Skip to content

Computer-Aided Drug Design: SBDD, LBDD, QSAR, Pharmacophores and ADMET

Chemical databases, receptor-ligand interactions, docking and SBDD, LBDD, QSAR, pharmacophores, drug-likeness and ADMET prediction for CSIR NET LS-BT.

  • 21 min read
  • Updated
On this page · 14 sections
  1. Prerequisites
  2. The Hook: Reading a Pocket
  3. The Discovery Pipeline and Where Computation Sits
  4. Receptor–Ligand Interactions
  5. Structure-Based Drug Design (SBDD)
  6. Ligand-Based Drug Design (LBDD)
  7. Drug-Likeness and ADMET
  8. Memory Tricks
  9. Don’t Confuse
  10. Common Mistakes
  11. Practice
  12. Quick Revision Table
  13. Key Takeaways
  14. What’s Next

Prerequisites

The Hook: Reading a Pocket

How CSIR NET asks it. New topic in the revised LS-BT syllabus, so there are no past CSIR papers. Expect Part B on which approach fits which situation (structure known or not), what a pharmacophore or a QSAR descriptor is, and Lipinski’s rule. Expect Part C on \(\Delta G\) from \(K_d\), a Hansch-type prediction, a Tanimoto coefficient, an enrichment factor, or counting rule-of-five violations.


The Discovery Pipeline and Where Computation Sits

graph LR
    T["Target identification
and validation"] --> H["Hit finding
HTS, virtual screening,
fragments"] H --> L["Hit to lead"] L --> O["Lead optimisation
potency, selectivity,
ADMET"] O --> P["Preclinical
studies"] P --> C["Clinical trials
Phase I, II, III"] C --> R["Approval;
Phase IV
surveillance"]
StageWhat it decidesComputational contribution
Target identification and validationWhich protein to act onGenomics and comparative genomics (targets present in a pathogen, absent in the host), expression and network analysis, druggability assessment
Hit findingWhich molecules bind at allVirtual screening of compound libraries; pharmacophore and similarity searching
Hit to lead, lead optimisationWhich analogue is betterDocking and scoring, free-energy methods, QSAR, selectivity checks against related proteins
PreclinicalWhether it can be a drugADMET prediction, toxicity alerts

Terms worth separating: a hit is any molecule showing activity in an assay; a lead is a validated, tractable hit with a usable structure–activity relationship; a candidate is the optimised compound taken into preclinical and clinical work.

Chemical and bioactivity databases

ResourceContent
PubChem (NCBI)Open archive of chemical substances and their bioassay results: Substance (SID), Compound (CID) and BioAssay (AID) records, with computed properties and cross-links
NCI resourcesThe US National Cancer Institute’s open compound collections and screening data, including the NCI-60 panel of sixty human tumour cell lines used for growth-inhibition profiling, and the NCI Open Database of structures available for research
ChEMBL (EMBL-EBI)Manually curated bioactivities (\(IC_{50}\), \(K_i\), \(K_d\)) extracted from the medicinal-chemistry literature
DrugBankApproved and investigational drugs with targets, pharmacology and interactions
ZINCCommercially available compounds prepared in ready-to-dock 3D form
BindingDBMeasured binding affinities for protein–ligand pairs
PDBbindPDB complexes paired with their measured affinities; used to train and test scoring functions
Cambridge Structural Database (CSD)Small-molecule crystal structures: the source of standard bond lengths and geometries

Chemical structures are exchanged as SMILES (a line notation, for example ethanol CCO), InChI and its hashed InChIKey (a canonical identifier), MOL/SDF files (atoms, bonds, optional 3D coordinates) and SMARTS (a pattern language for substructure queries, the chemical analogue of a regular expression).


Receptor–Ligand Interactions

Binding is a reversible equilibrium, \(\mathrm{R} + \mathrm{L} \rightleftharpoons \mathrm{RL}\):

$$K_d = \frac{[\mathrm{R}][\mathrm{L}]}{[\mathrm{RL}]}, \qquad \Delta G^{\circ} = RT\ln K_d = -RT\ln K_a$$

A smaller \(K_d\) means tighter binding; \(K_d\) equals the free ligand concentration at which half the receptor is occupied. Since \(RT \approx 2.48\) kJ mol⁻¹ at 298 K, each tenfold gain in affinity is worth about 5.7 kJ mol⁻¹ (1.36 kcal mol⁻¹).

Binding free energy splits into enthalpy and entropy, \(\Delta G = \Delta H - T\Delta S\):

ContributionPhysical origin
Hydrogen bonds, ionic and dipolar contactsDirectional, enthalpic; buried charges must pay a desolvation cost
Hydrophobic effectBurying non-polar surface releases ordered water: largely entropic, and the main driver for most drugs
Van der Waals / shape complementarityMany weak dispersion contacts; demands a close fit
Conformational entropy lossFreezing rotatable bonds and side chains costs entropy, so rigidifying a ligand can improve affinity
Metal coordination, halogen bonds, π-stacking, cation–πSpecific contacts exploited in design

Binding models: the rigid lock and key, induced fit (the site reshapes around the ligand) and conformational selection (the ligand captures one pre-existing conformation). Real sites show all three, which is why docking to a single rigid structure can mislead.

Pharmacological vocabulary (Part B favourites): agonist (binds and activates), partial agonist (submaximal response), inverse agonist (reduces basal activity), antagonist (blocks without activating; competitive at the same site, non-competitive/allosteric elsewhere), allosteric modulator (positive or negative). Potency is reported as \(IC_{50}\) or \(EC_{50}\); for a competitive inhibitor the concentration-independent constant is \(K_i\), related by the Cheng–Prusoff relation \(K_i = IC_{50}/(1 + [S]/K_m)\). Selectivity is the ratio of affinities for the intended target and an off-target; ligand efficiency is binding energy per heavy atom, \(LE = -\Delta G/N_{\text{heavy}}\), which stops chemists from buying potency with sheer size.


Structure-Based Drug Design (SBDD)

SBDD needs a 3D structure of the target: from X-ray crystallography, NMR or cryo-EM, or failing that a homology model or a predicted structure (with the caveats on the modelling page).

graph TD
    A["Target structure
(PDB, model)"] --> B["Prepare: add H,
assign protonation,
remove waters, fix
missing loops"] B --> C["Identify and
characterise the
binding site"] C --> D["Dock a library:
poses + scores"] D --> E["Rescore, cluster,
inspect interactions"] E --> F["Synthesise and
assay the best"] F --> G["New complex structure
→ next design cycle"] G --> C

Molecular docking

Docking has two parts: a search over the ligand’s position, orientation and conformation in the site, and a scoring function that ranks the poses.

Search strategyExamples
Systematic / incremental constructionDOCK, FlexX
Stochastic: genetic algorithms, Monte CarloAutoDock, GOLD, Glide (hierarchical), AutoDock Vina
Scoring function typeBasisWeakness
Force-field basedMolecular-mechanics energies (LJ + Coulomb), sometimes with a solvation termPoor treatment of entropy and desolvation
EmpiricalWeighted sum of terms (H-bonds, hydrophobic contact, rotatable bonds) fitted to measured affinitiesOnly as good as the training set
Knowledge-basedStatistical potentials from the frequency of atom-pair contacts in known complexesBiased by what has been crystallised
Machine-learned / consensusModels trained on affinity data, or several functions combinedNeeds careful validation against decoys

Practical points exam questions test:

  • Docking usually keeps the receptor rigid and the ligand flexible; ensemble docking or induced-fit protocols relax this.
  • Scoring reproduces the pose far better than the affinity; rank ordering of different chemotypes is unreliable.
  • Redocking a ligand from a known complex and measuring the RMSD to its crystal pose (commonly judged successful below 2 Å) validates the protocol.
  • Water molecules in the site may be displaceable or essential; protonation states and tautomers change the result.
  • MM-PBSA/GBSA and free-energy perturbation or thermodynamic integration give better energetics for a few compounds, at much higher cost.

Virtual screening and its metrics

Virtual screening ranks a library in silico so only the top fraction is assayed. Structure-based screening docks the library; ligand-based screening ranks by similarity or pharmacophore match. Performance is measured against a set of known actives mixed with presumed inactives (decoys):

$$\text{EF}_{x\%} = \frac{(\text{actives found in top } x\%)/(\text{compounds in top } x\%)}{(\text{total actives})/(\text{total compounds})}$$

An enrichment factor of 10 means the selected fraction is ten times richer in actives than the whole library. ROC curves and AUC are also reported. Related ideas: fragment-based design (screen very small fragments with weak affinity by biophysical methods, then grow or link them, watching ligand efficiency), de novo design (build molecules inside the site, now often with generative models), and drug repurposing (screen approved drugs against a new target, so ADMET and safety are already known).


Ligand-Based Drug Design (LBDD)

Used when no structure of the target is available, but several active molecules are known. It rests on the similar property principle: structurally similar molecules tend to have similar activity.

Molecular descriptors and similarity

Descriptor classExamples
0D/1D (constitutional)Molecular weight, atom and ring counts, rotatable bonds, H-bond donors and acceptors
2D (topological)Connectivity indices, topological polar surface area (TPSA), fingerprints
Physicochemicallog P (octanol–water partition, lipophilicity), log D, \(pK_a\), solubility, molar refractivity
ElectronicHammett σ constants, partial charges, dipole moment, HOMO–LUMO energies
StericTaft steric parameter \(E_s\), molar volume, Verloop sterimol parameters
3DShape, surface area, CoMFA interaction fields, pharmacophore fingerprints

Fingerprints encode substructures as bit strings; similarity between two fingerprints is usually the Tanimoto (Jaccard) coefficient:

$$T = \frac{c}{a + b - c}$$

with \(a\) and \(b\) the bits set in each molecule and \(c\) the bits set in both. \(T = 1\) is identity; in practice \(T \ge 0.85\) on common fingerprints is treated as “quite similar”, but the threshold depends on the fingerprint.

Pharmacophores

A pharmacophore is the 3D arrangement of features required for activity, not a molecule and not a functional group: for example a hydrogen-bond donor, an aromatic ring centre and a positively ionisable centre at fixed distances. Features are typed as H-bond donor/acceptor, hydrophobic, aromatic, positive or negative ionisable.

  • Ligand-based derivation: align several actives (flexible alignment of conformers) and keep the features they share; programs include Catalyst/HipHop-type methods, PHASE and LigandScout.
  • Structure-based derivation: read the complementary features out of the binding site.
  • Use: 3D database searching for new scaffolds (scaffold hopping), and as a filter before docking.
  • Bioisosteric replacement swaps a group for another with similar properties to keep the pharmacophore while improving stability or permeability.

QSAR

Quantitative structure–activity relationship: a statistical model relating descriptors to activity, usually on a log scale.

$$\log\frac{1}{C} = a\,(\log P) + b\,\sigma + c\,E_s + d$$

The classical Hansch (linear free-energy) approach uses physicochemical constants (log P, Hammett σ, Taft \(E_s\)); the Free–Wilson approach assigns each substituent at each position an additive contribution. Activity often depends parabolically on lipophilicity (\(\log 1/C = -a(\log P)^2 + b\log P + c\)): too polar and the molecule cannot cross membranes, too greasy and it is trapped in them or insoluble, so there is an optimum log P.

3D-QSAR places aligned molecules on a grid and correlates steric and electrostatic interaction fields with activity using partial least squares: this is CoMFA; CoMSIA uses similarity-index fields instead. Output is a contour map showing where bulk or charge helps or hurts.

Validating a QSAR model (a favourite Part C theme):

StatisticMeaning
\(R^2\)Fit to the training set; always improves as descriptors are added, so it is not evidence of predictivity
\(q^2\) (\(Q^2\), cross-validated \(R^2\))Leave-one-out or leave-many-out internal prediction; \(q^2 > 0.5\) is a common minimum
\(R^2_{\text{pred}}\) on an external test setThe real test
y-randomisationShuffle activities and refit; a good model must then collapse
Applicability domainPredictions are trusted only for molecules resembling the training set

The rule of thumb is at least about five compounds per descriptor; otherwise chance correlation is likely.


Drug-Likeness and ADMET

A potent molecule is useless if it never reaches its target or is cleared or toxic. ADME(T) = absorption, distribution, metabolism, excretion and toxicity. Predicting these early is “fail fast, fail cheap”.

Drug-likeness filters

Lipinski’s rule of five (oral, passively absorbed drugs; derived by Lipinski and colleagues from compounds that had reached clinical trials): poor absorption is more likely when two or more of these hold.

PropertyLimit
Molecular weight> 500 Da
log P> 5
Hydrogen-bond donors (sum of OH and NH)> 5
Hydrogen-bond acceptors (sum of N and O)> 10

Related filters: Veber’s rules (rotatable bonds ≤ 10 and TPSA ≤ 140 Ų for good oral bioavailability); the rule of three for fragments (MW < 300, cLogP ≤ 3, donors ≤ 3, acceptors ≤ 3); PAINS filters that flag frequent-hitter substructures giving assay artefacts; lead-likeness limits that keep starting points small, since optimisation adds mass.

Known exceptions matter: natural products, antibiotics, peptides and actively transported compounds routinely break the rule of five, and the rule says nothing about potency or toxicity.

What ADMET prediction covers

PropertyWhy it mattersTypical computational handle
Aqueous solubilityDissolution limits absorptionlog S models from descriptors
PermeabilityCrossing gut epithelium; Caco-2 and PAMPA assays are the referencesTPSA, log P, H-bond counts
Oral bioavailability (F)Fraction of the dose reaching circulationRule-based plus trained models
Plasma protein bindingOnly free drug actsLipophilicity-based models
Blood–brain barrierNeeded for CNS targets, avoided otherwiseLow TPSA and moderate log P favour entry; efflux by P-glycoprotein opposes it
MetabolismCytochrome P450 (notably CYP3A4, CYP2D6) oxidation in phase I, conjugation in phase II; inhibition or induction causes drug–drug interactionsSite-of-metabolism and CYP inhibition classifiers
Excretion, half-lifeDosing interval; \(t_{1/2} = 0.693\,V_d/CL\)Clearance models
ToxicityAttrition and withdrawalStructural alerts, hERG channel blockade (cardiac QT risk), hepatotoxicity, mutagenicity (Ames) models

Tools commonly named: SwissADME, pkCSM, admetSAR, ADMET Predictor, Derek-style expert systems; toxicity models are often (Q)SAR classifiers, and read-across from similar compounds is used in regulatory settings.


Memory Tricks

Don’t Confuse

ThisvsThat
SBDD: target structure requiredvsLBDD: known actives, no structure
Docking search: finds posesvsScoring function: ranks them (the weaker half)
\(K_d\)/\(K_i\): equilibrium constantsvs\(IC_{50}\): assay-dependent (Cheng–Prusoff links them)
PotencyvsLigand efficiency: potency per heavy atom
Pharmacophore: 3D feature patternvsFunctional group: specific chemistry
CoMFA (3D-QSAR): field-based, needs alignmentvsHansch QSAR: physicochemical constants, no alignment
\(R^2\): fitvs\(q^2\)/external \(R^2_{pred}\): predictivity
Rule of five: absorption risk flagvsPotency or safety: not covered by it
Competitive antagonist: same sitevsAllosteric modulator: different site
HitvsLead: validated, optimisable, with an SAR

Common Mistakes


Practice

All questions below are original practice questions written in CSIR NET style. They are not previous-year questions; this unit is new in the revised syllabus.

Level 1: Foundation

  1. Which approach is used when the 3D structure of the target is unknown but several active ligands are known?
  2. What does a pharmacophore describe?
  3. Name the four properties in Lipinski’s rule of five.
Answers
  1. Ligand-based drug design (similarity and pharmacophore searching, QSAR).
  2. The 3D arrangement of molecular features needed for activity.
  3. Molecular weight, log P, hydrogen-bond donors, hydrogen-bond acceptors.

Level 2: Part B Style

Q1 (practice). A ligand binds its receptor with \(K_d = 1\) nM at 298 K (\(RT = 2.48\) kJ mol⁻¹). The standard binding free energy is closest to:

  • (a) −12 kJ mol⁻¹
  • (b) −51 kJ mol⁻¹
  • (c) +51 kJ mol⁻¹
  • (d) −114 kJ mol⁻¹
Answer and explanation

(b). \(\Delta G^{\circ} = RT\ln K_d = 2.48 \times \ln(10^{-9}) = 2.48 \times (-20.72) \approx -51.4\) kJ mol⁻¹. Sign check: binding is favourable, so \(\Delta G\) is negative; (c) has the wrong sign and (d) would correspond to a far smaller \(K_d\).

Q2 (practice). A compound has MW 540 Da, cLogP 5.8, 2 hydrogen-bond donors and 6 acceptors. Regarding Lipinski’s rule of five:

  • (a) it passes, because donors and acceptors are within limits
  • (b) it has one violation, which the rule tolerates
  • (c) it has two violations (MW and log P), flagging likely poor oral absorption
  • (d) the rule cannot be applied without the \(IC_{50}\)
Answer and explanation

(c). MW > 500 and log P > 5 are two violations; the rule predicts poor absorption when two or more limits are exceeded. The rule uses only computed properties, so (d) is wrong, and potency is irrelevant to it.

Q3 (practice). Two molecules are compared with the same 2D fingerprint. Molecule 1 has 60 bits set, molecule 2 has 80 bits set, and 40 bits are set in both. The Tanimoto coefficient is:

  • (a) 0.40
  • (b) 0.50
  • (c) 0.67
  • (d) 0.80
Answer and explanation

(a). \(T = c/(a + b - c) = 40/(60 + 80 - 40) = 40/100 = 0.40\). Using 40/100 is the union in the denominator; dividing by either molecule alone (0.67 or 0.50) gives the Tversky-style asymmetric measures, not Tanimoto.

Q4 (practice). Which statement about docking scoring functions is correct?

  • (a) Knowledge-based functions use statistical potentials derived from contacts observed in known complexes.
  • (b) Empirical functions are derived purely from quantum-mechanical calculations.
  • (c) Force-field scoring handles desolvation and entropy rigorously.
  • (d) Scoring functions predict absolute binding affinities to within 1 kJ mol⁻¹.
Answer and explanation

(a). Empirical functions are weighted sums fitted to measured affinities, not QM results; force-field scoring treats solvation and entropy crudely; and absolute affinity prediction remains the weakest part of docking, with errors of several kJ mol⁻¹ or worse.

Level 3: Part C Style

Q1 (practice). A library of 100,000 compounds contains 200 known actives. A docking run is used to select the top 1% for assay, and 60 of the actives appear in that selection.

(i) Calculate the hit rate in the selection, the hit rate in the whole library, and the enrichment factor at 1%. (ii) How many actives would be missed? (iii) Give two reasons a well-enriched screen can still produce no usable lead.

Answer and worked solution
  1. Top 1% = 1,000 compounds, of which 60 are active: hit rate 6%. Library hit rate = 200/100,000 = 0.2%. \(\text{EF}_{1\%} = 0.06/0.002 = \mathbf{30}\).
  2. 200 − 60 = 140 actives missed (70% of them): good enrichment is not good recall, which is why the cut-off depends on how much assay capacity exists.
  3. For example: (a) the actives recovered may be weak, non-selective, or chemically intractable (PAINS-like frequent hitters, poor solubility); (b) scoring may rank a chemotype well for the wrong reason (a wrong pose, a displaced water, a wrong protonation state), so assay confirmation fails; (c) the enrichment was measured against decoys that are too easy to distinguish, overstating performance.

Q2 (practice). A series of analogues gives the QSAR equation

$$\log(1/C) = -0.35\,(\log P)^2 + 1.82\,\log P + 0.60\,\sigma + 3.10$$

with n = 24, \(R^2 = 0.91\), \(q^2 = 0.78\).

(i) At what log P is activity predicted to be maximal? (ii) Predict \(\log(1/C)\) for an analogue with log P = 3.0 and σ = 0.23. (iii) What does the positive σ coefficient suggest? (iv) Comment on the statistics and on the limits of using this equation for a new analogue with log P = 7.

Answer and worked solution
  1. Differentiate: \(-0.70\log P + 1.82 = 0\), so log P = 2.6 gives maximum activity. The negative quadratic term is the usual parabolic lipophilicity dependence.
  2. \(-0.35 \times 9 + 1.82 \times 3 + 0.60 \times 0.23 + 3.10 = -3.15 + 5.46 + 0.138 + 3.10 = \mathbf{5.55}\).
  3. A positive coefficient on the Hammett constant means electron-withdrawing substituents (σ > 0) raise activity, hinting at an electronic requirement in the interaction.
  4. With 24 compounds and 3 descriptors the ratio is acceptable (about 8 per descriptor), and \(q^2 = 0.78\) indicates genuine internal predictivity rather than fit alone; an external test set and y-randomisation would still be needed. log P = 7 lies far outside the range the model was built on, so the prediction would be outside the applicability domain and the parabola would extrapolate an implausible collapse in activity.

Q3 (practice). A team has a validated bacterial enzyme target with a 2.0 Å crystal structure, and 15 known inhibitors from the literature spanning three scaffolds. Consider these plans:

(A) Dock a commercial library into the site after preparing the structure (hydrogens, protonation states, retained structural waters), then rescore and visually inspect the top poses. (B) Build a ligand-based pharmacophore from the 15 inhibitors and use it to search a 3D database for new scaffolds. (C) Accept the top-scoring docked compound as a lead without assay, since the scoring function is a free-energy calculation. (D) Check the selected compounds for rule-of-five violations, PAINS substructures and predicted hERG liability before buying them.

Which statements describe sound practice?

  • (a) A, B and D only
  • (b) A and C only
  • (c) B, C and D only
  • (d) A, B, C and D
Answer and worked solution

(a) A, B and D.

  • (A) Sound: structure preparation, rescoring and human inspection of interactions are standard, because scores alone are unreliable.
  • (B) Sound: with several actives in hand, a pharmacophore search complements docking and can find new scaffolds.
  • (C) Wrong: docking scores are not free energies, and no computational result substitutes for an assay. The target’s homologue in the host should also be checked for selectivity.
  • (D) Sound: early property and liability filters are cheaper than late failures.

Quick Revision Table

ItemFact
PipelineTarget → hit → lead → optimisation → preclinical → Phase I–III → approval
Chemical databasesPubChem (SID/CID/AID), NCI collections and the NCI-60 cell-line panel, ChEMBL, DrugBank, ZINC, BindingDB, PDBbind, CSD
Structure formatsSMILES, InChI/InChIKey, MOL/SDF, SMARTS (substructure patterns)
Affinity\(K_d\) small = tight; \(\Delta G^{\circ} = RT\ln K_d\); 10× affinity ≈ 5.7 kJ mol⁻¹
\(IC_{50}\) vs \(K_i\)Cheng–Prusoff: \(K_i = IC_{50}/(1 + [S]/K_m)\)
Binding forcesH-bonds and ionic (enthalpic, desolvation cost), hydrophobic effect (entropic, dominant), van der Waals, conformational entropy loss
Binding modelsLock and key, induced fit, conformational selection
Ligand efficiency\(LE = -\Delta G/N_{\text{heavy}}\)
SBDDNeeds target structure; prepare → site → dock → rescore → assay → new structure
DockingSearch (systematic, genetic, Monte Carlo) + scoring (force-field, empirical, knowledge-based, ML); receptor usually rigid; redocking RMSD < 2 Å validates
Better energeticsMM-PBSA/GBSA, FEP, thermodynamic integration
Virtual screening\(\text{EF}_{x\%}\); ROC/AUC; fragment-based, de novo, repurposing
LBDDSimilar property principle; descriptors 0D–3D; Tanimoto \(T = c/(a+b-c)\)
Pharmacophore3D features (donor, acceptor, hydrophobe, aromatic, ±ionisable); scaffold hopping; bioisosteres
QSARHansch (log P, σ, \(E_s\)), Free–Wilson (additive substituents); parabolic log P optimum
3D-QSARCoMFA (steric + electrostatic fields, PLS), CoMSIA
QSAR validation\(R^2\) vs \(q^2 > 0.5\) vs external \(R^2_{pred}\); y-randomisation; applicability domain; ~5 compounds per descriptor
Rule of fiveMW > 500, log P > 5, donors > 5, acceptors > 10; two or more = poor absorption likely
Other filtersVeber (rot. bonds ≤ 10, TPSA ≤ 140 Ų), rule of three (fragments), PAINS
ADMETSolubility, permeability (Caco-2, PAMPA), bioavailability, protein binding, BBB and P-gp, CYP450 metabolism and interactions, clearance (\(t_{1/2} = 0.693V_d/CL\)), toxicity (hERG, Ames, hepatotoxicity)

Key Takeaways

What’s Next