Module 2: Target Identification, Validation & Druggability
Every other module in this course assumes you already have a target. This one asks the question that comes first — and is the single highest-leverage decision in the entire pipeline: which protein should we drug at all?
Learning outcomes
- State the three independent conditions a genuine therapeutic target must satisfy.
- Assemble an evidence chain for target validation and rank the strength of its links.
- Compare geometric, energetic, and knowledge-based binding-site detection methods.
- Compute and interpret a SiteMap Dscore, and explain why PPI interfaces score badly.
- Explain how polypharmacology turns target prediction into a similarity search.
1. What Is a Therapeutic Target?
A therapeutic target is a biomolecule — almost always a protein — whose activity can be modulated by a drug to produce a desired clinical effect. To be a genuine target it must satisfy three independent conditions, and failing any one of them is fatal:
1. Disease-linked
Modulating it must actually change the disease. This is a question about biology, and it is where most drugs die.
2. Druggable
It must have a binding site that a drug-like molecule can actually occupy with high affinity. This is a question about structure — and it is what the rest of this module measures.
3. Safely modulable
Hitting it must not cause unacceptable harm — which depends on where else it is expressed and what else the drug touches (Module 12).
2. Why Target Selection Dominates Everything Downstream
The human genome encodes roughly 20,000 proteins. Estimates of the druggable fraction and approved-drug target counts depend on database date, definitions, and modality. A pocket score is not a universal census of what can be therapeutically modulated.
~20,000
protein-coding genes
the whole proteome
Context-dependent
therapeutic tractability
Depends on modality and evidence
Date-dependent
approved-drug target counts
Specify source, date, and definition
Lack of efficacy is a major cause of clinical attrition, but it does not uniquely diagnose a wrong target. Inadequate tissue exposure, insufficient functional modulation, patient selection, compensatory biology, and endpoint choice can all contribute. Target validation should test these alternatives.
3. Target Classes
Approved drugs cluster heavily into a handful of protein families, largely because those families happen to have well-formed binding pockets. Knowing the class tells you a great deal about which method to reach for.
| Class | Why it's druggable | Typical approach |
|---|---|---|
| Enzymes (esp. kinases, proteases) | Deep, enclosed catalytic/cofactor pockets evolved to bind small molecules | Structure-based docking; competitive inhibition |
| GPCRs | Orthosteric site buried in the 7-TM helical bundle | Historically ligand-based (few structures); now structure-based |
| Ion channels | Pore and gating sites; but selectivity across the family is hard | Often ligand-based; heavy off-target screening (see hERG, Module 12) |
| Nuclear receptors | Buried hydrophobic ligand-binding domain | Structure-based (see the estradiol pharmacophore, Module 7) |
| Transporters | Substrate sites, but large conformational cycles | Ensemble/state-specific modelling |
| PPIs & transcription factors | Usually not — flat, extended, solvent-exposed interfaces | Fragment-based, stabilizers/glues, or degradation (Module 6) |
4. Target Validation: Building an Evidence Chain
Validation asks: if I modulate this target, does the disease change? No single experiment settles it. Confidence comes from stacking independent, orthogonal lines of evidence — the more independent the sources, the lower the risk. Module 17 covers where that evidence comes from: omics data, interaction networks, pathway context, and perturbation screens, and how to weigh them without fooling yourself.
Human genetic evidence
The strongest signal available. GWAS hits, rare loss-of-function variants, and Mendelian disease genes tie the target to human biology directly — not to a mouse. Targets with supporting human genetic evidence succeed in the clinic roughly twice as often, which is why genetics-led target selection has reshaped the industry.
Genetic perturbation in models
Knockout/knockdown (CRISPR, RNAi) shows whether removing the protein produces the desired phenotype. Caveat: removing a protein is not the same as inhibiting it — a knockout also deletes scaffolding functions a drug would leave intact.
Chemical/pharmacological evidence
A selective tool compound reproduces the phenotype. This is the closest proxy for what a drug will actually do — but only if the tool is genuinely selective. A promiscuous 'tool' validates nothing.
Clinical/translational precedent
Human data: an existing drug on the same pathway, or a biomarker that tracks target engagement in patients.
5. Druggability: Can a Small Molecule Actually Bind?
Tractability is modality-dependent. Pocket descriptors estimate one aspect of small-molecule ligandability, not therapeutic feasibility. Antibodies, degraders, oligonucleotides, and other modalities have different requirements, including accessibility, tissue delivery, and mechanism.
The first step is finding the pocket at all. Tools fall into three families:
Geometric
Find concave cavities purely from shape, e.g. by rolling a probe sphere over the surface (fpocket, PockDrug). Fast and structure-only.
Energetic / probe-based
Computationally flood the surface with small organic probes and find where they cluster — "hot spots" (FTMap, SiteMap). Captures chemistry, not just shape.
Knowledge-based / ML
Learn what real ligand-binding sites look like from the PDB and predict new ones. Increasingly the default for large-scale proteome-wide assessment.
Once you have a pocket, you score it. The best-known scheme is SiteMap (Halgren, 2009), which reduces a pocket to three numbers and combines them linearly:
n is the number of site points, capped at 100 in the published function. e is enclosure, from 0 to 1. p is hydrophilic character. Size and enclosure raise the score; polarity lowers it. Halgren fit the three coefficients to a set of known sites, so the equation describes that fit. A flat, polar protein–protein interface scores badly for the same algebraic reason. A low Dscore does not, by itself, decide that the protein is undruggable: hot spots, pocket motion, and whether a small molecule can compete with the partner are outside the regression.
Interactive Playground: Druggability Scorecard
Load a pocket type and move one descriptor at a time. Start from the flat protein–protein interface and raise enclosure: SiteMap’s Dscore can cross its druggable cut-off from that change alone. Enclosure is one fitted descriptor. Hot spots, pocket adaptability, and competition with the partner protein are outside this equation.
Deep, enclosed, largely hydrophobic — the archetypal druggable pocket.
Capped at 100 by the scoring function.
0 = flat open surface, 1 = fully enclosed cavity.
Higher = more polar/solvent-friendly = worse for binding.
1.20
Dscore
This descriptor combination resembles higher-scoring sites in the fitted scheme; high affinity and therapeutic tractability remain untested.
≥ 0.80 — above this scheme’s reference threshold. Inspect local structure and ligand evidence; this is not proof of a binding site.
Because enclosure carries a +0.60 coefficient and hydrophilicity a −0.324 one, a pocket earns its druggability mainly by being buried and greasy. A PPI interface fails not because it is small — it is usually enormous — but because it is flat and wet. This is exactly why the field invented PROTACs and molecular glues (Modules 1 and 6): alternative modalities change the requirements, but PROTACs still need target engagement, productive ligase geometry, ubiquitination, and suitable cellular exposure. Antibodies and other biologics take the same escape route from the opposite direction — they bind flat, polar epitopes precisely because they present a large complementary surface rather than filling a cavity; Module 16 covers modelling those interfaces.
6. The Inverse Problem: Target Fishing & Polypharmacology
So far we have gone target → molecule. The reverse question is just as important: given a molecule, what does it hit? This is target prediction(or "target fishing"), and it matters for three reasons:
- Phenotypic hits need deconvolution. If a compound cures cells in a screen but you do not know its target, you cannot optimize it rationally.
- Off-targets cause toxicity. Predicting secondary targets early feeds directly into the safety work in Module 12.
- Polypharmacology is sometimes the point. Many effective drugs — notably kinase inhibitors and most CNS drugs — work because they hit several targets. Perfect selectivity is not always the goal.
The dominant computational approach rests on the similar property principle you will meet in Module 5: if your molecule is highly similar to a compound with a known target, it probably shares that target. This turns target prediction into a fingerprint similarity search against annotated bioactivity databases such as ChEMBL.
Knowledge check
Research decision checkpoint
A target has a high pocket score but weak functional perturbation evidence. Would you advance it?
Your output: Write a target evidence matrix with modulation direction, provenance, modality, and the next falsifying experiment.
Compare your reasoning
Keep pocket ligandability separate from disease causality. Seek selective perturbation and rescue in a relevant model; absent human genetics is not an automatic veto.