Drug discovery basics

Drug discovery basics

This page is optional background. It assumes you have heard of DNA, RNA and proteins but have not worked with them, and it explains just enough for the rest of RxCubed to make sense. Nothing here is specific to using the platform — the tutorial covers that.

✦ Need any of this explained further? Ask the AI assistant.

From gene to machine

Three things, in order:

• DNA is the stored instruction set. It sits in the cell and is copied faithfully, but it does not do any work itself. Think of it as the archive.
• RNA is the working copy. When a cell needs a particular protein, it transcribes that stretch of DNA into RNA and sends it off to be built. RNA is temporary by design.
• Protein is the machine. It is what actually does things: catalysing reactions, carrying signals, moving material across membranes, holding cells together.

The sentence that matters for drug discovery: disease usually shows up at the protein level. A gene may carry a mutation, but the reason someone is ill is that the protein built from it is the wrong shape, is made in the wrong amount, or is working when it should be idle. That is why most drugs act on proteins.

This is also why RxCubed's databases split the way they do. UniProt gives you a protein's sequence — the ordered list of parts. RCSB PDB and AlphaFold DB give you its shape. Both describe the same protein, and you usually need both.

Why shape is everything

A protein starts as a long chain of amino acids and folds up into one particular three-dimensional shape. That shape is not decoration — it is the function. Change the shape and you change what the protein does, which is exactly how a single mutation can cause disease.

Folded proteins have dents, grooves and channels on their surface. A pocket is one of those: a cavity the right size and chemistry for a small molecule to settle into. Drugs work by sitting in a pocket and staying there long enough to interfere.

The old description is "lock and key" — the drug is shaped to fit the pocket. It is a useful picture as long as you know where it breaks down: proteins are not rigid. They flex, breathe, and often change shape slightly when something binds. Two structures of the same protein can look meaningfully different.

In RxCubed this is what the 3D viewer, pocket detection and contact mapper are for: finding pockets, and seeing exactly which parts of the protein a bound molecule touches.

What a drug actually does

Almost everything reduces to two verbs.

Block. The drug occupies the site so the protein's normal partner cannot get in, or so the protein cannot do its job. Called an inhibitor for an enzyme, an antagonist for a receptor.

Activate. The drug binds and switches the protein on, imitating whatever normally would. Called an agonist.

The clearest illustration is two drugs on closely related targets:

• Propranolol blocks the beta-1 adrenergic receptor, stopping adrenaline from binding, which slows the heart.
• Salbutamol activates the beta-2 adrenergic receptor, imitating the natural signal, which relaxes airway muscle so someone can breathe during an asthma attack.

Same receptor family, opposite actions, completely different medicines.

Binding tightly is not enough. The hard part is selectivity — binding your intended target and not the several hundred human proteins that look broadly similar. A compound that sticks to everything is not a drug; it is a poison. Most compounds that fail, fail here or on safety, not on potency.

The four target families

Almost all small-molecule drugs act on one of four kinds of protein.

GPCRs (receptors). Sit in the cell membrane and pass signals from outside the cell to inside. The largest single family among approved drug targets. Example: propranolol, above.

Enzymes. Catalysts that speed up chemical reactions. Usually blocked by a molecule that occupies the site the real substrate would use. Example: atorvastatin blocks HMG-CoA reductase, slowing cholesterol production. Kinases — enzymes that attach phosphate groups to other proteins — are a very heavily targeted sub-family, especially in cancer.

Ion channels. Gated pores that let charged particles flow across a membrane. Example: lidocaine blocks sodium channels in pain-sensing nerves, so the pain signal never reaches the brain.

Transporters. Pumps that carry molecules across membranes. Example: fluoxetine jams the serotonin reuptake pump, leaving more serotonin in the synapse.

One sourced figure worth knowing: about 70% of small-molecule drugs act through just four "privileged" families — GPCRs, ion channels, kinases and nuclear receptors (Santos et al., Nature Reviews Drug Discovery, 2017, which mapped 1,578 FDA-approved drugs onto 893 targets). Precise per-family percentages vary a lot between sources depending on whether you count drugs or targets, so treat any exact breakdown you see with caution.

These families are not equally easy to work with computationally, which is worth knowing before you pick a target:

• Enzymes are the comfortable case. Many have well-solved structures with the natural substrate or an inhibitor already bound, which tells you exactly where the pocket is.
• GPCRs are the largest drug target family but among the hardest here. They sit in membranes, which makes them much harder to crystallise, and they change shape depending on what state they are in.
• Ion channels and transporters are also membrane proteins and share the same difficulty, often with large moving parts.

So the most-drugged family and the most computationally tractable family are not the same family. That is a real constraint, not a gap in the software.

Which diseases this kind of work suits

Beginners usually expect the answer to be a list of diseases. It is not — it is a property called druggability, and it cuts across disease areas. Three questions decide it:

1. Is there a specific protein that drives the disease?
2. Can we get its 3D structure, either solved experimentally or predicted?
3. Is there a pocket a small molecule or short peptide could bind?

Three yeses and this kind of work is productive. A no anywhere and it becomes very hard, however important the disease is.

Where it works well

Infectious disease, including antimicrobial resistance. Bacterial, viral and parasite proteins are often well characterised, structurally distinct from ours (which helps selectivity), and abundantly solved. HIV protease inhibitors and the SARS-CoV-2 main protease inhibitor in Paxlovid both came from exactly this approach.

Cancers driven by one identifiable mutated protein. Where a single altered protein drives the disease, it is a clean target. Imatinib for BCR-ABL in chronic myeloid leukaemia is the classic case; the KRAS G12C inhibitors are a recent one. Note the qualifier — cancers without a single driving protein are much harder.

Metabolic and enzyme disorders. An enzyme doing too much or too little is close to the ideal case. Statins on HMG-CoA reductase; acetazolamide on carbonic anhydrase — which is the example the tutorial works through end to end.

Rare and single-gene disorders. Often overlooked, and a genuinely good fit: when one gene causes the disease, the target is unusually well defined. There is rarely enough clinical data for large statistical models, but structural reasoning does not need a large dataset — it needs one structure.

Repurposing an existing drug. Asking whether an approved, already-safe drug might act on a different target is a structural question, and a cheap one. Its safety profile is known. ChEMBL is the natural starting point, since it holds what compounds have actually been measured doing.

Targets best hit with a peptide or a protein rather than a small molecule. A broad, flat interface is a poor fit for a small molecule but often a good one for a designed peptide, a nanobody or an antibody, which can cover area instead of filling a pocket. This is still structural work, and RxCubed's large-molecule tools cover it: predict a structure from sequence, dock two proteins together, map the contacts across the interface, estimate the binding affinity of the complex, design a sequence for a given backbone, and run molecular dynamics to see whether the result holds together. Worth knowing where this stops — the antibody-specific engineering that follows, such as humanisation and developability assessment, is not part of the platform.

Where it works poorly, or not at all

Complex, polygenic disease. Most psychiatric illness, most common hypertension, much of type 2 diabetes: hundreds of genes each contributing a little, and no single protein to aim at. Question 1 fails.

Targets with no pocket. Some proteins, especially many transcription factors, have flat surfaces or no fixed shape at all (intrinsically disordered). There is nowhere for a small molecule to grip. These are what people mean by "undruggable" — and it means undruggable by small molecules, not unimportant. A flat surface with nothing to grip is often exactly the case for a larger binder, which is the entry above.

Diseases treated with cell therapy, gene therapy or RNA drugs. These do not work by a molecule binding a protein surface at all, and their discovery turns on vector design, delivery and manufacturing rather than on structure. That is a different discipline, and the tools here do not apply to it.

Anything where the target is unknown. If nobody knows which protein is at fault, structural work has nothing to aim at yet. That is a biology problem, and it comes first.

The useful takeaway: when someone asks whether a platform like this suits their disease, the answer turns on the protein and its pocket, not on the name of the illness.

How potency is measured

You will meet these constantly, so it is worth five minutes.

Potency is reported as a concentration — how much of the compound you need for an effect. Common labels are IC50 (the concentration that inhibits half the activity), Ki and Kd (binding strength), and EC50 (the concentration producing half the maximum effect).

Because it is a concentration, smaller means stronger. A drug that works at 1 nanomolar is a thousand times more potent than one needing 1 micromolar. This catches people out constantly.

To avoid reading backwards, ChEMBL publishes pChEMBL, which flips it round so that bigger means stronger:

• pChEMBL 9 = 1 nM — very potent
• pChEMBL 7 = 100 nM — potent
• pChEMBL 5 = 10 µM — weak

It also puts different measurement types on one scale, so rows can be compared. Do still read the assay description: two numbers from very different experiments are not automatically comparable.

What the computer can and cannot tell you

This is the most important section on the page.

Computational results are hypotheses that help you decide what to test next. They are not evidence that a drug works. A docking score is a guess at how well a molecule might sit in a pocket, made by an approximate model of physics; a predicted structure is a guess at shape; a predicted property is a guess at behaviour.

Used properly, that is genuinely valuable — it turns "test ten thousand compounds" into "test fifty, in this order", which is an enormous saving. Used improperly, it produces confident-sounding numbers about a molecule that would do nothing whatsoever in a cell.

Two habits keep you honest:

• Prefer measured over predicted when both exist. An experimental structure beats a predicted one; a measured activity from ChEMBL beats a computed score.
• Say which is which in your write-up. RxCubed keeps every result linked to its inputs precisely so a reader can see where a number came from — see How RxCubed organises work.

Where RxCubed fits

Drug discovery runs roughly: find a target → find molecules that hit it (hits) → improve them (leads) → test in animals → test in people → approval. It usually takes over a decade, and the great majority of projects stop somewhere along the way.

RxCubed covers the first part: target to hit. Finding and reading the literature, collecting sequences and structures, looking for pockets, gathering candidate compounds, checking basic properties, and early docking and contact analysis — with the provenance to show how you got there.

It does not do lead optimisation, preclinical safety decisions, or anything clinical. If you find something interesting here, the next step is a laboratory.

Glossary

• Target — the protein a drug is meant to act on.
• Ligand — any molecule that binds a protein. A drug is a ligand chosen for a useful effect.
• Binding site / pocket — the place on the protein where a ligand binds.
• Affinity — how tightly a ligand binds. Reported as a concentration, so lower is tighter.
• Selectivity — binding your target and not others. Usually the harder problem.
• Hit — a compound showing real activity against the target. A starting point, not a drug.
• Lead — a hit improved enough to be worth serious investment.
• Assay — a laboratory experiment measuring what a compound does.
• In silico / in vitro / in vivo — on a computer / in a dish / in a living organism, in increasing order of how much the result means.
• Apo / holo — a structure solved without anything bound / with something bound. Holo structures show you the pocket in its occupied shape.
• Druggable — having a pocket a small molecule can bind well enough to matter.

What to read next

The tutorial puts most of this into practice on one real drug–protein pair. How RxCubed organises work explains evidence and provenance. The AI Assistant can explain any of the above in more depth as you go.