Cancer immunotherapy and single-cell genomics, explained from zero

0.1 · A cell, an instruction book, and why two cells differ

Your body is built from roughly thirty trillion cells. A cell is a bag of water and machinery, wrapped in a membrane, with a control room in the middle called the nucleus. Inside the nucleus sits DNA: a very long chemical chain whose sequence of four letters — A, C, G, T — spells out instructions for building every protein the body can make.

A gene is one stretch of that chain: the instructions for one product. Humans have roughly twenty thousand genes. The complete set of DNA is the genome.

Recipes do not leave the library. To use one, the cell makes a disposable working copy called messenger RNA (mRNA) — a single-stranded chemical copy of that one gene. The copy travels out of the nucleus to a machine called a ribosome, which reads the letters three at a time and hooks together the corresponding chain of amino acids. That finished chain folds into a protein — the actual working object: a structural beam, an enzyme, a pump, a receptor.

The two steps have names you will see constantly. Copying DNA into RNA is transcription. Building protein from RNA is translation. Together, DNA → RNA → protein is called the central dogma of molecular biology.

0.2 · Badges on the outside: how cells recognise each other

Some proteins are not made for the inside of the cell. They are pushed into the outer membrane so that part of them sticks out into the world. These are surface proteins, and they are how cells identify one another — like a uniform with rank badges.

Biologists number these badges with a standard naming system: CD followed by a number. CD3 marks T cells. CD4 and CD8 mark the two main T-cell subtypes. CD19 marks B cells. When a paper says "CD19-positive", it simply means "wearing the CD19 badge".

Two more words you need. An antigen is anything the immune system can recognise and react to — usually a specific surface protein. An antibody is a Y-shaped protein made by the immune system whose two arm-tips are shaped to grip exactly one antigen and nothing else. That grip is the most precise recognition event in biology, and Module 3 will show you how engineers steal it.

0.3 · The immune system, in one page

The immune system has two halves. The innate half is fast, blunt and reacts to general danger signals. The adaptive half is slow, specific, and learns. Cancer immunotherapy is almost entirely about the adaptive half, and within it, about one cell: the T cell.

A T cell carries a single type of receptor on its surface, the T-cell receptor (TCR), and each T cell's TCR has a randomly generated shape. Your body holds millions of T cells with millions of different TCR shapes, so that whatever new threat appears, some T cell somewhere already happens to fit it.

But a TCR cannot simply touch any protein. It only reads protein fragments that have been loaded into a display case on the target cell's surface, called MHC (in humans also written HLA). Every cell continuously chops up samples of its internal proteins and displays the fragments in MHC — effectively holding up a summary of what it is making inside. A T cell patrols, reads those displays, and reacts if it sees something it recognises as wrong.

The two main T-cell types divide the work. CD8 T cells are the killers: when their TCR recognises a displayed fragment, they punch holes in the target cell using a protein called perforin and inject enzymes called granzymes that force the target to self-destruct. CD4 T cells are the coordinators: they release chemical messages that recruit and amplify everyone else. Those chemical messages are cytokines — small signalling proteins such as interferon-gamma (IFN-γ), tumour necrosis factor alpha (TNF-α) and interleukin-2 (IL-2). Measuring cytokines is one of the three pillars of Module 5.

T cells also age. A freshly recruited one is naïve; after fighting it can become a long-lived memory cell that responds fast next time; and after fighting for too long without rest it becomes exhausted — still present, still bearing its receptor, but no longer killing well and no longer dividing. Exhaustion is the central failure mode of cancer immunotherapy and reappears in every module of this guide.

0.4 · What cancer actually is

Cells divide under strict control. A mutation is a typo in the DNA — one letter changed, a chunk deleted, a piece moved. Most typos are harmless. A few land in genes that control division: an oncogene is an accelerator stuck on, a tumour suppressor is a brake that has been cut. A cell with enough of these typos divides when it should not, ignores stop signals, refuses to die on schedule, and produces a growing population of copies of itself. That population is a tumour.

A tumour is not only tumour cells. It is a small disordered organ containing blood vessels, structural cells, and large numbers of immune cells that have been talked into helping rather than attacking. That whole neighbourhood is the tumour microenvironment — and mapping it is exactly what Modules 1 and 2 exist to do.

Tumours survive the immune system by several concrete tricks: stopping production of the MHC display case so T cells cannot read them; wearing "stand down" badges such as PD-L1 that switch off any T cell that touches them; recruiting suppressive cell types; and flooding the area with inhibitory signals such as TGF-β.

0.5 · What a CAR-T cell is

A chimeric antigen receptor — CAR — is an artificial receptor that does not exist in nature. "Chimeric" means assembled from parts of different origins, and that is literally what it is: the gripping tip of an antibody, fused to the internal signalling machinery of a T-cell receptor.

Read from outside to inside, a CAR has four parts:

• The recognition tip — an scFv (single-chain variable fragment), which is an antibody's two gripping regions joined by a flexible linker into one compact piece. This is the part that decides which antigen the cell will attack.
• The hinge and transmembrane segment — a stalk that sets how far the tip reaches and anchors the whole thing through the membrane.
• The costimulatory domain — usually taken from 4-1BB or CD28. This is the "and really mean it" signal; without it the cell activates once and then quits.
• CD3ζ (CD3 zeta) — the piece that normally transmits "target found" from a real T-cell receptor into the cell's interior. It is the ignition wire.

The effect is a T cell that binds its target directly, the way an antibody does, without waiting for the target cell to display anything in MHC. This bypasses the tumour's favourite escape route in one stroke — but it also means the CAR can only ever see antigens physically present on the cell surface, which is why target selection in Modules 1 and 2 matters so much.

The clinical process is a loop: take T cells from the patient's blood (autologous) or from a healthy donor (allogeneic), install the CAR gene, grow the cells to a dose of hundreds of millions, test them, and infuse them back. Modules 3, 4 and 5 are the three stages of that loop.

0.6 · The vocabulary, collected

Everything above, in one place. Come back to this table whenever a later module uses a word you have lost.

1.1 · The question this method answers

You have a piece of tumour. It contains cancer cells, several kinds of immune cells, blood-vessel cells and structural cells, all mixed together. You want to know: which cell types are in there, in what proportions, and what is each type doing right now?

The older approach was to grind the whole piece up and measure the RNA of the mixture. That method is called bulk RNA sequencing, and it has a fatal flaw for this question.

1.2 · Words first

Read this table before the rest of the module. Every term below appears in the sections that follow.

1.3 · How the machine physically separates forty thousand cells

The problem is separation. Forty thousand cells must each end up in their own private container, and each container must add a different sticker. Doing that with forty thousand real test tubes is impossible. The solution is to make the test tubes out of droplets, and make them by the thousand per second.

Inside the chip, three channels meet at a T-shaped junction. One channel carries the cell suspension. A second carries the gel beads. The third carries oil, moving crosswise. Where the water streams meet the oil, the oil pinches the water into separate droplets, in the same way a dripping tap breaks a stream into drops — except this happens hundreds of times per second, and each droplet holds about one nanolitre, roughly a millionth of a raindrop.

Two streams of water meet a stream of oil, and the oil chops the water into a rapid train of identical bubbles. Each bubble is a reaction vessel.

Every gel bead is coated with about 100 million short DNA primers. All the primers on one bead are identical, and all differ from those on every other bead. Each primer is built from four segments joined end to end:

[ sequencing handle ]-[ CELL BARCODE ]-[ UMI ]-[ poly-T ]
       ~29 letters        16 letters     12 letters   30 T's
   lets the sequencer   same on every    different    grabs the
     grab the piece     primer on this   on every     poly-A tail
                        one bead         primer       of any mRNA

Now follow one droplet. It contains one cell, one bead, and a chemical that dissolves the cell membrane. The cell bursts; its mRNA floats free; the poly-T end of each primer catches a poly-A tail. Reverse transcription copies each captured RNA into DNA — and because the copy is made from the primer, the finished DNA copy permanently carries that bead's cell barcode and one UMI at its front end. Then the oil is broken, everything from all droplets is pooled into one tube, amplified, and sequenced together.

Each cell's messages get stamped with the name of the cell they came from, before everything is mixed. That is why mixing afterwards is harmless: the computer sorts the reads back out by reading the stamp.

Why some droplets contain two cells

Cells are loaded deliberately dilute — so dilute that most droplets get no cell at all. Typically about 88 in every 100 droplets are empty, 11 hold exactly one cell, and fewer than 1 hold two or more.

That sounds wasteful, and it is, but it is the only way to keep doublets rare. Loading cells more densely would fill more droplets, and would also put two cells together far more often. Since two cells in one droplet share a barcode, they arrive in the data looking like a single cell that is somehow both a T cell and a tumour cell — a chimera that does not exist. Wasting 88% of droplets is the price of trusting the remaining 11%.

Rain falling on a grid of cups: to be confident that no cup catches two drops, you accept that most cups catch none.

In practice the doublet rate scales with how many cells you recover: about 0.8% per 1,000 cells. Load for 10,000 cells and roughly 8% of your "cells" are actually pairs. That is the number you trade against cost.

1.4 · What actually happens at the bench

1. Turn tissue into single floating cells Chop the sample and add enzymes that dissolve the glue between cells. This is the step that decides whether the experiment is any good. Target more than 90% living cells and a concentration of roughly 700–1,200 cells per microlitre. For fragile samples such as brain tumours, use a protease that works at 6 °C — the cold stops cells from reacting to the handling.
2. Load the chip Cells, beads and oil go into their wells. Load for the number of cells you intend to recover, remembering the doublet cost of roughly 0.8% per 1,000 recovered cells.
3. Make the copies inside the droplets Warm the emulsion to about 53 °C for 45 minutes so reverse transcription runs, then to 85 °C for 5 minutes to stop it. At this point the information is safely stamped and the droplets have done their job.
4. Break the oil and pool Recover all the stamped cDNA into one tube. From here on it is ordinary molecular biology.
5. Amplify 11–14 cycles of PCR. Fewer cycles if you loaded many cells, more if you loaded few. Too many cycles wastes sequencing on duplicates; too few gives too little material.
6. Build the library Cut the cDNA into sequencer-sized pieces and attach the adapters the machine needs to grip them.
7. Sequence The machine is told to read 28 letters on one end — exactly the barcode plus UMI — and about 90 letters from the other end, which is the actual gene sequence. Budget 20,000–50,000 read pairs per cell.

1.5 · What comes out, and how it becomes biology

The sequencer produces a very large text file. Software — Cell Ranger is the standard — sorts it by cell barcode, collapses duplicate UMIs, matches each read to a gene, and produces the count matrix: genes down the side, cells across the top, and in each square, the number of original molecules of that gene found in that cell. That table is the raw material for everything afterwards.

One subtlety: the machine produces hundreds of thousands of barcodes, but most correspond to empty droplets that captured only ambient RNA. Distinguishing "a real cell" from "an empty droplet with background" is a statistical decision, made by an algorithm called EmptyDrops. It is a decision, not a measurement, and it can be wrong in either direction.

The analysis chain, in the order it is performed

1. Remove bad cells Drop cells with too few genes (broken), too many genes or UMIs (probably doublets), or a high mitochondrial percentage (dying). Set thresholds by looking at your own data's distribution, never by copying numbers from a paper on a different tissue.
2. Normalise One cell may yield twice the material of another for purely technical reasons. Normalisation converts raw counts into comparable values so that "more RNA recovered" is not read as "more active cell".
3. Choose the informative genes Most genes are equally expressed everywhere and carry no information about identity. Keep the 2,000–3,000 that vary most between cells, and work only with those.
4. Reduce the dimensions Even 2,000 measurements per cell cannot be drawn. PCA compresses them into 30–50 summary axes that capture most of the variation; UMAP then compresses those into the two axes you actually plot.
5. Correct for batch, if needed Cells from two patients processed on different days will separate by day rather than by biology. Integration methods (Harmony, scVI) align the datasets — but every correction risks erasing real differences alongside technical ones, so it must be checked, not assumed.
6. Cluster An algorithm (Leiden) groups cells with similar profiles. The number of groups depends on a resolution setting you choose — so the number of cell types you "find" is partly your decision, and should be justified by markers, not by the plot looking tidy.
7. Name the clusters For each cluster, find the genes it expresses far more than the others, and match them against known biology: CD3E means T cell, CD8A means killer T cell, CD19 or MS4A1 means B cell, CD14 or LYZ means monocyte or macrophage. Automated tools (Azimuth, SingleR, CellTypist) propose names, but a human must confirm them.

2.1 · The question this method answers

Module 1 begins by dissolving the tissue into a soup of floating cells. That step is unavoidable for droplet capture, and it permanently destroys the architecture. You end up knowing that the tumour contained 22% killer T cells — but not whether those T cells were inside the tumour doing their job, or stuck at the edge, unable to get in.

That distinction is not a detail. For a cell therapy it is frequently the whole result.

2.2 · Words first

2.3 · Two completely different ways to keep the address

Every spatial platform belongs to one of two families, and they solve the problem from opposite directions. Understanding which family a method belongs to tells you almost everything about its strengths and its failure modes.

Family 1 — bring the RNA to a labelled grid (capture-based)

The slide is printed in advance with thousands of spots, each carrying millions of primers, and each spot's primers share a barcode that encodes that spot's coordinates. A tissue section is placed on top and photographed. The section is then permeabilised, RNA falls straight down onto whatever spot is beneath it, and is copied there — picking up that spot's positional barcode exactly the way a droplet's bead picked up a cell barcode in Module 1.

You lay the tissue on numbered graph paper, let its contents drop through, and afterwards read off which square each message landed on.

The strength is that you measure every gene, because nothing had to be chosen in advance. The weakness is resolution. On the standard Visium slide there are 4,992 spots, each 55 µm across with 100 µm between centres — and 55 µm is several cells wide. One spot may contain a tumour cell, a T cell and a macrophage, and their signals arrive mixed. Recovering the composition requires computation, not measurement. Newer versions push the grid much finer (Visium HD at 2 µm; Stereo-seq at 220 nm), which relieves but does not abolish the problem.

Family 2 — leave the tissue alone and photograph the genes (imaging-based)

Nothing is extracted. Fluorescent probes are added that bind their target RNA molecules inside the intact section, and the slide is photographed. Then those colours are washed away, a new set added, and another photograph taken — round after round.

No single photograph identifies a gene. The pattern across rounds does. If gene A is designed to light up in rounds 1, 3 and 6, and gene B in rounds 2, 3 and 5, then a dot that shone in rounds 1, 3 and 6 was gene A. This is the codebook, and it is the same principle as Morse code: information in the pattern, not the flash.

You photograph the tissue many times with different colour sets, and identify each gene by the sequence of rounds in which its dot lit up.

The codebook is also the error check. Some patterns are deliberately left unassigned — they correspond to no gene at all. If dots are being read as those blank patterns, you are measuring your own error rate directly, which is a rare and valuable thing in biology.

The strength is resolution: you see individual RNA molecules at their true positions, well inside single cells. The weakness is that you only ever measure the genes you chose in advance — typically 300 to 5,000 — and choosing wrongly means you will never see what you did not ask for.

2.4 · At the bench: the step everything depends on

For a fresh-frozen capture experiment, the sequence is:

1. Freeze the tissue fast Cold isopentane, not a domestic freezer. Slow freezing grows ice crystals that tear cells apart and destroy both the architecture and the RNA.
2. Cut and mount 10 µm sections placed carefully within the printed capture area. Anything hanging outside the area is simply not measured.
3. Fix and photograph Methanol fixation, then a standard stained image. This photograph is not decoration — it is the anatomical reference onto which every later measurement is laid.
4. Find the right permeabilisation time — the critical step Holes must be punched in the cells so RNA can fall onto the slide. The optimal duration differs for every tissue type and must be determined in a separate experiment beforehand, testing a series such as 3, 6, 12, 18, 24 and 30 minutes.
   • Too short: little RNA escapes. Low counts. Obvious, and therefore harmless — you repeat it.
   • Too long: RNA escapes and then spreads sideways before it lands. Counts look excellent. The signal is now in the wrong place.
5. Copy the RNA on the slide Reverse transcription happens in place, so every copy picks up the barcode of the spot beneath it.
6. Build the library and sequence Budget roughly 50,000 read pairs for each spot that actually has tissue over it.

For archived hospital tissue (FFPE), the chemistry is different: the RNA is too damaged to copy, so instead pairs of probes are laid onto each target gene and joined together only if both halves found their match. The joined probe is what gets captured and counted. This covers around 18,000 genes and works on decades-old blocks — which is what makes retrospective study of rare paediatric tumours possible at all.

For imaging platforms, the section is fixed, probes are hybridised, the signal is amplified into visible dots, and the instrument runs its rounds of photograph-and-wash automatically — typically overnight — producing a list of transcripts with x, y coordinates, plus a stain outlining nuclei to help with segmentation.

2.5 · Turning the pictures into biology

Capture data must first be unmixed. Deconvolution methods take a single-cell dataset from Module 1 as a reference — a catalogue of what each cell type's expression looks like — and estimate, for each spot, what mixture of those types would produce the observed signal. The obvious implication is usually missed: a deconvolution result is only as good as its reference. If a cell type is absent from your reference, it cannot be assigned, and its signal will be misattributed to whichever type it happens to resemble.

Imaging data must first be segmented. Classical methods find nuclei and expand outwards by a fixed distance to guess the cell body; newer methods use the transcripts themselves to decide where one cell ends and the next begins. Neither is exact, and in dense tissue the error is substantial.

After that, both families support the same analyses:

• Spatial domains — grouping neighbouring spots or cells into regions, which typically recovers tumour core, invasive margin, stroma and normal tissue without being told they exist.
• Neighbourhood statistics — asking whether two cell types are found together more often than chance allows. This is the claim most papers actually rest on.
• Cell–cell communication — looking for a signal molecule made by one cell and its matching receptor on a neighbouring cell.

3.1 · The question this method answers

You have decided the T cell needs a new receptor, and you have decided that some of its own genes should be switched off. Both require changing DNA at a chosen position, in a living human cell, without wrecking the rest of the genome.

3.2 · Words first

3.3 · How Cas9 finds one address in three billion letters

The human genome is roughly three billion letters long. A twenty-letter search string is, by itself, plenty to be unique. But Cas9 cannot read the whole genome and compare — it must search physically, by bumping into DNA at random.

The trick is that it does not check the guide first. It checks for the PAM first: a short fixed sequence, NGG, occurring roughly every eight letters. Cas9 collides with DNA, asks only "is there a PAM here?", and if not, lets go immediately. Only at a PAM does it pause and begin prying the two strands apart to compare them against its guide — and even then it starts from the PAM end and gives up at the first mismatch.

Cas9 does not search for your sequence directly. It searches for a cheap marker that appears everywhere, and only bothers to read the address at those points. That is what makes finding one site in three billion physically possible.

This explains two things you need to know. First, you cannot target any sequence you like — there must be a PAM in the right place, which constrains guide design. Second, mismatches close to the PAM kill the reaction, while mismatches at the far end are often tolerated. So an off-target site that differs from your guide in three distant letters may still be cut, while one that differs in a single PAM-proximal letter will not be. Off-target prediction is about position, not just count.

When the match holds, two separate cutting modules in the protein each cut one strand, three letters in from the PAM, producing a clean double-strand break.

3.4 · The part you do not control: repair

A double-strand break is, for a cell, an emergency on the level of a fire alarm. It has three ways to deal with it.

Route 1 — fast, sloppy rejoining (NHEJ)

The dominant route, available at all times. The two ends are trimmed and pushed back together, frequently gaining or losing a few letters in the process. Roughly two-thirds of such joins produce a frameshift and therefore destroy the gene. This is how you knock a gene out — not by removing it, but by making its instructions unreadable.

Route 2 — repair using a nearby repeat (MMEJ)

If short repeated sequences flank the break, the cell can line them up and delete everything in between. Predictable enough that software can forecast the outcome from the target sequence alone.

Route 3 — accurate repair from a template (HDR)

The cell copies from a template with matching flanks. If you supply your own template, whatever you put between those flanks is written into the genome. This is how you knock something in — but it only operates in cells actively preparing to divide, which is why resting T cells must be woken up before editing (Module 4) and why HDR is always the minority outcome.

3.5 · Why cutting is more dangerous than it looks

Standard analysis of an edit amplifies a few hundred letters around the cut and counts the small insertions and deletions. It is quick and it is systematically reassuring, because it can only detect damage small enough to remain inside the amplified window.

Cutting also produces damage that this method is blind to by construction:

• Large deletions — thousands of letters removed. Both ends of the amplified window disappear, so no product forms, and the cell silently drops out of the analysis instead of being counted as damaged.
• Loss of one whole chromosome arm, which removes the healthy copy of every gene on it.
• Translocations — two breaks in different chromosomes repaired to each other, fusing them. The risk rises with the number of simultaneous cuts, which is exactly what multiplexed editing of allogeneic products requires.
• Chromosome shattering after a break in a chromosome that gets isolated from the nucleus.
• p53 activation: cells that repair the break best are the ones with a weakened damage-response — so editing quietly selects for cells with impaired tumour suppression.

3.6 · Editing without cutting

If the cut is the source of the danger, the obvious move is to remove the cut.

Base editors — change one letter chemically

Take Cas9, disable one of its two cutting modules so it can only nick a single strand, and bolt on an enzyme that chemically converts one DNA letter into another. The complex still finds the target with the same guide, but instead of cutting, it performs chemistry on the letters that happen to lie in a narrow window — usually positions 4 to 8 along the guide.

Two families exist: cytosine base editors convert C to T, and adenine base editors convert A to G. Between them they cover a large share of useful single-letter changes, with no double-strand break, no template needed, and no dependence on cell division.

Base editors bring their own failure modes, and they are different in kind from Cas9's. A second letter of the same type inside the window is changed too — a bystander edit. The chemistry can also act on DNA that Cas9 never visited, wherever a strand happens to be temporarily single-stranded; this off-target damage is scattered across the genome and cannot be predicted from the guide sequence at all, which means guide redesign does not fix it and a genome-wide assay is the only way to see it.

Prime editing — genuine find and replace

Prime editing takes the nickase further: it is fused to an enzyme that writes DNA from an RNA template, and the guide itself is extended to carry that template. The guide now says both where to go and what to write.

The guide finds the site, the nickase cuts one strand, one end of the extended guide grips the freed DNA end, and the copying enzyme writes the new sequence directly onto it. The cell then resolves the two competing versions, and about half the time it keeps the new one.

This performs any small substitution, insertion up to roughly forty letters, or small deletion — without a double-strand break and without a separate template molecule. The cost is efficiency: prime editing is generally less efficient than base editing and needs more optimisation per target.

3.7 · Building the gene you are going to install

Before any of this can be delivered, the CAR must exist as a physical piece of DNA, assembled in the right order and carried in a plasmid. Read from the front, the construct is:

[promoter] [signal] [ scFv: VH -linker- VL ] [hinge] [membrane] [costim] [CD3ζ] [ 2A ] [marker]
    the       sends     the antibody grip        stalk    anchor    "mean    ignition   splits  lets you
   on-switch  it to     that picks the target                        it"       wire      one    see which
              the                                                                       gene    cells took
            surface                                                                    in two    the gene

Two constraints dominate the design. First, the delivery virus in Module 4 has a payload ceiling of about 8–9 kilobases including its own required sequences, so every extra feature competes for space with the CAR itself. Second, every junction between parts must preserve the reading frame; a single misplaced letter between two domains shifts the frame and produces a protein that is nonsense from that point on.

Three assembly methods are in common use:

Cut and paste with restriction enzymes

The classical method. Enzymes cut at fixed recognition sequences, the pieces are mixed with a joining enzyme, and the result is put into bacteria. Reliable, cheap, and limited by the requirement that the necessary cut sites exist in the right places and nowhere else.

Gibson assembly

Each piece is made with 20–40 letters of overlap with its neighbour. One mixture of three enzymes chews back the ends, lets the overlaps find each other, fills the gaps and seals them. One tube, one hour, no cut sites required, several fragments at once. This is the default for building a new CAR from scratch.

Golden Gate assembly

Uses enzymes that cut outside their own recognition sequence, so the cut site is removed as the pieces join and cannot be re-cut. Cutting and joining therefore run simultaneously in one tube, driving the reaction toward the finished product. This allows ten to twenty fragments to be assembled in a defined order in a single reaction, which makes it the method of choice for building libraries of receptor variants that differ in one domain.

Verification, in the order it is done

1. Cut with enzymes and look at the fragment sizes — five minutes, catches gross assembly errors.
2. Sequence the junctions — confirms the pieces joined in frame.
3. Sequence the entire plasmid — the only step that detects backbone rearrangement.
4. Check protein expression on the cell surface — the construct can be perfect on paper and still not reach the membrane.
5. Check it functions — surface expression does not prove signalling. That test is Module 5.

4.1 · The question this method answers

Module 3 produced a gene on a plasmid. It is useless there. It has to end up inside a T cell, integrated into that cell's own DNA so the cell keeps it permanently and passes it to its descendants — and there must eventually be several hundred million such cells, all still healthy enough to work.

4.2 · Words first

4.3 · Getting the passengers ready

Separating the cells you want

Blood is layered gently onto a liquid of precisely chosen density and spun. Red cells and granulocytes, being denser, drive through to the bottom. The immune cells you want are less dense and stop as a thin white band right on top of the layer, from where they are pipetted off.

Two details in the protocol are not optional. The spin is done with the centrifuge brake switched off, because braking creates turbulence that remixes the layers you just separated. And blood should be processed within about eight hours; after that, cell viability and function fall measurably.

Spin the blood on a liquid of a chosen weight, and the cells sort themselves into layers. Take the band you want. Do not brake, and do not leave the blood overnight.

Waking them up — with two signals, not one

A resting T cell will not divide and will not accept the courier. Waking it requires two signals at the same time: signal 1 says a target has been detected, signal 2 says the situation is genuine. Signal 1 alone does not produce a weak response — it produces the opposite, a state of unresponsiveness. This is a genuine safety mechanism, and in the laboratory it simply means both signals must always be supplied together, which is why the standard reagent is a bead carrying both antibodies on the same surface.

The beads are removed after five to seven days. Leaving them in place means the cells are being shouted at continuously, and continuous stimulation is the fastest route to exhaustion.

The choice that decides whether the product works

The medium contains cytokines, and which ones you choose is the most consequential decision in the whole manufacturing process.

4.4 · Building the courier

A lentiviral vector is a virus with everything dangerous deleted. What remains is the ability to enter a cell and insert cargo into its DNA. Two design principles make this safe enough for clinical use.

Split the parts across separate plasmids. The genes for the virus's internal machinery, its copying enzymes, its regulator and its coat are placed on four different plasmids, and only the transfer plasmid carries the signal that says "package me". A particle therefore carries only your gene, and there is no single molecule anywhere that contains a complete virus. Reconstructing one would require several improbable recombination events in the same cell.

Delete the virus's own promoter. The repeated sequences at the ends of the inserted cargo have their promoter removed, so once integrated into the target cell the inserted DNA cannot drive its neighbours. This prevents the inserted vector from switching on a nearby gene — a failure that caused leukaemias in early gene-therapy trials and is the reason the design exists.

1. Transfect production cells All four plasmids are pushed into a laboratory cell line grown to 70–80% coverage of the flask. These cells assemble the particles for you.
2. Collect the supernatant The particles are released into the liquid, which is harvested at about 48 and again at 72 hours and filtered. Use a 0.45 µm filter of a material that does not bind protein; the wrong filter silently removes much of your product.
3. Concentrate Spin the filtered liquid through a dense sugar layer at very high speed, then resuspend in a small volume. Freeze in single-use portions — every freeze–thaw cycle costs roughly half the working titre, so a tube opened four times is close to worthless.
4. Measure the titre Count working particles by transducing test cells at several dilutions and measuring what fraction took up the gene.

The other courier: AAV

AAV is smaller, milder, and mostly does not insert into the genome — its cargo persists as a separate circle, which is safer but gradually lost in dividing cells. Its defining limitation is size: about 4.7 kilobases including the sequences it needs itself, which leaves too little room for most CAR constructs.

Its role in this field is therefore specific and important: it is the best way to deliver a repair template to a cell that has just been cut by Cas9 (Module 3). The template does not need to be permanent — it only needs to be present in large amounts at the moment the cell is repairing the break.

4.5 · The meeting point: transduction

The virus must physically reach the cell, and in a liquid it mostly does not — diffusion is slow and the particle decays. Every technique below exists to solve that one problem.

Checking that it worked — and when to check

You need three numbers. What fraction of cells carry the receptor, measured by staining for the receptor itself. How many copies each cell received, measured by a DNA assay — regulators require an average below five, and a product above that is not releasable regardless of how well it performs. Whether expression lasts, checked again at two and four weeks, because inserted genes can be switched off by the cell over time.

5.1 · The question this method answers

You have a bag of engineered cells. Three things are unknown, and they are genuinely separate questions that no single experiment answers:

• What is in the bag? How many cells carry the receptor, and what state are they in — young, or already worn out?
• Do they kill? Put them with cancer cells and count what survives.
• What do they say while doing it? Which chemical messages are released, in what amounts — which predicts both efficacy and the toxicity a patient may experience.

5.2 · Words first

5.3 · Flow cytometry: measuring cells one at a time

The instrument forces cells into single file. The sample is injected into a faster-moving sheath fluid, which squeezes the stream into a thread narrow enough that cells must pass in single file — a physical arrangement, not a filter, which is why it can process thousands of cells per second without clogging.

Each cell crosses one or more laser beams. Two measurements come free: light scattered forwards, which tracks size, and light scattered sideways, which tracks internal complexity. Beyond that, you stain the cells beforehand with antibodies carrying coloured dyes, each antibody gripping one surface badge. If a cell wears CD8, its CD8 antibody is bound, and its dye fires when the laser hits. Detectors record how much light of each colour came from each cell.

Cells go past a laser one by one. Size and internal texture are measured automatically; everything else is measured because you attached a coloured tag to it in advance.

Why compensation is necessary, and what it does not fix

Dyes are not clean. A dye meant for the green detector also emits some light that the orange detector picks up. Compensation measures that leak using single-stained samples and subtracts it arithmetically.

The controls that decide whether your result means anything

• An unstained sample — shows the cells' own background glow.
• Single-stained samples — one per dye, required to calculate compensation.
• An FMO control — stained with everything except the marker in question. This, not an isotype control, is what tells you where the boundary between negative and positive actually falls in your specific panel. The isotype control was designed to answer a question about antibody stickiness, and it does not answer the question people use it for.
• Untransduced cells from the same donor — the only honest baseline for how bright a CAR stain looks in a cell that definitely has no CAR.
• A cancer cell line lacking the target antigen — the control that separates real recognition from general stickiness, and the one most often missing from otherwise impressive data.

5.4 · Do they kill?

Engineered cells are mixed with cancer cells at a chosen ratio, and how many cancer cells remain is measured. The arithmetic is simple, and the subtraction matters: some target cells die on their own, so the assay reports killing caused by the treatment, not total death.

                    (measured death) − (death with no killer cells)
specific killing =  ─────────────────────────────────────────────── × 100
                    (death when all are killed) − (death with none)

Required control wells:
  targets alone .......... spontaneous death  → must stay below ~20–25%
  targets fully killed ... the maximum possible
  killer cells alone ..... their own signal, subtracted out
  untransduced T cells ... kills not caused by the receptor

5.5 · What do they say while doing it?

Killing is one output; the chemical messages released are another, and they carry information the killing assay cannot. Four methods, answering different questions:

ELISA — how much of one protein is in this liquid?

Antibodies capture the protein, a colour develops in proportion to the amount, and the amount is read against a standard curve. Precise, cheap, one protein at a time. The routine error is measuring a single dilution: if the true concentration sits off the top of the curve, the number you record is not the concentration, it is the ceiling.

ELISpot — how many cells released it?

Cells sit on an antibody-coated membrane; each responding cell leaves a spot. You count responders rather than total output. This distinguishes "a few cells shouting" from "many cells speaking" — two situations with identical ELISA readings and completely different biology.

Multiplex bead assays — what is the whole pattern?

Dozens of proteins measured at once from a very small volume. Essential when sample is limited, as it always is with paediatric material, and the only practical way to see the shape of a response rather than one of its components.

Intracellular staining — which individual cells make which messages?

Release is blocked chemically so the proteins accumulate inside, then the cells are permeabilised and stained and read by flow cytometry. This is the only method that shows polyfunctionality — single cells making several cytokines at once, which correlates with clinical response better than the total amount of any one of them.

6.1 · The loop

6.2 · The controls nobody should have to remind you about

• Module 1 — a check that your interesting cluster is not defined by stress genes; a doublet assessment; background correction using the empty droplets.
• Module 2 — a permeabilisation series run in advance; a sharply bounded marker used to prove the signal has not smeared; a distance limit on every claim about cells signalling to one another.
• Module 3 — whole-plasmid sequencing, not just junctions; an assay capable of seeing damage larger than the amplified window; a translocation test whenever more than one cut is made.
• Module 4 — titre determined on the cell type you will actually use; transduction read at day 8 or later; copy number measured, not assumed.
• Module 5 — untransduced cells from the same donor; a target line lacking the antigen; FMO rather than isotype; spontaneous and maximum release wells in every killing assay.

6.3 · Short answers worth having ready

Why single-cell rather than bulk sequencing?

Bulk gives an average across a mixture, which can correspond to many different mixtures. Single-cell keeps each cell separate, so composition and cell state are both recoverable.

What is the difference between a read and a UMI?

Reads count photocopies produced by amplification; UMIs count the original molecules. Only UMI counts are biologically meaningful.

Why add spatial work if you already have single-cell data?

Dissociation destroys position. Whether engineered cells penetrated the tumour, whether the target is uniform, and which neighbours suppress the response are all questions that only survive if position is kept.

Why is knock-in harder than knockout?

Knockout only needs the cell's always-available sloppy repair to be imprecise. Knock-in needs accurate template-based repair, which operates only in dividing cells and always competes with the sloppy route.

Why use base editing instead of Cas9 for multiple knockouts?

Each Cas9 cut is a double-strand break, and the risk of two breaks being joined together rises steeply with the number of simultaneous cuts. Base editors disable genes without producing breaks at all.

Why is a large cell expansion not good news?

Vigorous expansion in culture correlates with differentiation and exhaustion. Cells that have already divided many times divide less after infusion, which is where division matters.

Why is transduction efficiency measured after day 8?

Because virus particles deliver protein as well as genes, and that protein makes cells appear positive for two to three days without any gene having integrated.

Why does a construct that kills perfectly in vitro fail in vivo?

Because the in-vitro assay was usually short, run at a high killer-to-target ratio, and performed once. Long, low-ratio, repeatedly rechallenged assays discriminate far better.

What is the first thing to check when a patient relapses?

Whether the tumour still carries the target. If it does not, the receptor must change. If it does, the cells failed — and that is a persistence problem, not a targeting one.