechecked on request.

Low-copy noise (Ruijter et al. 2013; MYCN, 93–94 replicates per level). The Monte Carlo model is run at the same nominal copy numbers, with nothing fitted to these data. Two reference columns: Poisson sampling alone (the SD of log₂ k for k copies drawn from a Poisson distribution, no amplification model; log₂ assumes perfect doubling, so at 90 % efficiency it would read about 8 % higher), and the model plus one constant technical SD added in quadrature, taken from the 15 000-copy level where sampling noise is negligible (one fitted number). The observed SD carries a 95 % χ² interval, which assumes normally distributed Cq; at 15 copies Cq is skewed (log-Poisson), so that interval is somewhat optimistic. † The 15 000-copy row is where the technical SD comes from, so it matches by construction.

CopiesReplicatesObserved Cq SD (95 % CI)Poisson onlyModelModel + technical
15940.45 (0.40–0.53) 0.390.410.44
150940.19 (0.16–0.22) 0.120.120.19
1,500930.20 (0.17–0.23) 0.040.040.15
15,000940.14 (0.13–0.17) 0.010.010.14†

At 15 copies the model's 0.41 cycles sits inside the observed interval, but Poisson sampling alone already gives 0.39: the branching process adds only a few hundredths of a cycle, so this row is mostly a check of the sampling arithmetic, not evidence for the amplification model. From ~150 copies the observed spread levels off at 0.14–0.20 cycles, which is pipetting and instrument noise. With that noise added the 15- and 150-copy rows fit, while the 1 500-copy row (observed 0.20) is noisier than a constant technical SD explains. The copy numbers are nominal values from a dilution series, not counted, so a few-fold error in them would shift the low-copy rows noticeably. Technical noise is not part of the simulator's Monte Carlo, which therefore understates replicate spread above ~100 copies. Data note: in qpcR's copy of this dataset the copy-number labels run opposite to the curves (the "15" curves rise first and vary least); copy numbers were assigned by Cq order.

Predicting a dilution series. Four assay parameters (input copies, polymerase capacity, intrinsic efficiency and the fluorescence scale) are fitted to the most concentrated dilution only. The Cq of every other dilution is then predicted from the dilution factor. Baselines extrapolate from the same top Cq, Cqtop + i·log(d)/log(1+E), with E taken from the top curve's log-linear phase (single-curve efficiency), or simply assumed to be 100 % or 90 %. A leave-one-out standard curve, fitted to every other dilution, shows what a real standard curve achieves. It is a reference, not a ceiling: a fixed 90 % does as well or better on some series, because these assays happen to run near 90 %. Errors are mean absolute Cq errors in cycles; shape is the RMS deviation of predicted from observed curves.

DatasetDilutionsModelSingle-curve EE = 100 %E = 90 %Standard curve (LOO)Shape (% of plateau)
Guescini et al. 20087 × 10-fold1.090.760.880.070.057.1
Rutledge 20046 × 10-fold1.081.530.970.180.189.7
Boggy & Woolf 20106 × 10-fold0.840.510.700.220.3316.3
Lievens et al. 20125 × 5-fold0.342.650.390.290.364.1

The model averages 0.84 cycles of error, against 1.36 for the single-curve efficiency, 0.73 for assuming perfect doubling and 0.19 for assuming 90 %. It beats perfect doubling on 1 of 4 series and a fixed 90 % on 0. A single curve's shape does not pin down the efficiency well enough to extrapolate across dilutions: use the model to explore how conditions change a reaction, not to quantify. Each series mean rests on only 4–6 predicted dilutions, and there are four series, so differences of a few tenths of a cycle between columns are not meaningful. On that scale the model is no better than assuming perfect doubling (a gap of 0.10 cycles) and clearly worse than assuming 90 %. A fixed efficiency only works when it happens to be right; when the efficiency is unknown, quantification needs a real standard curve, which the last column shows is accurate to a few tenths of a cycle. Batsch et al. (2008) SYBR series were excluded: their Cq steps (~1 cycle per annotated 4-fold dilution) imply efficiencies above 200 %, so the annotation does not describe the curves. That exclusion was decided after looking at the data. The single-curve efficiency here uses a 3–30 % of plateau window; the simulator's own qPCR panel uses a LinRegPCR-style window, so the two efficiencies are not interchangeable.

Long-DNA melting: the melt model's absorbance midpoint matches Marmur & Doty's (1962) relation for genomic DNA within 0.41 °C on average (max 0.68 °C) at five GC levels from 30 to 70 %. That check was used to choose the anchoring, so it is a consistency check, not an independent test.

Whole-genome melting. The melt model run on complete genome sequences from NCBI, against measured melting temperatures compiled by Khandelwal & Bhyravabhotla (2010) from Mandel et al. (1970), Owen & Jackman (1982), Blake (1987), Ussery (2001) and Schildkraut & Lifson (1965). Na⁺ only.

SetnRangeMean abs. errorMaxBias
Other genomes (20 organisms) 2326–73 % GC, 0.016–0.15 M Na⁺0.87 °C2.72 °C−0.37 °C
E. coli at several salts120.01–0.6 M Na⁺2.53 °C4.68 °C+2.37 °C

Across 23 measurements of 20 organisms (several measured at two salt levels or by two labs; each genome is one sequenced strain, not necessarily the one measured), spanning 26–73 % GC, the model's absorbance midpoint is within 0.9 °C on average. The E. coli salt series is predicted about 2.4 °C too high. That comes from the 10 1965 Schildkraut & Lifson measurements (on average +2.94 °C); E. coli measured by other labs agrees within 0.6 °C, and the labs themselves differ by 5.0 °C (4.6 °C after allowing for the salt difference) at 0.015–0.016 M, so part of the gap is measurement spread. Against that one series the model runs +2.94 °C high on average (SD 1.2 °C, range +0.61 to +4.68), with no clear trend across all 10 points (slope −0.28 ± 0.75 °C per decade of Na⁺). That depends on the 0.6 M point: without it the slope is −1.74 ± 0.63, a clear trend. So these data cannot settle the model's salt dependence: an error of up to about 3.2 °C per decade is possible (95 %, t distribution). The genomic data partly overlap the data such polymer relations were fitted on (Schildkraut & Lifson's series among them), so this is a consistency check against real genomes rather than a fully independent test. Several organisms appear more than once and the E. coli rows are one organism, so the rows are not independent; with so few organisms no interval is given. Reproduce with npm run test:genomes after fetching the sequences (~75 MB; each is checked against its recorded NCBI version and checksum).

Primer Tm against measured melting temperatures. 443 UV-melting measurements of 127 DNA duplexes (15–30 bp) at 0.07–1.02 M Na⁺, each at its stated strand concentration. They come from the tables published with Khandelwal & Bhyravabhotla (2010): the benchmark compiled by Panjkovich & Melo (2005) and 15-mers from Owczarzy et al. (2004). 5 measurements of one sequence were left out because its stated length does not match its sequence.

Salt modelMean abs. errorMedian90th pct.MaxBias
Owczarzy 2004/2008 salt (default)1.83 °C1.42 °C4.09 °C6.51 °C+1.00 °C
SantaLucia 1998 salt (Primer3 default)1.33 °C1.03 °C2.81 °C6.15 °C-0.11 °C

Repeated measurements of one duplex at several salts are not independent, so the uncertainty is estimated by resampling the 127 sequences (cluster bootstrap): mean absolute error 1.56–2.11 °C for Owczarzy and 1.21–1.46 °C for SantaLucia (95 % intervals); weighting each duplex once instead gives 2.04 and 1.55 °C. That per-duplex figure is a different quantity, not a contradiction: the 48 duplexes measured only once have larger errors and count more when every duplex weighs the same. The SantaLucia salt term is better on these Na⁺-only data by 0.29–0.72 °C, a real difference. The default (Owczarzy) model predicts about 1.0 °C too high on average over all rows, a figure set mostly by the design data. The bias is already present near 1 M Na⁺, where both salt models make almost no correction (+1.09 and +1.16 °C there), so it comes from the nearest-neighbour parameters, or from the strand concentrations: halving every stated concentration would bring the Owczarzy bias to +0.03 °C. The fixture uses the published total strand concentration CT (CT/4 in the equation), so a convention slip would have to be in the source tables. Owczarzy's correction carries it to every salt level (bias +1.09 at 1.02 M, +1.32 at 0.62 M, +1.11 at 0.22 M, +0.49 at 0.12 M, +1.23 at 0.07 M). The SantaLucia salt term over-corrects at intermediate salt (+1.16, +0.03, −1.17, −0.81, +0.56), so its small overall bias is partly cancellation; its lower error here is real but not evidence of a better salt model. The default stays because these measurements contain no Mg²⁺, and PCR buffers do: Owczarzy 2008 is the better-supported treatment of Mg²⁺, which this dataset cannot test. No offset was fitted to these data. No single offset applies, though: on the Owczarzy design the default model reads +1.20 °C (mean absolute error 1.8 °C), but on the 48 measurements that were not used to fit it, it reads −0.61 °C with an error of 2.5 °C (SantaLucia salt: −1.12 and 2.1 °C). For a sequence outside the training data, expect a typical (mean absolute) error of about 2.5 °C in predicted Tm, with larger misses for some. Much of this data helped derive the model: 395 of the 443 measurements are the Owczarzy et al. (2004) design (79 duplexes at five salts), the data the Owczarzy salt correction was fitted on, so this is agreement with the measurements behind the model, not a fully independent test.

Not modelled yet

Heteroduplexes (and their melt curves), other qPCR chemistries (molecular beacons, dual-labelled hybridisation probes), betaine, DMSO's effect on the polymerase (only its effect on Tm is modelled), the effect of high Mg²⁺ on fidelity and enzyme activity, mismatch type at priming sites, indel-aware chimera estimates, genomic DNA in RT-PCR, and outer-primer carry-over in nested PCR. Simplified: DMSO lowers long-DNA melting by the oligonucleotide rate (0.75 °C per %); denaturation is the equilibrium separated fraction at the setpoint (no sample-temperature overshoot or hold kinetics), and long-DNA melting in PCR buffer is uncertain by about ±2 °C, so predictions for GC-rich products or very high Mg²⁺ (beyond ~5 mM, outside the range the Mg²⁺ equivalence was derived for) are the least reliable; probe occupancy is taken at the extension temperature and does not compete with re-annealing product strands; Monte Carlo Cq values are computed for the target alone, so with a probe or strong primer dimers their absolute Cq can differ from the displayed curve. The template is only the pasted sequence: off-target sites elsewhere in a genome are not searched.

Working on AI for healthcare? Let’s talk.