Gibson Assembly vs. Golden Gate: Thermodynamic Overlap Design & Troubleshooting
Homology arm design, SantaLucia nearest-neighbour melting temperatures and molar ratio optimisation for isothermal Gibson assembly and Golden Gate cloning.
Most failed Gibson assemblies fail for one of four reasons, and three of them are visible on paper before you set up the reaction. Here is how the chemistry constrains the design, and where the common rules of thumb are wrong.
**The short answer.** Gibson assembly joins fragments that share identical sequence at their ends. The overlap has to be long enough to hold the two pieces together at 50 °C while the polymerase fills the gap — that is a melting-temperature floor, not a target window. Longer is not automatically better, molar ratio depends on how many fragments you are joining, and if your overlaps are fine but you still get nothing, the problem is almost always the vector, not the assembly.
**What is actually happening in the tube.** Three enzymes work simultaneously at 50 °C for an hour (Gibson et al., *Nature Methods* 2009). A 5'→3' exonuclease — T5 in the original method — chews back the 5' ends, exposing single-stranded 3' overhangs. Where two fragments share end sequence, those overhangs find each other and anneal. A high-fidelity polymerase fills the remaining gaps, and a thermostable ligase seals the nicks.
The choice of a *thermostable* ligase is what makes the whole thing work in one tube. Taq ligase is active at 50 °C, which is hot enough that short spurious annealing events fall apart before they can be sealed, while your designed 20-nt overlap stays paired. Run the same reaction at 16 °C with T4 ligase and you would seal every accidental pairing in the tube.
**Overlap length: use the manufacturer's numbers, not "longer is safer".** The rule most people carry around is "40 bp overlaps", which comes from the original 2009 paper. Modern HiFi formulations need considerably less, and the guidance depends on fragment count because it is tied to reaction time:
- **2–3 fragments:** 15–20 nt overlaps, 15-minute reaction. - **4–6 fragments:** 20–30 nt overlaps, 60-minute reaction.
Those are NEB's numbers for NEBuilder HiFi. The longer reaction is what allows the longer overlaps to be used, not the other way round. Designing 40-nt overlaps for a two-fragment assembly does not make it more robust — it makes your primers longer, more expensive, and more likely to carry secondary structure.
**Melting temperature is a floor, not a window.** The overlap has to stay annealed at the reaction temperature, so it needs a Tm at or above roughly 48 °C. There is no upper limit. A GC-rich overlap that melts at 58 °C is not "too hot" — it is simply more stable than it needs to be, which costs you nothing.
This matters because tools that enforce a 48–55 °C *band* will flag perfectly ordinary overlaps as suboptimal. SciKeep's planner enforced exactly that band until September 2026 and now checks only the floor; if you have been told by any tool that a 58 °C overlap is a problem, check whether that tool is applying a ceiling the chemistry does not have.
SciKeep computes the overlap Tm with the SantaLucia (1998) nearest-neighbour thermodynamics — the same unified ΔH°/ΔS° table used by the primer designer, with a monovalent-salt correction — rather than the 4×GC + 2×AT rule, which is only valid below about 20 nt and badly overestimates longer overlaps.
**Molar ratio changes with fragment count, and this is where most protocols are misquoted.** The commonly repeated "2:1 or 3:1 insert to vector" is right for a simple two- or three-piece assembly. It is wrong for a multi-part build:
- **2–3 fragments:** roughly 2-fold molar excess of each insert over vector, 0.03–0.2 pmol total DNA. - **4–6 fragments:** approximately equimolar, 1:1, with 0.2–0.5 pmol total.
The logic is that a multi-fragment assembly has to get every junction right in one reaction. Flooding it with excess insert biases towards partial products and insert concatemers rather than the full-length construct.
**When to use Golden Gate instead.** Gibson does not care what your sequence contains — it only needs shared ends. That makes it the right choice for a one-off construct, for fragments straight off a PCR, and for anything with awkward internal sequence.
Golden Gate wins when you will build the same construct repeatedly from a parts library. A type IIS enzyme such as BsaI or BsmBI cuts *outside* its own recognition sequence, so the enzyme site sits in the primer tail and disappears from the product, leaving a scarless four-base junction. Many parts can be assembled in a defined order in one cut-and-ligate cycle.
**Golden Gate's real failure mode is an internal site.** Because the enzyme cuts outside its recognition sequence, an extra copy of that sequence *inside* one of your parts is fatal: the enzyme cuts the part itself, and the assembly quietly produces something other than what you designed. Before ordering parts, scan every fragment for the enzyme's site **on both strands** — a reverse-strand BsaI site reads as GAGACC on the strand you are looking at, and a forward-only search will not see it. If a part carries one, either domesticate it with a silent mutation or switch to an enzyme whose site the part lacks.
One more practical point: BsmBI Golden Gate reactions cycle at 42 °C, not the 55 °C listed as that enzyme's plain-digest temperature. T4 ligase is essentially dead at 55 °C, so a protocol that cycles there will cut your parts thirty times and never ligate them — a reaction that produces no colonies for a reason nothing in the output explains.
**Troubleshooting, in the order worth checking.**
**Lots of colonies, all wrong — usually background from uncut vector.** Your linearisation was incomplete, or template plasmid survived. Run the linearised vector on a gel and look for residual supercoiled band; treat PCR-amplified vector with DpnI to destroy methylated template; include a no-insert control in the same experiment, because that control tells you the background rate directly.
**No colonies at all.** Check the overlaps for hairpins and for runs of G. Secondary structure in the overlap region competes with the intended annealing, and it is the single most common reason a design that looks fine on paper does nothing. Lengthening the overlap slightly or adding DMSO can help, but redesigning the junction a few bases along is usually faster.
**Correct size, wrong sequence at a junction.** Suspect repeated or near-identical overlaps. If two junctions in a multi-part assembly share 12+ bases of similarity, fragments can swap. Design junctions to be mutually dissimilar, not merely individually well-formed.
**What the planner does and does not do.** SciKeep's assembly planner computes overlap Tm from real nearest-neighbour thermodynamics, screens every part for internal type IIS sites on both strands, and checks that four-base Golden Gate overhangs pair directionally. It does not design your overhangs for you, does not choose junction sequences to maximise fidelity, and does not model secondary structure in the overlap — so a design it passes is one whose arithmetic is right, not one guaranteed to assemble. Run it, then read the sequences yourself.