Structure-Based Drug Design: From Protein Structure to Lead Compound

AlphaFold gave us the structure. Now what? With over 200 million predicted protein structures available, the bottleneck in drug discovery has shifted from "Can we get the structure?" to "Can we find druggable binding sites and design molecules that bind them?" Here's how structure-based drug design transforms a PDB file into a therapeutic candidate.

Key Takeaways

• Structural information has contributed to the discovery or optimization of a large fraction of modern small-molecule drugs (~50% often quoted), especially in kinases, proteases, and antivirals

• Main advantage: Rational design based on atomic-level interactions vs phenotypic screening

• Success rate: Multiple studies report 2-5× improvements in hit-to-lead efficiency when SBDD is layered on top of HTS, rather than relying on HTS alone

• Modern workflow: AlphaFold structure → Binding site prediction → Virtual screening → Lead optimization

• Critical gap: AlphaFold predicts apo (ligand-free) states; drugs often require holo (ligand-bound) conformations

What Is Structure-Based Drug Design?

Structure-based drug design (SBDD) uses the three-dimensional structure of a protein target to guide the discovery and optimization of drug molecules. Instead of randomly testing millions of compounds hoping one sticks, we use atomic-level knowledge of the binding site to design molecules that should bind.

The concept is elegantly simple: If you know the shape and chemistry of a protein's binding pocket, you can design a molecule that fits like a key in a lock.

SBDD vs Phenotypic Screening: When to Use Each

Structure-Based Drug Design (SBDD):

• Requires: High-resolution structure (X-ray, Cryo-EM, NMR, or AlphaFold)

• Best for: Known target with defined binding site (kinases, proteases, GPCRs)

• Advantages: Faster optimization, mechanistic understanding, selectivity tuning

• Limitations: Requires structural information, may miss allosteric mechanisms

Phenotypic Screening:

• Requires: Cell-based or organism-based assay

• Best for: Unknown mechanism, complex pathways, orphan indications

• Advantages: No target knowledge needed, captures entire biological system

• Limitations: Slower optimization, unknown mechanism, off-target effects

The Modern Approach: Hybrid workflows that combine both—use phenotypic screening to identify hits, then use SBDD to optimize them.

The SBDD Workflow: Five Critical Steps

Quick Overview:

1. Get structure (X-ray, Cryo-EM, AlphaFold)

2. Find druggable sites (FPocket, Orbion)

3. Virtual screening (dock millions of compounds)

4. Optimize hits (improve potency, selectivity, ADMET)

5. Validate experimentally (biochemical + cell assays)

Below is the detailed breakdown of each step:

Step 1: Obtain (or Predict) the Protein Structure

Traditional approach:

• X-ray crystallography (atomic resolution, but requires crystals)

• Cryo-EM (large complexes, membrane proteins, but lower resolution for small molecules)

• NMR (solution state, dynamics, but limited to <50 kDa proteins)

Modern approach: AlphaFold

• 200M+ predicted structures available

• Near-experimental accuracy for many targets

• Critical limitation: Predicts apo state (no ligand), not holo state (ligand-bound)

Why this matters: Proteins are dynamic. The pocket that binds a drug may not exist in the apo state—it might only open when the ligand approaches (cryptic pockets) or may adopt a different conformation when bound (induced fit).

The workaround:

• Use experimental holo structures from PDB as templates

• Run molecular dynamics (MD) simulations to sample conformations

• Use MSA subsampling in AlphaFold to generate alternative states

• Start with apo structure for initial docking, then refine with MD

Step 2: Identify Druggable Binding Sites

Not all pockets are equal. A "druggable" binding site has specific properties:

Size: Large enough for a drug-like molecule (~300-500 Da)

• Too small: Cannot fit a molecule

• Too large: Hard to achieve selectivity

Hydrophobicity: Balance of hydrophobic and polar regions

• All hydrophobic: Poor solubility

• All polar: Poor membrane permeability

Shape complementarity: Well-defined pocket with clear boundaries

• Deep pockets: Higher affinity (more contacts)

• Shallow grooves: Harder to target (protein-protein interfaces)

Accessibility: Can a drug molecule reach the pocket?

• Buried sites: Need to cross membranes or enter cells

• Surface sites: Easier access but often less specific

Hot Sites vs Cold Sites

Hot sites (high druggability):

• ATP-binding pockets in kinases

• Protease active sites (HIV protease, NS3/4A in Hepatitis C)

• GPCR orthosteric pockets

Cold sites (low druggability):

• Flat protein-protein interfaces (historically "undruggable")

• Large, solvent-exposed grooves

• Sites with no clear pocket

Tools for Binding Site Prediction

FPocket / DoGSiteScorer: Geometry-based pocket detection SiteMap (Schrödinger): Druggability scoring P2Rank: Machine learning-based site prediction AlphaFill: Predicts ligand positions based on homology Orbion: Combines structure analysis with evolutionary conservation to identify functional binding sites

Example: Identifying cryptic pockets Cryptic pockets don't appear in static structures—they only open transiently. These are increasingly important drug targets because they often provide selectivity (not all homologs have the same cryptic pocket).

Case: BCL-2 family inhibitors

• Static structure: No clear druggable pocket

• Molecular dynamics: Transient pocket opens near BH3 binding groove

• Result: Venetoclax (FDA-approved for CLL) targets this cryptic site

Step 3: Virtual Screening & Molecular Docking

Once you have a binding site, you need molecules that fit. Virtual screening tests millions of compounds computationally before synthesizing anything.

Molecular Docking Basics

The goal: Predict how a small molecule binds to the protein

• Pose prediction: Where does the ligand sit?

• Affinity prediction: How tightly does it bind?

Common docking programs:

• AutoDock Vina (free, widely used)

• Glide (Schrödinger) (commercial, high accuracy)

• GOLD (genetic algorithm-based)

• rDock (open-source, flexible)

Scoring functions: Docking programs estimate binding affinity using simplified energy functions:

• Force-field based: Sum of van der Waals, electrostatics, H-bonds

• Empirical: Trained on experimental binding data

• Machine learning: Neural networks trained on protein-ligand complexes

The accuracy problem: Docking scores correlate with binding affinity with R² ~0.5-0.7. They're good for ranking candidates but not precise affinity predictions.

Virtual Screening Strategies

1. High-Throughput Virtual Screening (HTVS)

• Screen 10⁶–10⁹ compounds from libraries (ZINC, PubChem, vendor catalogs)

• Fast docking (seconds per compound)

• Goal: Identify top 100-1,000 candidates for experimental testing

2. Fragment-Based Drug Discovery (FBDD)

• Screen small fragments (~150 Da) with weak binding (mM range)

• Crystallography or NMR to determine binding mode

• Link or grow fragments into larger, higher-affinity molecules

• Success: Vemurafenib (melanoma, targets BRAF V600E)

3. Structure-Based Virtual Screening (SBVS)

• Use known binders as starting point

• Similarity searching or scaffold hopping

• Optimize based on docking scores

Step 4: Lead Optimization

You've found a hit (IC₅₀ ~10-100 μM). Now make it better.

The Three Pillars of Lead Optimization

1. Potency (lower IC₅₀)

• Increase contacts with binding site

• Optimize H-bonds, hydrophobic packing

• Add functional groups that fit sub-pockets

2. Selectivity (hit your target, not others)

• Exploit differences between homologs

• Target unique residues in the binding site

• Check off-target binding (hERG for cardiotoxicity, CYP450s for metabolism)

3. ADMET properties (drug-like characteristics)

• Absorption: Oral bioavailability (Lipinski's Rule of Five)

• Distribution: Can it reach the target tissue?

• Metabolism: Is it stable or rapidly degraded?

• Excretion: Half-life, clearance

• Toxicity: No major liabilities

Computational Tools for Optimization

FEP (Free Energy Perturbation):

• Predicts ΔΔG of binding for mutations/modifications

• Accuracy: ~1 kcal/mol (good for ranking)

• Example: Schrodinger FEP+ improved kinase inhibitor affinity 100×

MMGBSA/MMPBSA:

• Post-docking energy refinement

• Accounts for solvation, entropy

• Faster than FEP, less accurate

AI/ML scoring:

• DeepDTA, KDEEP, OnionNet

• Trained on experimental binding data

• Emerging as more accurate than classical scoring

Step 5: Experimental Validation

Computational predictions must be tested experimentally.

Biochemical assays:

• IC₅₀, Ki, Kd measurements

• SPR (surface plasmon resonance), ITC (isothermal titration calorimetry)

• Crystallography or Cryo-EM to confirm binding mode

Cell-based assays:

• Does the compound work in cells?

• Off-target effects?

• Cytotoxicity?

Iterative cycle: Structure → Predict → Test → Solve structure of complex → Refine → Repeat

Case Study: HIV Protease Inhibitors

One of the greatest SBDD success stories.

The Target

HIV-1 protease: A homodimeric aspartic protease essential for viral replication. Without it, HIV produces non-infectious viral particles.

The Challenge

• Small, symmetric active site

• Highly flexible flaps cover the active site

• Drug resistance mutations emerge quickly

The Solution (1990s SBDD)

Step 1: Crystal structure of HIV protease solved (1989)

• Revealed symmetric active site with catalytic Asp25-Asp25' dyad

• Identified substrate binding cleft

Step 2: Design peptidomimetic inhibitors

• Mimic the natural peptide substrate

• Replace scissile bond with non-cleavable transition state analog

Step 3: Optimize for drug-like properties

• First-generation: Saquinavir (1995) - large, poor oral bioavailability

• Second-generation: Ritonavir, Indinavir - improved pharmacokinetics

• Third-generation: Darunavir (2006) - maintains activity against resistant mutants

Impact:

• 10+ FDA-approved HIV protease inhibitors

• Part of HAART (highly active antiretroviral therapy)

• Transformed HIV from death sentence to manageable chronic disease

Key lesson: Iterative structure-based optimization over 15+ years, combining X-ray crystallography of every new inhibitor with rational design, led to drugs that saved millions of lives.

Modern SBDD: The AlphaFold Era

AlphaFold changes the game by making structures accessible for previously "unstructurable" targets.

What AlphaFold Enables

1. Orphan targets Proteins with no experimental structure can now be modeled with high confidence.

Example: KRAS G12C (oncogenic mutant)

• Long considered "undruggable" (no deep pockets in static structure)

• Discovery: Fragment-based crystallography combined with medicinal chemistry revealed a cryptic pocket (switch II pocket) adjacent to Cys12

• AlphaFold now enables similar analyses for proteins without experimental structures

• Result: Sotorasib (FDA-approved 2021), Adagrasib (FDA-approved 2022)

2. Homology-based druggability assessment If a homolog is druggable, AlphaFold structure can reveal if your target has a similar pocket.

3. Allosteric site discovery Sites distant from the active site that regulate function. Often missed in traditional screens.

What AlphaFold Doesn't Give You

1. Holo (ligand-bound) conformations AlphaFold is trained on apo structures. Induced fit and cryptic pockets are missed.

Workaround: Use experimental holo structures of homologs as templates, or run MD simulations.

2. Binding site identification AlphaFold predicts structure, not function. You still need to identify which pocket to target.

Solution: Use binding site prediction tools (FPocket, Orbion) or evolutionary conservation analysis.

3. Protein dynamics Static structure doesn't capture breathing, domain motions, or transient states.

Solution: Molecular dynamics or enhanced sampling (metadynamics, accelerated MD).

Where Orbion Fits in the SBDD Workflow

While AlphaFold provides the structure, Orbion adds the biological and functional context needed for effective drug discovery:

1. Binding Site Prediction & Ranking

What Orbion does:

• Analyzes structure for potential binding sites

• Scores sites by druggability (size, shape, conservation)

• Identifies allosteric sites based on evolutionary conservation

Why it matters: Not all pockets are functional. Orbion filters out non-functional pockets and highlights those likely to regulate protein activity.

Example workflow:

1. Upload AlphaFold structure of kinase

2. Orbion identifies ATP pocket (expected) + allosteric site in hinge region (unexpected)

3. Virtual screening targets both sites

4. Allosteric site provides selectivity advantage

Orbion platform's binding site predictions on Adenosine Receptor A2a

2. PTM-Aware Drug Design

What Orbion does:

• Predicts post-translational modifications (phosphorylation, glycosylation, ubiquitination)

• Maps PTM sites onto structure

Why it matters: PTMs regulate protein activity and localization. A drug targeting the inactive state might fail if the cell modifies the protein to the active state.

Example: Kinases are activated by phosphorylation. If your AlphaFold model is unphosphorylated (inactive), but the cellular target is phosphorylated (active), your docking predictions will be wrong.

Orbion solution: Predicts phosphorylation sites, allows you to model both states.

Orbion platform's post-translational modification predictions on Adenosine Receptor A2a

3. Stability & Expression Optimization

What Orbion does:

• Suggests thermostabilizing mutations for construct design

• Recommends expression system (E. coli, yeast, insect, mammalian)

Why it matters: To validate computational hits experimentally, you need protein. If your target is unstable or doesn't express, you can't run biochemical assays.

Example: GPCR is computationally predicted to have a druggable allosteric site, but it aggregates during purification.

Orbion solution: Suggests stabilizing mutations (from construct design module), enabling expression of stable protein for screening.

Orbion platform's expression protocol suggestion on Adenosine Receptor A2a

The Economics of SBDD

Traditional Drug Discovery Timeline

• Target identification: 1-2 years

• Hit discovery (HTS): 1-2 years (test 10⁶ compounds)

• Lead optimization: 2-4 years

• Preclinical: 1-2 years

• Clinical trials: 5-10 years

• Total: 10-15 years, $1-2 billion

SBDD Impact

• Hit discovery: Reduced to 6-12 months (virtual screening before HTS)

• Lead optimization: Faster (structure-guided modifications)

• Attrition reduction: Better selectivity and ADMET early

ROI: SBDD reduces costs by 30-50% in early stages, increases probability of success by 2-3×.

Practical Workflow: From AlphaFold to Candidate

Here's a realistic, step-by-step workflow for a structural biologist or medicinal chemist:

Week 1-2: Structure Preparation

1. Obtain AlphaFold structure or retrieve from PDB

2. Prepare protein (add hydrogens, assign charges, check for errors)

3. Identify binding sites (Orbion, FPocket, or manual inspection)

4. Validate site by checking conservation and known ligands in homologs

Week 3-4: Virtual Screening Setup

1. Choose docking program (AutoDock Vina for free, Glide if you have license)

2. Define search space (binding site box)

3. Select compound library (ZINC, vendor catalogs, or fragment library)

4. Run test docking with known ligands (if available) to validate parameters

Week 5-8: High-Throughput Virtual Screening

1. Dock 10⁴–10⁶ compounds

2. Rank by docking score

3. Visual inspection of top 100-500 poses (filter out unrealistic binding modes)

4. Apply ADMET filters (Lipinski, PAINS filters, synthesizability)

5. Select 20-50 compounds for experimental testing

Week 9-12: Experimental Validation

1. Purchase or synthesize compounds

2. Biochemical assay (IC₅₀ determination)

3. Hit confirmation (repeat, dose-response curves)

4. Crystallography or Cryo-EM of top hits bound to protein

Month 4-12: Lead Optimization

1. Solve structure of hit-protein complex

2. Analyze binding mode (which residues interact? Can we improve?)

3. Design analogs (add H-bonds, fill sub-pockets, improve selectivity)

4. Iterative rounds of synthesis + testing + structure determination

5. Optimize ADMET properties

Month 12+: Preclinical Development

1. Scale-up synthesis

2. In vivo testing (animal models)

3. Toxicology studies

4. Formulation development

5. IND filing

Common Pitfalls in SBDD (And How to Avoid Them)

Pitfall 1: Over-Reliance on Docking Scores

Problem: Docking scores are noisy. A score of -10 kcal/mol doesn't guarantee binding.

Solution:

• Manually inspect poses (does the binding mode make chemical sense?)

• Use consensus scoring (multiple programs)

• Validate with experimental data (even a few known binders helps calibrate)

Pitfall 2: Ignoring Protein Flexibility

Problem: Proteins move. A rigid docking may miss the real binding mode.

Solution:

• Use flexible docking (allow side chains to move)

• Run molecular dynamics after docking (induced fit refinement)

• Ensemble docking (dock to multiple protein conformations)

Pitfall 3: Targeting the Wrong State

Problem: You optimize a drug for the inactive state, but the cell has the active state.

Solution:

• Understand the biological state (is the protein phosphorylated? Bound to a cofactor?)

• Use Orbion to predict PTMs and model both states

• Check if homologs have structures in active/inactive states

Pitfall 4: Neglecting Selectivity Early

Problem: You find a potent inhibitor, but it hits 50 other kinases.

Solution:

• Screen against homologs computationally (docking to off-targets)

• Exploit unique residues in your target's binding site

• Test selectivity panel early (before investing in optimization)

Pitfall 5: Poor ADMET Properties

Problem: Great affinity, but it doesn't cross membranes or is metabolized instantly.

Solution:

• Apply Lipinski's Rule of Five early (MW < 500, LogP < 5, H-bond donors < 5, acceptors < 10)

• Check for PAINS (pan-assay interference compounds)

• Predict metabolism (CYP450 interactions)

The Future of SBDD: AI-Native Drug Design

The next generation of SBDD doesn't just use AI for structure prediction—it uses AI for molecule generation.

Generative Models for Drug Design

Traditional: Dock existing compounds → Pick best

AI-Native: Generate novel compounds optimized for the binding site

Tools emerging:

• Insilico Medicine (Chemistry42): Generates molecules de novo for a given target

• Relay Therapeutics: Combines MD simulations + AI to find allosteric sites + design binders

• Exscientia: AI-designed drugs already in clinical trials (faster timelines)

How it works:

1. Train generative model (VAE, GAN, or Transformer) on millions of drug-like molecules

2. Condition generation on binding site structure

3. Model generates molecules predicted to bind

4. Filter for synthesizability and ADMET

5. Synthesize top candidates

Impact: First AI-designed drugs are entering clinics (e.g., DSP-1181 for OCD, designed by Exscientia in 12 months vs typical 4-5 years).

The Bottom Line

Structure-based drug design has matured from a niche technique to the dominant paradigm in drug discovery. With AlphaFold democratizing access to structures and AI tools accelerating virtual screening and lead optimization, the bottleneck is shifting from "Can we get the structure?" to "Can we interpret it correctly?"

The modern SBDD stack:

• Structure: AlphaFold (apo state) + experimental structures (holo state) + MD simulations (dynamics)

• Binding sites: Orbion (functional prediction) + FPocket (geometry) + evolutionary conservation

• Virtual screening: Glide/Vina (docking) + Machine learning scoring (refinement)

• Lead optimization: FEP (affinity prediction) + ADMET filters + iterative structure determination

• Validation: Biochemical assays → Cell assays → Structural confirmation → In vivo

The gap that AlphaFold left—PTMs, binding site identification, stability, druggability assessment—is where platforms like Orbion add critical value. Combining structural prediction with biological context is the key to turning a PDB file into a drug.