ML in Gene Therapy

Engineering AAV when the search space is the problem

Ravi Kalia

Four Sequences That Worked

Four medicines that changed the last decade. Each depends on a deliberately chosen biological sequence.

ChAdOx1: The Vector Was Already There

A chimpanzee adenovirus, gutted and reloaded.

• Spike gene swapped in for E1
• Replication-deficient
• Shipped at fridge temperature

≈35 kb of DNA

Billions of doses. The virus is the courier, not the medicine.

The payload was written, not discovered — a spike sequence pasted into a vector backbone.

Voysey et al. (2021)

Start concrete and move fast. ChAdOx1 is a chimpanzee adenovirus precisely because human adenovirus seroprevalence would neutralise it — already a hint of the immunology problem that dominates the back half of this talk.

The thing to plant: the vector was off the shelf. What changed was the sequence dropped into it, and that took days once the genome was published.

BNT162b2: The Courier Disappears

No virus at all. A sequence in a lipid droplet.

• Nucleoside-modified mRNA
• Lipid nanoparticle delivery
• Transient by design

4,284 nucleotides

Designed from a published genome in days.

Uğur ŞahinBioNTech

Özlem TüreciBioNTech

Polack et al. (2020)

The contrast with the previous slide is the point: strip the virus away entirely and the sequence still works, as long as something gets it into a cell.

Worth naming the pseudouridine modification if the room is interested — it is what keeps the RNA from being shredded by innate sensors, and it is the same class of problem as the CpG story later in this talk.

Change the Letter, Keep the Message

Unmodified mRNA is read by the innate immune system as infection. It triggers interferon, and the message is destroyed before it is translated.

N1-methylpseudouridine can reduce innate sensing and improve translation. RNA purification and delivery also shape the response.

RNA chemistry and delivery determine how much protein the cell makes.

Katalin Karikówith Drew Weissman — Nobel Prize, 2023

Polack et al. (2020)

A whole slide on this because it is the same idea the deck returns to three more times.

Innate immunity does not inspect what your nucleic acid codes for. It pattern-matches on chemical features that distinguish “foreign” from “self” — unmodified uridine in RNA, unmethylated CpG in DNA, double-stranded RNA. Karikó and Weissman’s insight was that you can keep the message and change the features.

Hold that thought. When we get to CpG depletion in the AAV genome and the TLR9 response, it is exactly the same manoeuvre on a different sensor — and the same reason it works.

Delandistrogene Moxeparvovec: A Gene Cut to Fit

Duchenne muscular dystrophy. One dose, intended to last.

• rAAVrh74 to skeletal muscle
• Micro-dystrophin, not dystrophin
• Full-length gene is 11 kb — it does not fit

≈5 kb cassette

Calf pseudohypertrophy, from Duchenne’s own 1868 plates.

The courier is fixed. The cargo was redesigned.

Mendell et al. (2023)

This is the slide that sets up the whole middle of the talk, so do not rush it.

Dystrophin is 11 kb of coding sequence. AAV packages 4.7 kb. So the therapy is not “dystrophin gene therapy” — it is a deliberately truncated protein, missing most of the central rod domain, designed to retain selected actin-binding and dystrophin-associated complex-binding functions and fit in a vector. This micro-dystrophin omits the full-length C-terminal domain.

That is a physical constraint dictating a biological design. Flag it now; we come back to it with the cassette budget.

Pembrolizumab: The Odd One Out

Not a vector at all — a protein, made by cells.

• Humanised IgG4 against PD-1
• Mouse-derived binding loops on a human frame
• Manufactured from a cell line

≈1.4 kb of coding sequence

Jimmy CarterAnnounced metastatic melanoma in August 2015; clear scans by December

Melanoma in liver and brain, treated with surgery, radiosurgery and this antibody.

The product is a protein. The asset is the sequence.

Emancipator (2020); Goldberg et al. (2016)

Included deliberately as the odd one out. No virus, no delivery problem, no packaging limit — and still fundamentally a sequence-engineering success.

Humanisation is the detail worth a sentence: take murine CDR loops, graft them onto a human framework, then fix what breaks. That is exactly the “different, but not too different” design problem that comes back on the ITR slide at the end.

Carter is here because he made the whole thing public himself, in August 2015 and again that December — so it is a case you can tell without touching anyone’s private medical information. Metastatic melanoma with brain lesions had a median survival measured in months. He had the liver lesion resected, stereotactic radiosurgery to four brain lesions, and pembrolizumab; he stopped treatment in 2016 and lived to 100.

Keep the causal claim honest if asked: he had three treatments at once, so his own case does not isolate the drug. The trial evidence does — Goldberg and colleagues reported durable intracranial responses to pembrolizumab in melanoma brain metastases the following year.

Four Sequences, Drawn to Scale

Impact has nothing to do with length. The design problem is the same in every case.

Let the blocks fill before talking over them. Drawn to scale, one bead per hundred bases.

The visual argument: ChAdOx1 is twenty-five times the size of pembrolizumab’s coding sequence and neither fact predicts anything about clinical impact. What they share is that somebody chose the sequence.

And once a therapeutic is a sequence, choosing it is a search problem over a space you cannot enumerate. That is the whole talk.

The Virus Worth Borrowing

AAV is a defective parasite of other viruses. That defect is exactly why we use it.

Adeno-Associated Virus, Up Close

• 4.7 kb single-stranded DNA
• 60 subunits, T = 1
• ≈26 nm across
• No established human disease from natural AAV infection

Drag it.

Xie et al. (2002)

Live render of the deposited AAV2 structure — one asymmetric unit, sixty symmetry operators applied in the browser. Colour runs from capsid interior to exterior, which is the convention in the structural literature.

Turn it to a 3-fold axis: those spikes are the receptor-binding surface. Then a 5-fold: that channel is how the genome gets in and out.

Everything engineered in the back half of this talk happens on the outside of this object.

Where the Symmetry Is

Sixty identical subunits, related by the rotation group of an icosahedron.

• 5-fold ×6 — a channel at each
• 3-fold ×10 — the protrusions
• 2-fold ×15 — a depression at each

Every axis passes through the centre. Each order is a different local environment, and they behave differently.

Xie et al. (2002)

The axes here are not drawn by hand. The runtime reads the sixty symmetry operators out of the deposited assembly, takes the rotation angle of each one — 72° and 144° mean a 5-fold, 120° a 3-fold, 180° a 2-fold — and recovers the axis direction from the matrix. Six, ten and fifteen lines, exactly as the icosahedral group requires.

Why a virologist should care about each:

• 5-fold — a pore runs down it. The VP1 N-terminus is externalised through it, and the genome is threaded in through it during packaging.
• 3-fold — the protrusions. Receptor binding, and most neutralising antibody epitopes. This is where almost all engineering happens.
• 2-fold — a depression. Shallow, and generally less tolerant of insertion.

The shell is at 45% opacity so you can see the axes pass right through. Drag it: line up a 5-fold and you are looking straight down the pore.

Everything in 4.7 Kilobases

Two genes. Three promoters. Two more proteins hidden inside cap.

Samulski and Muzyczka (2014)

Orientation slide — do not read it out. Point at four things and move on: the ITRs at each end, rep, cap, and the three promoters named for where they sit in map units.

The next three slides take each piece in turn.

Four Rep Proteins We Throw Away

From p5 — Rep78, Rep68

Nick the terminal resolution site. Drive integration at AAVS1 on chromosome 19.

From p19 — Rep52, Rep40

Motor proteins. Thread the genome into a preformed capsid.

Each pair is one unspliced and one spliced product. All four are SF3 helicases, 3′→5′.

Delete rep and you delete site-specific integration.

Kotin et al. (1990)

The splicing detail is worth stating precisely because it is commonly garbled: p5 gives Rep78 unspliced and Rep68 spliced; p19 gives Rep52 and Rep40 the same way.

The line that matters for later: wild-type AAV has a mechanism for permanent residence, and recombinant vectors throw it away. That trade — insertional mutagenesis risk against episomal dilution — is why expression fades in dividing tissue.

Three Capsid Proteins, Unequal Shares

One ORF, three start sites.

VP1	87 kDa	PLA₂ + NLS
VP2	73 kDa	dispensable
VP3	62 kDa	the shell

Roughly 1 : 1 : 10 across the 60 subunits — about five, five and fifty.

VP1 is a minority subunit, and endosomal escape depends entirely on it.

Five copies out of sixty decide whether the particle ever gets out of the endosome.

Girod et al. (2002)

The ratio is ≈1:1:10, not 1:1:5 — you will see 1:1:5 misquoted, and it contradicts VP3 being about 85% of the capsid. It also drifts with serotype and with prep, which is why it is a release-critical attribute.

The VP1 point is the one to land. Its unique N-terminal region carries a phospholipase A2 domain that is buried in the intact particle and externalised under endosomal acidification. That is what breaches the membrane — and there are only about five copies per capsid.

Two Genes in the Same Sequence

AAP — assembly-activating protein

Non-canonical CTG start, nested inside cap. Scaffolds assembly in the nucleolus.

MAAP — membrane-associated accessory protein

A +1 frame inside the VP3 region. Promotes egress.

MAAP was not found by looking. It fell out of a comprehensive mutational scan in 2019.

A previously unannotated gene emerged from systematic mutational data.

Sonntag, Schmidt, and Kleinschmidt (2010); Ogden et al. (2019)

This slide is a deliberate plant for the ML section, and in a virology room it lands harder than any engineering result.

Ogden and colleagues did a comprehensive single-mutant scan of the AAV2 capsid and measured fitness by sequencing. The mutational data did not behave the way a single-reading-frame model predicted — and out of that fell a previously unannotated viral gene, in the +1 frame, thirty years after the genome was first sequenced.

The mutational scan revealed biology that targeted experiments had missed. Hold that thought until What Deep Mutational Scanning Measures.

Defective by Design

AAV cannot replicate alone. It needs a helper.

Adenovirus — E1A, E1B55K, E2A, E4orf6, VA RNA

HSV-1 — UL5, UL8, UL52, UL29, UL30

For a vector this is a feature:

• Helper functions supplied from a plasmid
• No helper virus needed in plasmid-based production
• A vector that cannot replicate cannot spread

Replication incompetence is not a limitation. It is the product.

Samulski and Muzyczka (2014)

The pivot of the section. Natural AAV’s dependence on a co-infecting helper looks like weakness; for a vector it is the most useful property it has, and it is most of the safety argument.

Note the name: Dependoparvovirus. The genus is named for the defect.

Why Not Adenovirus?

Good vaccine

• ≈8 kb payload
• Efficient, high expression
• Strong immune response

Poor gene therapy

• That same response clears transduced cells
• High seroprevalence
• Inflammatory at therapeutic doses

A vaccine wants an immune response. A gene therapy does not.

Dunbar et al. (2018)

One sentence explains the entire divergence in vector choice, and it is the one on the slide.

The honest caveat, because this room will think it: AAV’s low immunogenicity is relative and dose-dependent. At the systemic doses used for muscle and liver it is emphatically not silent. We spend a whole act on that.

Serotypes Are Surface Chemistry

Same fold. Same T = 1 shell. Nine variable loops on the outside.

• AAV2 → heparan sulfate
• AAV9 → terminal galactose
• VR-IV differs: AAV9’s spikes are blunter

Tropism lives on these loops.

Xie et al. (2002); DiMattia et al. (2012); Summerford and Samulski (1998); Shen et al. (2011)

Both rendered identically and at the same scale, so the comparison is fair.

There is no “AAV2 protein” and “AAV9 protein” in any deep sense. Same β-barrel fold, different surface loops — and those loops decide which glycan it binds, which tissue it reaches, and which antibodies recognise it.

That is why capsid engineering is tractable at all. You are not redesigning a fold, you are editing loops on a scaffold that tolerates it.

From Virus to Vector

Delete everything that makes it a virus. Keep 145 base pairs at each end.

Strip It Down to the ITRs

Gray et al. (2011); Zanta-Boussif et al. (2009)

Everything competes for the same 4.7 kb, and the regulatory elements are not free — CAG alone eats more than a third of the budget.

Two things this audience will appreciate. Element order is not cosmetic: WPRE has to sit in the 3′ UTR upstream of the polyA signal or it is simply not transcribed. And WPRE is not free either — it carries a truncated Woodchuck hepatitis X ORF, which is why the field moved to WPRE3 and mut6 variants.

The Budget Decides the Medicine

Payload	Size	Fits?
Factor IX	1.4 kb	easily
micro-dystrophin	3.8 kb	compact cassette only
BDD Factor VIII	4.4 kb	nothing left over
dystrophin	11 kb	never

This is why Duchenne therapy is micro-dystrophin, and why haemophilia A is harder than haemophilia B.

A physical limit on a capsid is dictating protein design.

Mendell et al. (2023)

Call back to Delandistrogene Moxeparvovec explicitly. The packaging limit is not an inconvenience, it is the reason an entire therapeutic class looks the way it does.

Factor IX is small and comfortable, which is why haemophilia B got there first. B-domain-deleted Factor VIII at 4.4 kb leaves essentially no room for a promoter, and that is why that programme has been so much harder.

What the Budget Produced

Product	Vector	Payload	Route
Zolgensma (2019)	AAV9	SMN1 0.9 kb	intravenous
Itvisma (2025)	AAV9	SMN1 0.9 kb	intrathecal
Hemgenix (2022)	AAV5	F9 Padua 1.4 kb	intravenous
Kebilidi (2024)	AAV2	DDC 1.4 kb	intraputaminal
Fayuvi (2026)	AAV9	SGSH 1.5 kb	intravenous
Luxturna (2017)	AAV2	RPE65 1.6 kb	subretinal
Elevidys (2023)	AAVrh74	micro-dystrophin 3.8 kb	intravenous
Otarmeni (2026)	AAV1 × 2	OTOF 6.0 kb	intracochlear

Ordered by payload. Every one but the last sits well inside the 4.7 kb budget.

OTOF is 6 kb. It does not fit — so otoferlin ships as two vectors, each carrying part of the coding sequence, reconstituted inside the hair cell.

The budget did not move. The architecture did.

U.S. Food and Drug Administration (2026a); U.S. Food and Drug Administration (2026b)

The list is short, and that is the point. Read down the payload column against the table you just showed: every approved product but one is a small gene, exactly as the budget predicts. Nobody has approved an 11 kb transgene, because nobody can.

Otarmeni is the slide’s payoff and it is only months old — approved April 2026 for OTOF-related deafness. Otoferlin is 1,997 residues, about 6 kb of coding sequence, so it cannot be packaged. The product is two AAV1 vectors carrying overlapping halves, and the full-length coding sequence is reconstituted in the transduced hair cell. That is the packaging limit being designed around rather than obeyed — and it costs you a second vector, a second dose of capsid antigen, and a dependence on both halves reaching the same cell.

Now read the route column, because it is the quieter lesson and it sets up the dose ceiling. Subretinal, intraputaminal and intracochlear are small, enclosed, immune-privileged compartments: the dose is tiny and it stays where you put it. The intravenous products are the ones carrying the safety burden. Elevidys has carried a boxed warning since November 2025 for serious liver injury and acute liver failure, after two deaths in non-ambulatory patients, and its indication is now restricted to ambulatory patients aged four and over.

Zolgensma and Itvisma are worth a sentence: same vector, same transgene, different route. Intrathecal delivery was approved in November 2025 precisely so that older, heavier patients could be dosed without the systemic load an intravenous dose would require. When you cannot lower the dose, you change where it goes.

If asked what is missing: three AAV products have left the market. Glybera, the first gene therapy approved in the West, was withdrawn in 2017. Beqvez was discontinued by Pfizer in February 2025, having treated no commercial patients. Roctavian — the BDD Factor VIII product from the previous slide — was voluntarily withdrawn in February 2026 after BioMarin could not find a buyer. None of those three failed because the biology failed. Approval is not the finish line, and this is a field where working and selling are different problems.

ssAAV or scAAV: Capacity Against Onset

McCarty, Monahan, and Samulski (2001)

The mechanism is one sentence. Rep resolves each ITR by nicking the terminal resolution site. Delete the trs from one ITR, Rep cannot resolve it, and the replication intermediate is packaged as a covalently linked dimer — a genome carrying its own complement.

It snap-anneals on release, skipping second-strand synthesis entirely. Weeks become days. The cost is exactly half the capacity.

Straight trade: capacity against onset.

Getting In

Eight steps between the syringe and a transcribing genome. Most of the dose does not survive them.

From the Needle to the Cell

Mingozzi and High (2013)

Everything on the next slide happens inside a cell. This one is about getting there, and it is where most of the dose is already gone.

Three things to say over it. Pre-existing neutralising antibodies are the big one — a large fraction of adults are seropositive for common serotypes, which is why trials screen for it and why you cannot redose. Liver and spleen take vector regardless of where you wanted it to go. And the endothelium is a real barrier: fenestrated in liver, continuous in muscle, tight junctions at the brain.

The honest framing for the band at the bottom: it is qualitative. The exact fractions depend on serotype, route, species and assay, and anyone who quotes you a single number for “percentage reaching target” is selling something.

So the dose has to be enormous. Hold that thought for the toxicity slides.

The Journey In, Step by Step

Girod et al. (2002); Cataldi and McCarty (2010); Nakai et al. (2001)

Once playing, the journey loops; with reduced motion it starts paused, so press Play. Select any numbered stage to pause there. The coloured particles represent four independently delivered vector genomes. The red branch shows loss to degradation, not a measured survival fraction.

The entry route is simplified. Endosomal and Golgi trafficking precede nuclear entry, and the site of membrane escape and the precise mechanism of nuclear uncoating remain incompletely resolved. Capsid opening is a teaching schematic.

The two-genome and four-genome views illustrate possible circular concatemers. They do not mean that the vector replicates, or that every dimer becomes a tetramer. Monomers, linear forms and other concatemer arrangements also occur. Host DNA repair contributes to joining, with pathway use depending on cell type and context.

Transcription requires an accessible duplex template, not completion of a fixed circularization sequence. One polymerase and one transcript stand in for many: the mRNA leaves the nucleus 5′ end first, and in the cytoplasm several ribosomes read it at once, each releasing a chain that folds into protein. Persistent episomes become chromatinized, omitted here for clarity.

Primary evidence for nuclear entry before uncoating: https://pubmed.ncbi.nlm.nih.gov/10684294/ Genome repair and concatemer formation: https://pmc.ncbi.nlm.nih.gov/articles/PMC2918998/

Two Bottlenecks Inside the Cell

Endosomal escape

Needs VP1u’s PLA₂ domain, externalised at low pH.

Five copies per capsid.

Second-strand synthesis

The packaged strand is transcriptionally inert until it is made double-stranded.

scAAV exists to skip this.

Potency also depends on targeting, nuclear entry, uncoating and expression.

Girod et al. (2002); McCarty, Monahan, and Samulski (2001)

Pulled out of the pipeline because these are the two places the money goes.

These are two useful examples, not a universal ranking. Which step limits transduction depends on the capsid and target cell. Improving productive delivery can reduce the dose — which is the argument the whole back half of the talk rests on.

What the Genome Becomes

• Second strand synthesised
• Host repair joins vector DNA ends
• Circular monomers and concatemers
• Chromatinised with host nucleosomes

The damage response is not evaded. MRN and ATM/ATR sense those hairpin ends.

Nakai et al. (2001)

Four looping phases: delivered single strand, second-strand synthesis, circularisation, chromatinised episome.

Two things the literature gets sloppy about in talks. NHEJ can contribute strongly to circularization, but pathway use depends on cell type. Homologous recombination can also contribute to genome joining. And concatemers are not invisible to the damage machinery; the MRN complex senses those hairpin ends. “Chromatinised and tolerated” is accurate. “Undetected” is not.

From Hairpin to Chromatin

Nakai et al. (2001)

The printable version of the animation before it, and the slide to use if the room wants the mechanism written down.

The sequence to narrate: the strand arrives inert, gets a second strand, the two ITR ends undergo repair-mediated joining. Circular monomers and concatemers can form. NHEJ and homologous repair contributions vary by context. Most persistent vector DNA is episomal, although low-frequency integration can occur. Then host nucleosomes assemble on it and it is transcribed like any other chromatin.

The payoff is the right-hand box. Transcribed but never replicated is the whole clinical story: wonderful in a neuron or a cardiomyocyte, a dilution problem in a liver that is still growing. And because the first dose raises antibodies, you do not get a second go.

Why Expression Fades

The episome is not replicated when the cell divides.

Post-mitotic — neurons, cardiomyocytes, muscle. Episomes can support durable expression, depending on the tissue and vector.

Dividing — paediatric liver is the clinical case. Hepatocytes expand out from under the vector.

This is the bill for deleting rep.

Wild-type AAV integrated at AAVS1. Recombinant vectors gave that up for safety, and took dilution in exchange.

Decades of expression is the hypothesis under test, not a result.

Nakai et al. (2001)

Avoid promising lifelong expression. Duration varies by product, tissue and patient, and long-term follow-up is essential.

The paediatric liver problem deserves a beat because it is where the biology bites a real programme: you cannot simply redose, because the antibody response to the first dose blocks the second.

That closes the biology arc. If you only get one shot, the dose has to be right — and the dose has to be manufactured.

Making It at Scale

A dose is 10¹³–10¹⁴ genomes per kilogram. For an adult that is a bioprocess, not a bench prep.

Two Platforms, Three Components

Urabe, Ding, and Kotin (2002)

Three components have to reach the cell: the ITR-flanked cassette, rep and cap, and helper functions. Mammalian systems transfect them; insect systems infect with baculovirus and need no adenoviral helper at all.

Be careful with the full:empty claim. You will hear that mammalian systems give cleaner preparations. It does not survive scrutiny — full:empty is dominated by process and cassette design, not by host.

HEK293 Already Has E1

The helper plasmid carries E2A, E4orf6 and VA RNA.

It does not carry E1.

HEK293 is adenovirus-E1-transformed — that is what makes it HEK293.

E1 functions come from the producer cell. The helper plasmid supplies the remaining adenoviral functions used in this platform.

Helper functions come from both the cell and the plasmid.

Urabe, Ding, and Kotin (2002)

A whole slide for one fact, because it is the detail that separates someone who has read the protocol from someone who has run it — and because the error is extremely common in review figures.

Adenovirus E1 transformed the original HEK cells in 1977. The line carries it constitutively. So the triple transfection supplies E2A, E4orf6 and VA RNA, and nothing else of adenoviral origin.

On the insect side, the pleasant surprise: baculovirus infection provides a permissive environment by itself, so no adenoviral helper is required at all.

Downstream: Capture, Polish, Concentrate

Mingozzi and High (2013)

The affinity step captures every intact capsid, full and empty alike, because the ligand binds the surface and knows nothing about the contents.

So separating full from empty is a separate polishing step, working on the small charge difference the packaged genome contributes. CsCl and iodixanol do the same job at bench scale and do not scale to commercial batches.

Empty Capsids Are Not Free

An empty capsid delivers nothing and still:

• presents capsid antigen
• consumes immune tolerance
• counts toward total particle dose

Which is why full:empty ratio is a release-critical attribute and not a cosmetic one.

Measured by AUC, mass photometry, charge-detection MS.

Antigen without benefit.

Mingozzi and High (2013)

This is the bridge into the immunology act, so make it explicitly.

The patient’s immune system responds to capsid protein. It cannot tell a full particle from an empty one. So every empty capsid in the vial spends part of a budget that is strictly limited, and buys nothing with it.

That means a manufacturing attribute is directly setting a clinical ceiling.

The Dose Ceiling

Efficacy rises with dose. So does toxicity. The gap between them is the whole problem.

The Therapeutic Window

Manno et al. (2006); Hinderer et al. (2018)

The hinge of the talk. Everything before is setup; everything after is a response to this slide.

Say the argument out loud: you cannot move the toxicity curve much, because it is driven by total capsid dose. You can move the efficacy curve, by building a capsid that does more per particle. That is why capsid engineering is a commercial priority and not an academic one.

Three Ways the Dose Fails

Before

Pre-existing neutralising antibodies exclude a large share of patients. Worst for AAV2.

At dose

Capsid-specific CD8 T cells clear transduced hepatocytes — expression rises, then falls.

Above dose

Complement, hepatotoxicity, thrombotic microangiopathy, dorsal root ganglion toxicity.

Unmethylated CpG can activate TLR9 and promote adaptive responses. These toxicities have multiple mechanisms.

Mingozzi and High (2013); Faust et al. (2013); Hinderer et al. (2018)

Manno 2006 is the foundational observation for the middle column: in the first haemophilia B trial, Factor IX expression rose and then fell, tracking a T cell response to capsid — not to the transgene. The vector worked; the immune system removed the cells that had taken it up.

The right-hand column is why high-dose systemic programmes have had serious adverse events, and it is what sets the practical ceiling today.

The CpG line is the thread into the ITR engineering slide at the end.

Engineering the Modules

Three parts of the cassette can be designed: the capsid, the promoter, the ITRs. Each is a different kind of search.

Three Modules, Three Searches

Capsid

Where it goes, and whether antibodies see it.

A protein-surface problem.

Promoter

Which cells express, and how much.

A regulatory-grammar problem.

ITR

Packaging, persistence, and innate sensing.

A structure-constrained problem.

Set the map before walking it. These are genuinely different problems and they need different methods — which is the reason the back of this talk is not just “apply a transformer”.

The capsid has the most data and the most commercial pull, so it gets the most slides.

The Search Space Is the Problem

VP1 is 735 residues.

A 60-residue window is 20⁶⁰ ≈ 1.15 × 10⁷⁸ — about the atom count of the observable universe. It takes 62 to exceed it.

But nobody designs the whole capsid.

A 7-mer insertion at VR-VIII is 20⁷ ≈ 1.3 × 10⁹ — and that already outruns any in vivo screen.

Romero and Arnold (2009)

State this carefully, because the usual version of the slide is wrong. 20⁶⁰ is about 1.15 × 10⁷⁸, which is smaller than 10⁸⁰. You need 62 residues to pass the atom count. Getting it right is more impressive than overstating it, and this room will check.

The real argument is the bang line. The space we actually search is tiny by comparison — seven residues — and it is still a billion variants against a screen that handles maybe 10⁶ to 10⁸ in a pooled selection, with a noisy readout.

Nine orange peaks are the serotypes we know. Everything else is unmeasured.

Capsid Shuffling

Fragment the cap genes of several serotypes, reassemble, select. Every chimera is a mosaic of whichever parents crossed over.

Grimm et al. (2008)

DNA shuffling, applied to capsids. Digest the cap genes from a handful of natural serotypes, let them reassemble on each other as templates, and you get chimeras carrying blocks from several parents at once.

Then put the library through selection in the tissue you care about and sequence the survivors.

The strength is that it explores combinations no single point-mutation scan would reach. The weakness is on the next slide but one.

What Shuffling Produced

Grimm et al. (2008)

Give shuffling its due. It produced genuinely useful reagents, AAV-DJ among them, and dismissing it would be wrong.

But note what you get at the end: a sequence that survived, and no function that maps sequence to fitness. Run the screen again in a different tissue and you start from scratch.

That is the gap the machine-learning work is trying to close.

Peptide Display: One Loop, a Billion Variants

The screen was right. The model organism was the problem.

Deverman et al. (2016)

This is the single best cautionary tale in the field and it belongs in a virology talk.

AAV-PHP.B was found by selecting for variants that transduced neurons in mice, recovered through a Cre-dependent scheme. It works beautifully — in C57BL/6. It does not transfer to non-human primates, because the mechanism runs through LY6A, a GPI-anchored mouse protein with no functional primate equivalent.

So directed evolution gave a phenotype in one context and no understanding that generalises. Which raises the obvious question: can you learn something that does?

What a Random Forest Actually Computes

Tabular features, interpretable splits, no gradients — the baseline a sequence model has to beat.

Breiman (2001)

Spend a moment here, because the rest of the act assumes the audience knows what “fitted a model” means, and for a room of virologists that is worth ninety seconds rather than being skipped.

Walk the animation. A variant arrives as a handful of engineered columns — charge, hydrophobicity, bulk, how buried residue 588 is. One tree asks a sequence of threshold questions and lands in a leaf, and the leaf is just the average fitness of the training variants that fell into it. No mechanism, no structure, no gradient: a lookup refined by questions.

Then the two sources of disagreement, which are the whole reason this works. Each tree is fitted to a different bootstrap resample of the data, and at each split it is only allowed to consider a random subset of the features. So the three trees on screen genuinely disagree about the same variant — watch the three ticks land in different places — and averaging them cancels a large part of the variance of any one.

Be clear that the numbers on screen are illustrative. The routing and the average are really computed from the features shown, but the features and leaf values are invented for the slide.

Why this model class is in a virology talk: when your data is a few dozen engineered columns, a forest is hard to beat and it tells you which columns mattered. When your data is the sequence itself, it is the wrong tool — which is the next slide, and the reason the cross-species work used sequence models instead.

From Mouse to Macaque

Fit4Function — fit one model per measurable property, on a library that uniformly samples manufacturable sequence space.

Six properties: production fitness, mouse liver transduction, and binding and transduction in each of two human liver cell lines.

Six models combined to design a liver-targeted, manufacturable library. 88% of variants met all six criteria.

No single predictor transfers. The six together do.

Eid et al. (2024)

This is the direct answer to the AAV-PHP.B problem, and the reason it works is the library design rather than the model architecture.

Instead of screening for one phenotype and keeping the winner, they screened a library that uniformly samples sequence space for many separate, individually measurable functions, and fitted a model to each. Then they combined the models to specify a capsid satisfying all six at once.

The chart is their own transferability table. Read it bottom-up. Chance is 0.24. The best single predictor — transduction of a human liver cell line in a dish — gets you to 0.86. All six together, 0.91. The argument is not that any one model is clever; it is that stacking independently measurable functions is what survives the species change.

Two details worth having ready. Production fitness alone scores 0.27 against a 0.24 baseline — manufacturability tells you essentially nothing about where a capsid goes, which is why it has to be a separate objective rather than a proxy. And dropping the mouse in vivo predictor leaves the figure unchanged in their table, at 0.911 either way: four human in vitro assays are already carrying the tropism signal between them.

If asked about the model: it is not a forest. Each property got a two-layer LSTM over a one-hot encoded 7-mer. The input here is the sequence itself, so a sequence model is the right tool — the contrast with the previous slide is deliberate.

Predicting Fitness Before You Measure It

Mikos, Chen & Suh (Rice) asked a narrower question: can you predict production fitness from structure alone?

Features describing each residue’s structural microenvironment — no measured AAV fitness used as training labels at all.

Reported threefold enrichment in mutations beating wild-type production, compared with random mutagenesis. Preprint, p = 7.5 × 10⁻¹².

Production fitness is unglamorous and gates everything — a capsid you cannot manufacture is not a capsid.

The columns you can compute before you run anything.

Mikos, Chen, and Suh (2021)

Worth including because it attacks a different and underrated axis. Everyone optimises tropism; far fewer optimise whether the thing can be made at scale, and a variant that packages badly is worthless however well it targets.

The interesting methodological claim is that structural features alone carry enough signal to triple your hit rate, without any measured fitness data to train on. That matters when you are engineering a serotype nobody has screened.

This is the tabular-feature regime from two slides back: a residue’s burial, packing and neighbourhood are columns you can compute from a deposited structure before you have run a single experiment. That is what makes a tree ensemble the natural choice and what makes the result possible at all with no AAV fitness labels.

If asked: a preprint from the Suh lab at Rice, and still a preprint. Do not describe the specific algorithm from the podium unless you have read the methods — say “structural features” and move on. Check for a published version before citing it as settled.

Discriminative or Generative

Ng and Jordan (2001)

Do not let this become a generic ML slide. The reason the distinction matters here is the specific shape of the data: a vast space, a tiny labelled fraction, and a strong prior available from evolution in the form of sequences that exist and work.

A discriminative model can only tell you about the neighbourhood you have labels for. A generative model learns the shape of the viable region itself.

What a VAE Actually Does

Encoder squeezes a sequence into a few numbers.

Decoder rebuilds a sequence from them.

Training forces the middle to be a smooth, well-populated cloud — so you can sample it, and interpolate across it.

Ng and Jordan (2001); Bryant et al. (2021)

Explain it in one breath: a VAE is an autoencoder with a penalty that forces the latent distribution towards a simple prior. Without that penalty you get a lookup table with holes in it. With it, the space between training points is meaningful.

That is the whole value here. Sequence space is discrete and astronomically large. A latent space is continuous and small, and sampled points can decode to plausible candidates — so you can wander it, interpolate between two capsids, or sample a thousand candidates that all look like things evolution could have made.

The strip underneath is the decoded sequence. Watch it change smoothly as the sampler moves. That smoothness is the property you are paying for.

The Generative Toolkit

VAE — smooth latent space, easy to sample and interpolate

GAN — sharp samples, harder to train, no likelihood

Diffusion — strong sample quality, slower to generate

Autoregressive / protein language models — trained on everything nature has made, then conditioned

Good at “does this look like a real protein”, which is exactly the prior we want.

The model class matters less than the data it is fitted to.

Bryant et al. (2021)

Keep this brisk — the audience is virologists, not an ML conference. The useful framing is what each class buys you rather than how it is trained.

The bang line is the honest summary. The reason the AAV results improved was not a switch from VAE to transformer, it was deep mutational scanning producing enough labelled data to fit anything at all.

What Deep Mutational Scanning Measures

Counts before and after selection reveal both enrichment and depletion.

Ogden et al. (2019)

Spell out the mechanic, because it is what changed the field and it is simpler than people expect.

You are not screening and picking. You are counting. Each variant carries its own sequence as its barcode, so you read the whole population before and after selection, and the log-ratio is the fitness. One experiment, the entire library, enrichment and depletion estimated for adequately sampled variants. Missing output counts can indicate depletion but are limited by sampling depth.

That last point is the one to press. Selection hands you survivors. A fitness landscape needs the failures too, because that is what tells the model where viability ends. This is precisely why the ML results arrived when DMS did, and not before.

Machine-Guided Capsid Design

Sequence → fitness, learned rather than assumed.

Ogden et al. (2019); Bryant et al. (2021)

This is the slide that earns the title of the talk, and it closes the loop with Two Genes in the Same Sequence.

The honest caveat if someone asks: viability and packaging are the easy labels. Tissue tropism and immune evasion in humans are much harder to measure at library scale, and that is where the field actually is.

Promoter Engineering: The Underrated Module

The objective: express only in the target tissue, in as few bases as possible.

Every base here is a base not spent on the transgene.

Build from transcription-factor binding motifs and chromatin accessibility data rather than borrowing a viral promoter.

Why it matters beyond size:

• off-target expression is a safety problem
• ubiquitous promoters invite transgene-directed immunity
• a tissue-restricted promoter can replace a more exotic capsid

Specificity bought with sequence instead of tropism.

Gray et al. (2011)

Promoters are the underrated module. The instinct is to solve targeting with the capsid, but a liver-restricted promoter achieves much of the same thing far more cheaply, and it composes with whatever capsid you already have.

The design problem is a regulatory-grammar one: which motifs, in what order, at what spacing, in what chromatin context. There is a lot of transferable work from the broader genomics literature here — this is the module where the ML is least AAV-specific.

Synthetic Promoters That Work

Michael RobertsSynpromics → AskBio → Bayer · now Chromatin Bioscience

Built promoters from screened motif libraries assembled into compact, tissue-selective elements — rather than repurposing CMV or a native regulatory region.

AMT-191 — AAV5 carrying GLA for Fabry disease, under a Synpromics liver-selective promoter.

Initial four-patient cohort: 27–208× mean normal α-Gal A activity. Sponsor report, July 2025 data cutoff.

A designed promoter, not a borrowed one.

uniQure (2025)

This is the concrete answer to “does promoter engineering actually deliver”, and it is worth having a number.

Synpromics’ approach was to assemble promoters from screened motif libraries rather than repurposing CMV or a native regulatory region — compact, liver-selective, and designed to be strong enough that you can drop the dose.

This is a dated, sponsor-reported biomarker result in four patients, not an estimate of clinical benefit or a controlled comparison of promoters. It does not isolate the contribution of the promoter from dose, capsid and other cassette elements. Do not interpret high enzyme activity alone as evidence of safety or durable benefit.

Why CpG Reads as Danger

The innate signal is what turns a delivery event into an immune response.

Faust et al. (2013); Mingozzi and High (2013)

This is the mechanistic bridge between the manufacturing/immunology act and the ITR design problem, and it is the same logic as the Karikó slide at the start.

Why CpG specifically: over evolutionary time, methylated cytosine deaminates to thymine, so vertebrate genomes have been depleted of CpG and the survivors are mostly methylated. Bacterial and viral genomes have not been through that filter. So the dinucleotide, in the wrong compartment and unmethylated, is a genuine discriminator of non-self.

The chain to emphasise: this is an innate signal that ends in an adaptive response against the capsid. The transgene may be perfectly tolerated — it is the packaging that gets the cells killed.

So reducing CpG is not cosmetic. It is removing the thing that starts the cascade.

ITR Engineering: Change the Sequence, Keep the Fold

The hardest module. The ITR must still fold into the hairpin Rep nicks.

And it is a dense patch of CpG — the motif TLR9 reads as “infection”.

145 bases, and every change is a large fraction of them.

Faust et al. (2013)

Pull the thread from the dose-ceiling act. Unmethylated CpG dinucleotides are a pathogen signature; TLR9 in plasmacytoid dendritic cells reads them, drives type I interferon, and that licenses the anti-capsid CD8 response that clears transduced cells.

Depleting CpG from the transgene is routine now. The ITR is harder for two reasons: it is only 145 bases so every change is a large fraction of it, and its function is its secondary structure — a palindrome that has to fold into a T-shaped hairpin with a terminal resolution site Rep can nick.

So this is a constrained design problem in the strict sense: minimise CpG count subject to a fold constraint and a protein-recognition constraint. Very little prior art, and a natural fit for generative models with hard constraints.

Where This Goes

The binding constraint is the dose ceiling.

A better capsid reaches the same effect at a lower dose — below the toxicity curve, below the T cell threshold, and cheaper to make.

One objective, three problems solved.

What would actually move the field:

• labels that mean human tropism
• models that transfer across serotypes
• immune evasion as a trainable objective
• a regulatory path for sequences with no natural counterpart

The biology set the constraints. What is left is a search problem.

Ogden et al. (2019); Eid et al. (2024)

Close on the argument, not a list of adjectives.

The shape of the whole talk: AAV’s defectiveness made it a good vector; being a good vector meant giving up integration; episomal persistence means the dose has to work the first time; the dose is bounded by immunology; and the only lever that moves is potency per particle.

Machine learning is in this talk because that is the lever, and the space is too large to search by hand.

The honest limitation to end on: the models are good at predicting whether a capsid assembles and packages. They are not yet good at predicting what it does in a human liver. Closing that gap is the actual frontier.

References

Breiman, Leo. 2001. “Random Forests.” Machine Learning 45 (1): 5–32. https://doi.org/10.1023/A:1010933404324.

Bryant, Drew H., Ali Bashir, Sam Sinai, Nina K. Jain, Pierce J. Ogden, Patrick F. Riley, George M. Church, Lucy J. Colwell, and Eric D. Kelsic. 2021. “Deep Diversification of an AAV Capsid Protein by Machine Learning.” Nature Biotechnology 39: 691–96. https://doi.org/10.1038/s41587-020-00793-4.

Cataldi, Marcela P., and Douglas M. McCarty. 2010. “Differential Effects of DNA Double-Strand Break Repair Pathways on Single-Strand and Self-Complementary Adeno-Associated Virus Vector Genomes.” Journal of Virology 84: 8673–82. https://doi.org/10.1128/JVI.00641-10.

Deverman, Benjamin E., Piers L. Pravdo, Bryan P. Simpson, Sripriya Ravindra Kumar, Ken Y. Chan, Abhik Banerjee, Wei-Li Wu, et al. 2016. “Cre-Dependent Selection Yields AAV Variants for Widespread Gene Transfer to the Adult Brain.” Nature Biotechnology 34: 204–9. https://doi.org/10.1038/nbt.3440.

DiMattia, Michael A., Hyun-Joo Nam, Kim Van Vliet, Matthew Mitchell, Antonette Bennett, Brittney L. Gurda, Robert McKenna, et al. 2012. “Structural Insight into the Unique Properties of Adeno-Associated Virus Serotype 9.” Journal of Virology 86 (12): 6947–58. https://doi.org/10.1128/JVI.07232-11.

Dunbar, Cynthia E., Katherine A. High, J. Keith Joung, Donald B. Kohn, Keiya Ozawa, and Michel Sadelain. 2018. “Gene Therapy Comes of Age.” Science 359 (6372): eaan4672. https://doi.org/10.1126/science.aan4672.

Eid, Fatma-Elzahraa, Albert T. Chen, Ken Y. Chan, Qin Huang, Qingxia Zheng, Isabelle G. Tobey, Simon Pacouret, et al. 2024. “Systematic Multi-Trait AAV Capsid Engineering for Efficient Gene Delivery.” Nature Communications 15: 6602. https://doi.org/10.1038/s41467-024-50555-y.

Emancipator, Kenneth. 2020. “Keytruda and PD-L1: A Real-World Example of Co-Development of a Drug with a Predictive Biomarker.” The AAPS Journal 22: 35. https://doi.org/10.1208/s12248-020-00525-1.

Faust, Susan M., Peter Bell, Brett J. Cutler, Scott N. Ashley, Yanqing Zhu, Joseph E. Rabinowitz, and James M. Wilson. 2013. “CpG-Depleted Adeno-Associated Virus Vectors Evade Immune Detection.” Journal of Clinical Investigation 123 (7): 2994–3001. https://doi.org/10.1172/JCI68205.

Girod, Anne, Christiane E. Wobus, Zoltán Zádori, Martin Ried, Katrin Leike, Peter Tijssen, Jürgen A. Kleinschmidt, and Michael Hallek. 2002. “The VP1 Capsid Protein of Adeno-Associated Virus Type 2 Is Carrying a Phospholipase A2 Domain Required for Virus Infectivity.” Journal of General Virology 83 (5): 973–78. https://doi.org/10.1099/0022-1317-83-5-973.

Goldberg, Sarah B., Scott N. Gettinger, Amit Mahajan, Anne C. Chiang, Roy S. Herbst, Mario Sznol, Apostolos J. Tsiouris, et al. 2016. “Pembrolizumab for Patients with Melanoma or Non-Small-Cell Lung Cancer and Untreated Brain Metastases: Early Analysis of a Non-Randomised, Open-Label, Phase 2 Trial.” The Lancet Oncology 17 (7): 976–83. https://doi.org/10.1016/S1470-2045(16)30053-5.

Gray, Steven J., Sandra B. Foti, Joshua W. Schwartz, Lavanya Bachaboina, Bonnie Taylor-Blake, Jennifer Coleman, Michael D. Ehlers, Mark J. Zylka, Thomas J. McCown, and R. Jude Samulski. 2011. “Optimizing Promoters for Recombinant Adeno-Associated Virus-Mediated Gene Expression in the Peripheral and Central Nervous System Using Self-Complementary Vectors.” Human Gene Therapy 22 (9): 1143–53. https://doi.org/10.1089/hum.2010.245.

Grimm, Dirk, Joyce S. Lee, Lora Wang, Tushar Desai, Bassel Akache, Theresa A. Storm, and Mark A. Kay. 2008. “In Vitro and in Vivo Gene Therapy Vector Evolution via Multispecies Interbreeding and Retargeting of Adeno-Associated Viruses.” Journal of Virology 82 (12): 5887–5911. https://doi.org/10.1128/JVI.00254-08.

Hinderer, Christian, Nathan Katz, Elizabeth L. Buza, Cecilia Dyer, Tamara Goode, Peter Bell, Laura K. Richman, and James M. Wilson. 2018. “Severe Toxicity in Nonhuman Primates and Piglets Following High-Dose Intravenous Administration of an Adeno-Associated Virus Vector Expressing Human SMN.” Human Gene Therapy 29 (3): 285–98. https://doi.org/10.1089/hum.2018.015.

Kotin, Robert M., Michele Siniscalco, R. Jude Samulski, Xiao Dong Zhu, Lawrence Hunter, Catherine A. Laughlin, Susan McLaughlin, Nicholas Muzyczka, Mariano Rocchi, and Kenneth I. Berns. 1990. “Site-Specific Integration by Adeno-Associated Virus.” Proceedings of the National Academy of Sciences 87 (6): 2211–15. https://doi.org/10.1073/pnas.87.6.2211.

Manno, Catherine S., Glenn F. Pierce, Valder R. Arruda, Bertil Glader, Margaret Ragni, John J. Rasko, Margareth C. Ozelo, et al. 2006. “Successful Transduction of Liver in Hemophilia by AAV-Factor IX and Limitations Imposed by the Host Immune Response.” Nature Medicine 12: 342–47. https://doi.org/10.1038/nm1358.

McCarty, Douglas M., Paul E. Monahan, and R. Jude Samulski. 2001. “Self-Complementary Recombinant Adeno-Associated Virus (scAAV) Vectors Promote Efficient Transduction Independently of DNA Synthesis.” Gene Therapy 8 (16): 1248–54. https://doi.org/10.1038/sj.gt.3301514.

Mendell, Jerry R., Perry B. Shieh, Craig M. McDonald, Zarife Sahenk, Kelly J. Lehman, Linda P. Lowes, et al. 2023. “Expression of SRP-9001 Dystrophin and Stabilization of Motor Function up to 2 Years Post-Treatment with Delandistrogene Moxeparvovec Gene Therapy in Individuals with Duchenne Muscular Dystrophy.” Frontiers in Cell and Developmental Biology 11: 1167762. https://doi.org/10.3389/fcell.2023.1167762.

Mikos, Georgios, Weitong Chen, and Junghae Suh. 2021. “Machine Learning Identification of Capsid Mutations to Improve AAV Production Fitness.” bioRxiv. https://doi.org/10.1101/2021.06.15.447941.

Mingozzi, Federico, and Katherine A. High. 2013. “Immune Responses to AAV Vectors: Overcoming Barriers to Successful Gene Therapy.” Blood 122 (1): 23–36. https://doi.org/10.1182/blood-2013-01-306647.

Nakai, Hiroyuki, Stephen R. Yant, Theresa A. Storm, Sally Fuess, Leonard Meuse, and Mark A. Kay. 2001. “Extrachromosomal Recombinant Adeno-Associated Virus Vector Genomes Are Primarily Responsible for Stable Liver Transduction in Vivo.” Journal of Virology 75 (15): 6969–76. https://doi.org/10.1128/JVI.75.15.6969-6976.2001.

Ng, Andrew Y., and Michael I. Jordan. 2001. “On Discriminative Vs. Generative Classifiers: A Comparison of Logistic Regression and Naive Bayes.” In Advances in Neural Information Processing Systems 14.

Ogden, Pierce J., Eric D. Kelsic, Sam Sinai, and George M. Church. 2019. “Comprehensive AAV Capsid Fitness Landscape Reveals a Viral Gene and Enables Machine-Guided Design.” Science 366 (6469): 1139–43. https://doi.org/10.1126/science.aaw2900.

Polack, Fernando P., Stephen J. Thomas, Nicholas Kitchin, Judith Absalon, Alejandra Gurtman, Stephen Lockhart, et al. 2020. “Safety and Efficacy of the BNT162b2 mRNA Covid-19 Vaccine.” New England Journal of Medicine 383: 2603–15. https://doi.org/10.1056/NEJMoa2034577.

Romero, Philip A., and Frances H. Arnold. 2009. “Exploring Protein Fitness Landscapes by Directed Evolution.” Nature Reviews Molecular Cell Biology 10: 866–76. https://doi.org/10.1038/nrm2805.

Samulski, R. Jude, and Nicholas Muzyczka. 2014. “AAV-Mediated Gene Therapy for Research and Therapeutic Purposes.” Annual Review of Virology 1: 427–51. https://doi.org/10.1146/annurev-virology-031413-085355.

Shen, Shen, Kelli D. Bryant, Sarah M. Brown, Scott H. Randell, and Aravind Asokan. 2011. “Terminal N-Linked Galactose Is the Primary Receptor for Adeno-Associated Virus 9.” Journal of Biological Chemistry 286 (15): 13532–40. https://doi.org/10.1074/jbc.M110.210922.

Sonntag, Florian, Kristin Schmidt, and Jürgen A. Kleinschmidt. 2010. “A Viral Assembly Factor Promotes AAV2 Capsid Formation in the Nucleolus.” Proceedings of the National Academy of Sciences 107 (22): 10220–25. https://doi.org/10.1073/pnas.1001673107.

Summerford, Candace, and R. Jude Samulski. 1998. “Membrane-Associated Heparan Sulfate Proteoglycan Is a Receptor for Adeno-Associated Virus Type 2 Virions.” Journal of Virology 72 (2): 1438–45. https://doi.org/10.1128/JVI.72.2.1438-1445.1998.

uniQure. 2025. “uniQure Announces Initial AMT-191 Phase I/IIa Data Showing Sustained Increases in Alpha-Gal a Enzyme Activity in Patients with Fabry Disease.” https://uniqure.gcs-web.com/news-releases/news-release-details/uniqure-announces-initial-amt-191-phase-iiia-data-showing.

Urabe, Masashi, Chunping Ding, and Robert M. Kotin. 2002. “Insect Cells as a Factory to Produce Adeno-Associated Virus Type 2 Vectors.” Human Gene Therapy 13 (16): 1935–43. https://doi.org/10.1089/10430340260355347.

U.S. Food and Drug Administration. 2026a. “Approved Cellular and Gene Therapy Products.” https://www.fda.gov/vaccines-blood-biologics/cellular-gene-therapy-products/approved-cellular-and-gene-therapy-products.

———. 2026b. “FDA Approves First-Ever Gene Therapy for Treatment of Genetic Hearing Loss Under National Priority Voucher Program.” https://www.fda.gov/news-events/press-announcements/fda-approves-first-ever-gene-therapy-treatment-genetic-hearing-loss-under-national-priority-voucher.

Voysey, Merryn, Sue Ann Costa Clemens, Shabir A. Madhi, Lily Y. Weckx, Pedro M. Folegatti, Parvinder K. Aley, et al. 2021. “Safety and Efficacy of the ChAdOx1 nCoV-19 Vaccine (AZD1222) Against SARS-CoV-2.” The Lancet 397 (10269): 99–111. https://doi.org/10.1016/S0140-6736(20)32661-1.

Xie, Qing, Wenman Bu, Shaswati Bhatia, Jordan Hare, Thayumanasamy Somasundaram, Arnold Azzi, and Michael S. Chapman. 2002. “The Atomic Structure of Adeno-Associated Virus (AAV-2), a Vector for Human Gene Therapy.” Proceedings of the National Academy of Sciences 99 (16): 10405–10. https://doi.org/10.1073/pnas.162250899.

Zanta-Boussif, M. A., S. Charrier, A. Brice-Ouzet, S. Martin, P. Opolon, H. Thérance, A. Thérance, and A. Galy. 2009. “Validation of a Mutated PRE Sequence Allowing High and Sustained Transgene Expression While Abrogating WHV-X Protein Synthesis: Application to the Gene Therapy of WAS.” Gene Therapy 16 (5): 605–19. https://doi.org/10.1038/gt.2009.3.

Credits

Structures — RCSB PDB. AAV2 1LP3, AAV9 3UX1. Icosahedral assemblies expanded in-browser from the deposited asymmetric units.

Software — three.js and NGL Viewer, both MIT, vendored into this repository.

Diagrams — drawn for this deck; MIT with the rest of the source.

Portraits — via Wikimedia Commons. CC BY-SA 4.0:

• Uğur Şahin, Özlem Türeci — © BioNTech SE
• Katalin Karikó — © University of Szeged

Public domain: Jimmy Carter — Lauren Gerson, LBJ Presidential Library, 2014.

Plate — Duchenne de Boulogne (1868), public domain.

No photograph of an identifiable patient appears here without their own disclosure. President Carter announced his diagnosis and his treatment publicly, himself. The Duchenne figure is an 1868 engraving, not a photograph of a living child.

Keep this slide. CC BY-SA 4.0 requires attribution and a licence statement, and the portraits of Şahin, Türeci and Karikó are all under it.

The last line is deliberate. Using a real, identifiable patient’s photograph — especially a child’s — in a conference talk needs explicit consent that a slide deck cannot evidence. The 1868 engraving carries the same clinical point with none of that problem.

Carter is the one exception and it is a principled one: he held press conferences about his own diagnosis, his own treatment and his own scans. That is disclosure by the patient, not disclosure about a patient.

Questions

Likely questions:

• Why not redose? The antibody response to dose one blocks it. Plasmapheresis and IgG proteases are being tried.
• Why not integrate deliberately? Insertional mutagenesis; and rAAV integration is largely non-targeted anyway. Nucleases and base editors are the other answer.
• How real is DRG toxicity? Consistent across NHP studies, dose dependent, often subclinical but histologically unambiguous.
• Does generative design beat directed evolution? On viability at distance, yes. On human tropism, not yet demonstrated.
• Why is MAAP interesting? It is the clearest case of ML on sequence data revealing biology rather than just optimising it.