Targeting Macular Degeneration

The names come from what an ophthalmologist sees in the eye. In dry AMD the retina looks dry — just thinning and yellowish deposits. In wet AMD there's actual leaking fluid (and sometimes blood) pooling under and inside the retina. That fluid is what makes vision warp and disappear so fast. The technical name for the abnormal blood-vessel growth is choroidal neovascularization, abbreviated CNV.

VEGFA isn't a "bad" protein. It's essential. Embryos need it to build the circulatory system. Adults need it for wound healing (you grow new tiny vessels to feed a healing cut), for the menstrual cycle, for exercising muscle. The problem in wet AMD is too much of it in the wrong place. That's why anti-VEGF drugs are given as eye injections — you only block VEGFA right at the back of the eye, not everywhere in the body.

UniProt is the protein equivalent of Wikipedia for biologists — a free database with the full amino-acid sequence, structure, function, mutations, and references for every known protein. The entry for human VEGFA has the ID P15692. The protein is 395 amino acids long. That ID is the only thing you need to tell Boltz to pull the whole sequence — you'll see it happen in the walkthrough.

Before anti-VEGF drugs, most wet AMD patients went legally blind in the affected eye within two years. With regular anti-VEGF injections, the average patient now keeps their vision over the same period, and many actually gain back some lost vision. It's one of the clearest success stories in modern medicine. The catch: injections every 1–2 months for the rest of your life, and the drugs don't work for everyone.

Active areas of research include: (1) longer-lasting anti-VEGF formulations so patients need injections only twice a year (faricimab and brolucizumab are recent examples); (2) combination drugs that block VEGFA plus another growth factor (Ang-2) for harder cases; (3) gene therapy that turns the eye itself into a permanent anti-VEGF factory after a single injection; (4) oral small molecules that block VEGFA signaling — none approved yet for AMD, but the biggest prize, because pills are far easier than eye injections.

UniProt accessions are 6-character codes for most proteins (newer ones are 10). They're permanent and unique — P15692 will mean human VEGFA forever, even if the protein gets renamed or reclassified. Scientists cite accession numbers in papers exactly the way you'd cite a book's ISBN. This standardization is what lets databases and software all talk to each other.

Working backward from the disease. We knew: wet AMD = abnormal blood vessel growth in the retina. We asked: what biological signal drives new blood vessel growth? Decades of basic research had answered that — VEGFA. So VEGFA became the prime suspect. This is the standard pattern in drug discovery: identify the disease mechanism first, then identify the protein at the controlling node of that mechanism, then drug that protein. Picking the wrong target is the most common reason new drugs fail in clinical trials.

Each letter is a one-letter abbreviation for one amino acid. M = methionine (every protein starts with M — it's the "start" amino acid coded by the start codon AUG in your mRNA). T = threonine, D = aspartate, R = arginine, Q = glutamine, and so on. There are 20 standard amino acids and each has a single-letter code. The order of letters is what determines the protein's 3D shape, and the 3D shape determines what the protein does. A 395-residue protein has a one-in-20-to-the-395th-power chance of existing by chance — astronomically unlikely. Every detail of that sequence was shaped by evolution.

Proteins move. They often look slightly different when nothing is stuck to them ("unbound" or "apo" form) than when a drug or another protein is sitting in their pocket ("bound" or "holo" form). Boltz starts by predicting the unbound shape so you can see the natural resting state of VEGFA, including the pocket where a drug would eventually go. Later steps will show the protein with candidate drugs docked in.

VEGFA in real life doesn't float around as one chain — two copies of it lock together (one upside down relative to the other) to form a working "dimer." That's why the receptor on the blood-vessel cell binds two VEGFA chains at once. Boltz can model this if you tell it to add two copies of the sequence. For high-school purposes the single chain is fine, but it's a good reminder that real biology is usually messier than the simple picture.

Drug-discovery projects can take 10+ years and cost billions. Every step you take should narrow down the space of "what we still don't know." Writing a hypothesis up front ("a molecule that fits in the switch-II pocket of VEGFA should block its binding to VEGFR2") makes it possible months later to look back and ask: did we test what we thought we were testing? Even AI tools work better when you tell them a clear goal — "design molecules that bind here, with these constraints" gives much better candidates than "make me something cool."

Generative chemistry models are AI systems trained on millions of known molecules. They learn the grammar of chemistry — which atoms go next to which, which functional groups stabilize which others, what makes a molecule drug-like. Then, given a target protein's binding pocket, they propose new molecules that should fit that specific shape. Modern generative models can produce molecules nobody has ever seen before — but that obey all the rules of plausible chemistry. It's analogous to how language models generate sentences nobody has written, by learning the grammar of English.

SMILES stands for Simplified Molecular Input Line Entry System. It's a way of writing a molecule's structure as a string of characters. For example, caffeine is "CN1C=NC2=C1C(=O)N(C(=O)N2C)C" and aspirin is "CC(=O)OC1=CC=CC=C1C(=O)O." Every chemical database speaks SMILES. A CSV file with one SMILES per row is the universal way to hand a list of molecules to any computational chemistry tool. Boltz reads SMILES, builds the 3D shape of each molecule, and tries to dock it into your protein's pocket.

A single 50,000-molecule screen gives you maybe 100 plausible hits. But the top hit usually isn't drug-good enough on its own — it might bind, but be too greasy, too large, or have a flaw that makes it toxic. What real drug chemists do is take the best hit and vary it: swap one atom, add a ring, shrink a tail. Each variation gets re-scored. After 5-10 cycles of "best hit → 100 variations → new best hit," you've optimized into something genuinely drug-like. This is called lead optimization, and it's where most of the real intellectual work of medicinal chemistry happens. Boltz can do this loop in days; the old way took years.

About 90% of molecules that look great in computational screens fail in the lab — they don't actually bind, or they bind but can't get into cells, or they get into cells but get destroyed by liver enzymes within minutes, or they cause unrelated side effects. Drug discovery has gotten much better with AI but it's still fundamentally an empirical business. The goal isn't "find the perfect molecule on the computer" — it's "find the top 50 worth spending lab time on, instead of the top 50,000 you'd have to test blindly." That's the speedup.

Three reasons. First, the target was a small 147-residue protein with a fairly shallow pocket — challenging compared to a deep enzyme active site. Second, we ran Tiny (100 molecules); the model needs more samples to find genuinely good binders. Third, we used no expert constraints — no known binders as starting points, no specified pocket residues, no chemistry filters. A real project would seed the AI with a known active molecule (a "warhead") and ask it to vary that, which dramatically improves results. So 0.421 from a no-guidance Tiny run is actually a reasonable starting point — it proves the system works end-to-end and gives you a feel for what to optimize next.

Decoding the molecule: at the center is a pyridine ring (a benzene with one nitrogen) — a very common drug scaffold because pyridines often hydrogen-bond nicely with protein side chains. Hanging off it: a nitrile group (C≡N), another phenyl ring, an isobutyl group, and an amide linker (NH-C=O) connecting to a chlorophenyl group. The molecular weight (403 Da) is in the typical drug range (Lipinski's rule of 5 caps it at 500). The slight ⚠ on CLogP means it's a touch too oily — a real chemist would swap one of the phenyl rings for something more polar before going further.