Structural Biology from Zero · 6: Antibodies and the Immune System

The adaptive immune system recognizes a broad and changing set of molecular surfaces using a finite genome. These targets include components of pathogens as well as synthetic molecules with no evolutionary history in the host. Antibodies address this problem by combining a conserved protein architecture with genetically diversified binding loops and cellular selection. This article develops the structural basis of that repertoire and the mechanisms that refine useful binders.

The antibody: one fold, built for variation

An antibody, or immunoglobulin, in its common IgG form is a Y-shaped protein of about 150 kilodaltons. It contains two identical heavy chains and two identical light chains connected by disulfide bonds. Each tip forms a Fab ("fragment, antigen-binding") arm that contacts the target. The Fc ("fragment, crystallizable") stem interacts with receptors and other components of the immune system.

Every antibody chain consists of immunoglobulin domains. Each domain adopts the immunoglobulin fold, a β-sandwich of two packed antiparallel β-sheets stabilized by a conserved internal disulfide. A light chain has one variable ( $V_L$ ) and one constant ( $C_L$ ) domain. A heavy chain has one variable ( $V_H$ ) and three constant domains ( $C_{H}1$ , $C_H2$ , $C_H3$ ). The paired $V_H$ and $V_L$ domains form the antigen-binding site at each tip.

This is the key architectural idea: the fold is conserved and rigid, a dependable scaffold, and the variation that creates specificity is confined to small, tolerant regions on top of it. The immunoglobulin domain is the frame; the picture changes, the frame does not.

Where the diversity lives: the six loops

Look at a variable domain and you find that its β-sandwich core — the framework regions — is nearly the same from antibody to antibody. All the sequence variation is concentrated in three loops at one end of the domain, the loops that connect the β-strands at the tip. Elvin Kabat and Tai Te Wu discovered this in 1970 by lining up antibody sequences and simply plotting variability position by position: three sharp spikes per variable domain stood out against a flat, conserved background. They named the spikes the hypervariable regions; structurally, they are the loops now called complementarity-determining regions, or CDRs.

Each variable domain contributes three CDRs. Together, CDR-H1, H2, and H3 from the heavy chain and CDR-L1, L2, and L3 from the light chain form a continuous surface at the tip of the Fab. This surface is the paratope, the part of the antibody that contacts antigen. Variation in loop sequence, length, and conformation produces flat, grooved, and pocketed binding sites on a conserved framework.

One loop deserves singling out. CDR-H3, sitting right at the center of the paratope, is by far the most variable of the six in both sequence and length, and it usually dominates antigen contact. The reason is genetic, and it takes us to how the diversity is generated in the first place.

Generating the repertoire: combinatorics, then evolution

The genome stores antibody gene segments that are assembled combinatorially, the mechanism Susumu Tonegawa uncovered and that led to the 1987 Nobel Prize. Two stages provide much of the initial diversity and subsequent affinity improvement, alongside clonal selection, tolerance mechanisms, and receptor editing.

Stage one — build a diverse starting library (V(D)J recombination). The variable domains are not encoded as single genes but as libraries of interchangeable gene segments. The heavy-chain variable region is assembled by picking one V segment, one D segment, and one J segment from arrays of dozens each and splicing them together; the light chain does the same with V and J segments. Three sources of diversity stack here:

Combinatorial diversity — the sheer number of ways to choose one segment of each type, multiplied again by the pairing of any heavy chain with any light chain.
Junctional diversity — the joins between segments are deliberately sloppy: nucleotides are chewed back and random ones are added by the enzyme terminal deoxynucleotidyl transferase, scrambling the sequence exactly at the junctions.
The V–D–J junction is precisely where CDR-H3 is encoded, which is why that loop is the most variable of all six — it is built on top of the most chaotic joint in the process.

Together these processes can generate a potential repertoire exceeding $10^{11}$ sequences from a modest set of gene segments. Selection and tolerance then determine which B-cell receptors persist and respond to accessible antigens. The genome thereby encodes a generative process rather than one complete antibody gene per target.

Stage two — improve the winners (somatic hypermutation and affinity maturation). A starting binder pulled from the combinatorial library is usually mediocre, binding its target only weakly. So the immune system runs evolution on it. When a B cell's antibody binds an antigen, the cell is stimulated to proliferate in a structure called a germinal center, and during proliferation an enzyme (activation-induced cytidine deaminase) introduces point mutations into the variable-region genes at a furious rate — somatic hypermutation. B cells whose mutated antibodies happen to bind the antigen better are preferentially selected to survive and divide; those that bind worse are discarded. Iterated over days, this affinity maturation improves binding affinity by factors of a hundred to a thousand, turning a weak initial grip into a tight, specific one — often reaching sub-nanomolar dissociation constants. It is Darwinian selection, compressed from geological time into a week and running inside your lymph nodes. (This process — reading and optimizing the sequence–affinity relationship — is exactly where machine learning now meets antibody engineering, and it is close to my own day-to-day work.)

Combinatorial assembly provides broad initial diversity, while selection enriches clones that bind a particular antigen more effectively. Both processes contribute to the mature response.

The interface: recognition as complementarity

What does binding actually look like at the atomic level? An antibody grips its antigen through the same non-covalent forces that fold a protein (Part 3) — hydrogen bonds, salt bridges, van der Waals packing, and the hydrophobic effect — acting across a patch of contact. Two features of that patch matter.

First, complementarity. The paratope and the epitope (the patch on the antigen that it contacts) fit together in both shape and chemistry: bumps against hollows, a donor across from an acceptor, a positive charge facing a negative one. A typical antibody–antigen interface buries roughly $600$ – $900\ \text{Å}^2$ of surface on each side and involves on the order of fifteen to twenty residues from each partner, most of the antibody's contribution coming from the six CDR loops — and disproportionately from CDR-H3.

Second, the epitope is usually conformational, not linear. Because the antibody binds a surface, the epitope residues are typically brought together by the antigen's fold rather than being contiguous in its sequence — a discontinuous epitope. This is why antibodies raised against a folded protein often fail to recognize the same protein once it's denatured: the epitope only exists when the antigen is folded. Recognition is of three-dimensional shape, not of sequence — a direct consequence of everything in Parts 3 and 4.

The viewer below shows a real, classic complex: an antibody bound to hen egg-white lysozyme, the same small enzyme we've used throughout. Switch to the interface view and the abstraction becomes concrete — you can see the exact residues, almost all of them in the CDR loops, that the antibody uses to grip its target, and the matching epitope on the lysozyme surface.

Interactive. The HyHEL-10 antibody Fab bound to hen egg-white lysozyme (PDB 3HFM): heavy chain in blue, light chain in teal, lysozyme antigen in orange. "Highlight CDRs" colors the six complementarity-determining loops; "Interface" computes the contacting paratope and epitope residues live (heavy-atom contacts within 4 Å) and shows they fall almost entirely in the CDRs. Opens full-screen in a new tab.

The constant regions do the rest: isotypes and effector function

So far we've dwelt on the variable domains, because that's where recognition lives. But recognizing a target is useless unless something happens as a result, and that is the job of the constant regions — the rest of the heavy chain, which forms the Fc stem of the Y. The Fc is where the antibody stops being a binder and becomes a signal: once the Fab arms have flagged a target, the Fc is the handle that the rest of the immune system grabs to destroy it. The structural separation is clean and deliberate — variable domains for what to bind, constant domains for what to do about it — and it is connected by a flexible hinge region between the Fab arms and the Fc, which lets the two arms splay and swivel to grip two copies of an antigen at once.

The identity of the constant heavy chain defines the antibody's class, or isotype, and mammals make five: IgG, IgM, IgA, IgE, and IgD. They share the same variable-domain recognition machinery but carry different constant regions, and the differences are functional:

IgG is the workhorse of the bloodstream — the classic Y, the most abundant circulating antibody, and the scaffold of nearly every antibody drug. Its Fc is read by Fc receptors on immune cells (marking a target for engulfment) and triggers the complement cascade (tagging a target for destruction). It is also the isotype that crosses the placenta, giving a newborn its mother's immunity.
IgM is the first antibody made in a new response, and it assembles into a pentamer — five Y's joined at their Fc stems into a ten-armed star. Any single variable domain fresh out of V(D)J recombination binds weakly, but ten arms binding at once produce enormous avidity, so IgM compensates for low per-arm affinity with sheer multiplicity — a structural solution to the same weak-binder problem that affinity maturation solves by evolution.
IgA guards mucosal surfaces (gut, lungs, secretions, and breast milk), usually as a dimer.
IgE, though rare, binds mast cells and drives the allergic and antiparasitic responses.
IgD sits mostly on the surface of naïve B cells as a receptor.

Crucially, a single B cell can switch which constant region it uses — class switching — while keeping its variable domains, and therefore its hard-won specificity, intact. It swaps the effector module without touching the recognition module. This is the frame-versus-loops logic operating at yet another level: the same paratope can be redirected from complement activation (IgG) to mucosal defense (IgA) to allergic signaling (IgE) simply by changing the constant region it is attached to.

One more structural payoff of the Fc deserves mention because it explains why antibody drugs work at all: a receptor called FcRn binds the IgG Fc and recycles it away from degradation, giving IgG an unusually long serum half-life of about three weeks. That long half-life — a direct property of the constant-region structure — is a large part of why engineered IgG antibodies make practical, infrequently-dosed drugs, which is where we turn next.

From immunity to medicine: engineering antibodies

The modular organization of antibodies makes them amenable to therapeutic engineering. Frameworks, binding loops, Fc regions, and molecular format can each be modified for a particular application.

Humanization. Early monoclonal antibodies were made in mice, and their non-human sequence could provoke an immune response in patients. CDR grafting, introduced by Greg Winter's group in 1986, transfers mouse CDRs to a human framework. Framework selection and back-mutations are often needed to preserve loop conformation, affinity, and stability. Most current antibody drugs are humanized or fully human.
In vitro affinity maturation. The germinal-center loop — diversify, then select for binding — can be reproduced in a test tube. Display technologies like phage display present a library of antibody variants on the surface of phage particles, physically linking each variant to the gene encoding it, so that selecting for binding automatically selects the gene. Iterating diversification and selection matures affinity outside any animal. This is the in vitro analog of what your lymph nodes do, and it is increasingly guided by machine-learning models trained to predict which mutations will improve binding.
Smaller formats. The full Y is often more than you need. A single-chain variable fragment (scFv) fuses just $V_H$ and $V_L$ into one small binding module. Even more strikingly, camelids (camels, llamas) make antibodies with no light chain at all, whose single variable domain — a nanobody — is a fully functional ~15 kDa binder, small enough to reach clefts a full antibody cannot, and a favorite tool for stabilizing and crystallizing other proteins.

Each of these is a direct payoff of the structural logic: because specificity is localized in transplantable loops on a stable frame, antibodies can be cut down, reframed, and optimized almost like software.

Where this leaves us

The immune system solves an unbounded problem with a bounded genome by separating the invariant from the variable. One durable fold — the immunoglobulin β-sandwich — supplies a rigid, reliable scaffold; six loops on its tip supply near-limitless variation; combinatorial gene assembly generates enough starting breadth to bind almost anything; and somatic hypermutation with selection sharpens the good binders into tight ones. The molecule that recognizes the unknown is not a miracle of design but a beautifully organized application of the forces, folds, and selection we've built up over this series. It also sits squarely on the frontier of protein modeling: predicting how a CDR-H3 loop will fold, and which mutations will tighten an interface, is one of the places where structural biology and machine learning most urgently need each other.

We have now used protein structure to build machines, catalyze reactions, and recognize molecules. Throughout, we have trusted that we actually know these structures — that the atomic coordinates we've been rendering are real. In Part 7 we finally earn that trust, returning to the wall we hit in Part 1: you cannot see a protein with light. We'll see how X-ray crystallography, cryo-electron microscopy, and NMR get around that limit, straight from the physics.