Classical restriction cloning hit a wall around three-fragment assemblies. Each fragment needed a unique pair of compatible restriction sites at its ends, the host plasmid had to be free of those sites, and every junction left a scar — the recognition sequence itself, deposited at the seam between fragments. For pathway-scale work (five, ten, twenty parts), the combinatorics of finding non-clashing site pairs collapsed.
Golden Gate, introduced by Engler, Kandzia, and Marillonnet in 2008, sidesteps both constraints by using Type IIS restriction enzymes — BsaI, BbsI, SapI, and the like — instead of the more familiar Type II enzymes (EcoRI, BamHI).
Type II enzymes (most of what you learned in undergrad) bind a palindromic recognition site and cut inside it. The cut is symmetric; the recognition sequence is preserved on both products.
Type IIS enzymes bind a non-palindromic site and cut outside it, at a defined offset. BsaI, for example, recognizes GGTCTC and cuts one base downstream, leaving a 4-nucleotide 5′ overhang:
5'... G G T C T C N | N N N N ... 3'
3'... C C A G A G N N N N N | ... 5'
The four-base overhang is whatever you designed it to be. There’s no constraint that it match the recognition site, because the enzyme isn’t cutting inside its site at all.
Three consequences follow:
Scarless joins. Because the overhang is arbitrary, you can design adjacent overhangs in your fragments to encode functional sequence — a ribosome binding site, an operator, a junction inside a gene. The seam carries information instead of carrying a useless scar.
One-pot, one-enzyme, all fragments simultaneously. Add all your insert fragments (each flanked by BsaI sites pointing inward), the destination vector (with BsaI sites pointing outward), BsaI, ligase, and ATP to a single tube. The enzyme cuts every BsaI site; ligase joins compatible overhangs; products that re-incorporate a BsaI site get re-cut; only the desired final product, which has no remaining BsaI sites, is stable. The reaction runs to completion in a thermocycler with alternating cut/ligate temperatures.
Unique-overhang assembly scales linearly. With four-base overhangs, you have 256 possible sequences, of which ~240 are usable after excluding palindromes and the recognition site itself. Multi-part assemblies of 10–20 fragments are routine. Standard “parts kits” (MoClo, GoldenBraid) reserve canonical overhang sequences for canonical part roles — promoter, RBS, CDS, terminator — so parts from one lab compose cleanly with parts from another.
Internal sites are forbidden. Any BsaI site inside the fragment you’re trying to assemble gets cut. Domestication — silent mutations to remove internal Type IIS sites — is a required preprocessing step. Modern gene synthesis vendors will do this automatically if you flag the assembly standard.
Overhang design is now a real problem. Choosing 10–20 mutually distinguishable four-base overhangs is non-trivial; certain pairs ligate cross-reactively. Tools like NEB’s Golden Gate Assembly Tool optimize the set.
Not free. Type IIS enzymes are pricier than the workhorse Type II enzymes, and quality varies.
Gibson assembly is the obvious comparison — overlap-based, sequence-agnostic, no scar concerns at all, but harder to multiplex cleanly. The right next read is the side-by-side benchmarks in Casini et al. (2015), which still hold up.