Skip to content
LumenaryAcademy
Life & Earth 70 min

Medicinal chemistry

How a medicine finds its target and how new medicines are found — binding and affinity, the long road from idea to approval, structure–activity, ADME and drug-likeness, antibiotics and resistance, and the ethics of access.

Code tested · 6AI-draftedAwaiting review by a chemist

Your guides: Chloe · The Chemist, Margaret · The Analyst, Gideon · The Sceptic

MotivationH

What a tablet has to do

An aspirin tablet is a small disc of white powder. Within an hour of swallowing it, some of its molecules have dissolved in the stomach, crossed the wall of the gut, travelled in the blood, and reached an enzyme called cyclooxygenase (COX) in cells throughout the body. There each molecule transfers its acetyl group to one particular serine residue in the enzyme's channel, blocking it permanently, so that the enzyme can no longer make the prostaglandins that signal pain, fever and inflammation. John Vane showed in 1971 that this is how aspirin works, seventy years after chemists at Bayer first sold it, and more than two thousand years after people first chewed willow bark, which contains a related compound.

Every medicine has to solve the same problems. It must bind the right biological target tightly enough to change what it does; it must reach the target, which means surviving digestion, crossing membranes and not being cleared too soon; and it must not harm the patient more than it helps. Medicinal chemistry is the design and optimisation of molecules that do all three at once. It draws on everything in this course, especially Biochemistry, the equilibria of Chemical equilibrium and the handedness of Stereochemistry.

Note

This lesson explains the chemistry of well-known approved medicines and how new ones are discovered. It gives no medical or dosing advice: decisions about taking any medicine belong to patients with a doctor, pharmacist or other qualified health professional.

IntuitionH

Targets, binding and affinity

Most medicines act on proteins. About a third of approved medicines act on one family alone, the G-protein-coupled receptors that sit in cell membranes and relay signals such as adrenaline (Hauser and colleagues counted 34 % of FDA-approved drugs in 2017). Others target enzymes (aspirin and ibuprofen on COX; statins on HMG-CoA reductase, which makes cholesterol), ion channels, transporters, and occasionally DNA or RNA. A molecule that binds a receptor and activates it is an agonist; one that binds and blocks it is an antagonist; one that blocks an enzyme is an inhibitor, often a competitive one of the kind met in Biochemistry.

Binding is usually non-covalent and reversible. The medicine fits into a pocket on the protein and is held by many weak interactions at once: ionic attraction between opposite charges, hydrogen bonds, and the hydrophobic effect, which rewards burying oily surfaces out of water. A few medicines, aspirin and the penicillins among them, go further and form a covalent bond with their target.

H₃N⁺ONH₂CH₃Asp–COO⁻HO–SerLeu, Val side chainsionichydrogen bondhydrophobic contactbinding pocket of the target protein
A medicine (the ligand, centre) in the binding pocket of its target protein, drawn schematically. Each contact is weak on its own: an ionic bond between a protonated amine and a carboxylate side chain, a hydrogen bond from a serine to the ligand's carbonyl oxygen, and hydrophobic contact between a methyl group and oily side chains. Together they make the binding tight and selective.

Definition— dissociation constant and affinity

For a ligand L binding a protein P, P+L⇌PL\mathrm{P} + \mathrm{L} \rightleftharpoons \mathrm{PL}, the dissociation constant is Kd=[P][L]/[PL]K_\text{d} = [\mathrm{P}][\mathrm{L}]/[\mathrm{PL}], a concentration. A smaller KdK_\text{d} means tighter binding, or higher affinity; chemists often quote pKd=−log⁡10(Kd/M)\mathrm{p}K_\text{d} = -\log_{10}(K_\text{d}/\text{M}). The standard free energy of binding, relative to the 1 mol/L standard state, is ΔGbind∘=RTln⁡(Kd/c∘)\Delta G^\circ_\text{bind} = RT \ln (K_\text{d}/c^\circ).

Theorem— Fractional occupancy

If the free ligand concentration is [L][\mathrm{L}], the fraction of target molecules with ligand bound is

θ=[L]Kd+[L].\theta = \frac{[\mathrm{L}]}{K_\text{d} + [\mathrm{L}]}.

Proof

Write [PL]=[P][L]/Kd[\mathrm{PL}] = [\mathrm{P}][\mathrm{L}]/K_\text{d} from the definition, so θ=[PL]/([P]+[PL])=([L]/Kd)/(1+[L]/Kd)\theta = [\mathrm{PL}]/([\mathrm{P}] + [\mathrm{PL}]) = ([\mathrm{L}]/K_\text{d})/(1 + [\mathrm{L}]/K_\text{d}); multiply top and bottom by KdK_\text{d}.

∎

It is the same hyperbola as the Michaelis–Menten equation, for the same reason: a site that is either empty or full. In a laboratory assay, potency is usually measured as an IC50\text{IC}_{50}, the inhibitor concentration that halves an enzyme's activity. For a competitive inhibitor the IC50\text{IC}_{50} depends on how much substrate was used, so it is converted to the inhibition constant KiK_\text{i} (the inhibitor's dissociation constant) with the Cheng–Prusoff equation, Ki=IC50/(1+[S]/KM)K_\text{i} = \text{IC}_{50}/(1 + [\mathrm{S}]/K_\text{M}).

python
import numpy as np

R, T = 8.314, 310.15                 # J/(mol K); body temperature
c_std = 1.0                          # standard concentration, 1 mol/L

print(" Kd         pKd   ΔG°(binding) / kJ/mol")
for Kd in [1e-3, 1e-6, 1e-9, 1e-12]:
    dG = R * T * np.log(Kd / c_std) / 1000
    print(f"{Kd:8.0e}   {-np.log10(Kd):4.1f}   {dG:7.1f}")
print(f"each tenfold gain in affinity is worth {R * T * np.log(10) / 1000:.2f} kJ/mol")

# Fraction of receptors occupied by a ligand with Kd = 10 nM
Kd = 10.0                                            # nM
L = np.array([1, 3, 10, 30, 100, 300, 1000.0])       # free ligand, nM
theta = L / (L + Kd)
print("\n [L]/nM  occupancy")
for l, th in zip(L, theta):
    print(f"{l:7.0f}   {th:6.1%}  " + "#" * int(round(40 * th)))

# Cheng–Prusoff: an IC50 measured in an enzyme assay depends on the substrate
# concentration used; Ki does not.  Ki = IC50 / (1 + [S]/Km) for a competitive inhibitor.
Km = 25.0                                            # µM
for S, IC50 in [(5.0, 48.0), (25.0, 80.0), (100.0, 200.0)]:   # µM, nM
    print(f"[S] = {S:5.1f} µM, IC50 = {IC50:5.1f} nM  ->  Ki = {IC50 / (1 + S / Km):5.1f} nM")
Runs entirely in your browser, in an isolated worker. Nothing is sent to a server.

A typical modern medicine binds with KdK_\text{d} in the nanomolar range, a binding free energy of about −53 kJ/mol at body temperature. That is the sum of many small terms: once the cost of stripping water from both partners is paid, a hydrogen bond in water contributes only a few kJ/mol, so many contacts must cooperate. Every tenfold gain in affinity is worth only 5.94 kJ/mol. The occupancy table shows why concentration matters so much: 50 % occupancy at [L]=Kd[\mathrm{L}] = K_\text{d}, 91 % at ten times KdK_\text{d}, 99 % only at a hundred times. And the three assays, run at different substrate concentrations, give three different IC50\text{IC}_{50} values but the same KiK_\text{i} of 40 nM, which is the number to compare between laboratories.

Selectivity matters as much as affinity. Human cells contain hundreds of related kinases, proteases and receptors, and binding the wrong one causes side effects. The story of imatinib, approved in 2001 for chronic myeloid leukaemia, shows what selectivity can achieve. That cancer is driven by a single abnormal enzyme, the BCR-ABL kinase, made by a gene fusion; imatinib binds BCR-ABL's ATP pocket in an inactive shape that most other kinases rarely adopt, and it turned a fatal disease into a manageable one for most patients (Capdeville et al., 2002).

ExplanationH

From target to approval

A new medicine passes through a long sequence of stages, each designed to remove candidates that will not work or are not safe.

  1. Target identification and validation. Evidence from genetics, biology and patients that changing a particular protein's activity will help a disease. Support from human genetics, such as people with a natural loss-of-function variant who are healthy and protected, roughly doubles a target's later chance of success (Nelson et al., 2015).
  2. Hit finding. Searching for any molecule that binds: high-throughput screening of a million or more compounds, fragment-based screening of very small molecules by crystallography or NMR, virtual screening by computed docking, or DNA-encoded libraries of billions of molecules, each carrying a DNA barcode.
  3. Hit-to-lead and lead optimisation. Years of design–make–test cycles, the core of medicinal chemistry (next section), until one molecule has the potency, selectivity, ADME and safety to become a development candidate.
  4. Preclinical studies. Pharmacology and toxicology in cells and animals, and manufacturing to the regulators' quality standards, before an application to start human trials.
  5. Clinical trials. Phase I, usually a few dozen healthy volunteers (patients, for most cancer medicines), tests safety and how the body handles the medicine. Phase II, a few hundred patients, tests whether it works and at what exposure. Phase III, often thousands of patients in many countries, confirms benefit against the current standard treatment or a placebo, in randomised, controlled and usually double-blind trials.
  6. Regulatory review of all the evidence, and after approval, pharmacovigilance: monitoring for rare side effects that even large trials could not detect.

How often does it work? The most widely used numbers come from industry databases. The block below uses phase-transition rates published by the biotechnology trade association BIO with Informa and QLS, covering development programmes from 2011 to 2020 (rates of reaching US approval).

python
import numpy as np

# Phase-transition success rates and average phase durations for 2011-2020,
# from BIO, Informa Pharma Intelligence and QLS, "Clinical Development Success
# Rates 2011-2020" (2021). Rates are the fraction of programmes that advanced.
phases = ["Phase I", "Phase II", "Phase III", "Regulatory review"]
p_advance = np.array([0.520, 0.289, 0.578, 0.906])
years = np.array([2.3, 3.6, 3.3, 1.3])

remaining = np.cumprod(p_advance)
print(f"{'stage':18s} {'advance':>8s} {'still alive':>12s} {'years':>6s}")
for name, p, r, y in zip(phases, p_advance, remaining, years):
    print(f"{name:18s} {p:8.1%} {r:12.1%} {y:6.1f}")
loa = remaining[-1]
print(f"\nlikelihood of approval from Phase I: {loa:.1%}")
print(f"Phase I candidates per approved medicine: {1 / loa:.1f}")
print(f"average time from Phase I to approval: {years.sum():.1f} years")

# Where do the failures happen?  Out of 1000 programmes entering Phase I:
alive = 1000.0
for name, p in zip(phases, p_advance):
    print(f"  {name:18s} {alive:6.0f} enter, {alive * (1 - p):5.0f} stop")
    alive *= p
print(f"  approved: {alive:.0f}")
Runs entirely in your browser, in an isolated worker. Nothing is sent to a server.
enter Phase Ienter Phase IIenter Phase IIIsubmitted for approvalapproved10005201508779
Attrition in clinical development, from the phase-transition rates in the block above: of every 1000 programmes that enter Phase I, about 79 reach approval. The largest loss is in Phase II, where a medicine is first tested for whether it actually works in patients.

Only 7.9 % of candidates entering Phase I were approved, so roughly 13 enter human trials for each success, and the clinical stages alone took 10.5 years on average. Phase II is the great filter: just 28.9 % pass, because it is the first real test of whether the target hypothesis holds in patients. The same report found large differences between fields, from 5.3 % for cancer to 23.9 % for blood disorders. An independent MIT analysis of public trial records from 2000 to 2015 (Wong, Siah and Lo, 2019), counted differently, found a higher overall figure but only 3.4 % for cancer. Such numbers depend on definitions, so quote them with their source.

Worked examplesH

Structure–activity relationships

Lead optimisation rests on a simple, powerful idea: change one part of a molecule at a time, measure what changes, and learn which parts matter. The result, a table of structures against activities, is a structure–activity relationship (SAR). Medicinal chemists usually work with pIC50=−log⁡10(IC50/M)\text{pIC}_{50} = -\log_{10}(\text{IC}_{50}/\text{M}), because a logarithmic scale is proportional to free energy and differences add.

Example— Reading an SAR table

A team optimising an inhibitor of an enzyme makes four related compounds from a common core, varying two positions. (The data are invented for teaching.)

CompoundPosition APosition BIC₅₀ (nM)pIC₅₀
1HH20005.7
2ClH2006.7
3HOCH₃6306.2
4ClOCH₃637.2

The chlorine at A adds 1.0 log unit whether or not the methoxy group is present, and the methoxy group at B adds 0.5 either way: the effects are additive, which suggests the two positions bind independent parts of the pocket. Compound 4 is 32 times more potent than compound 1. Additivity often fails, and when it does that is informative too: two groups that each help alone but not together are probably competing for the same space. Analysing a large table this way, by fitting a contribution to each substituent, is the Free–Wilson method, an early form of the machine learning now common in drug discovery.

Potency is never the only goal. Adding greasy groups usually raises potency, because the hydrophobic effect rewards burying them, but it also lowers solubility, raises binding to the wrong proteins and speeds metabolism. Teams therefore track ligand efficiency, the binding free energy per heavy (non-hydrogen) atom, so that each added atom must earn its place. In imatinib's development, one added methyl group improved selectivity for the target kinase, and a solubilising group made the molecule absorbable by mouth.

Two other common moves are bioisosteric replacement, swapping a group for another with similar shape and charge but better properties (a tetrazole ring for a carboxylic acid in several blood-pressure medicines), and conformational restriction, locking a flexible molecule into its binding shape so that less entropy is lost on binding (Entropy and free energy). The 3D structure of the target bound to the ligand, from crystallography or cryo-EM, now guides most of these choices, and since 2024 predicted complexes from tools like AlphaFold 3 are used alongside them (Biochemistry).

ComputationC

ADME and drug-likeness

A molecule that binds beautifully in a test tube can still fail as a medicine. ADME names the four things the body does to it: absorption (crossing from the gut into the blood), distribution (where it goes: tissues, the brain, bound to plasma proteins), metabolism (mostly in the liver, where cytochrome P450 enzymes oxidise foreign molecules to make them easier to excrete) and excretion (mainly by the kidneys). Common genetic variants in P450 enzymes make some people fast and others slow metabolisers of particular medicines.

In 1997 Christopher Lipinski and colleagues at Pfizer looked at compounds that had reached Phase II and asked what the orally absorbed ones had in common. Their summary, the rule of five because every limit is a multiple of five, says that poor absorption is more likely when a molecule has more than one of these features: molar mass over 500 g/mol, a calculated octanol–water partition coefficient (log P) over 5, more than 5 hydrogen-bond donors, or more than 10 acceptors. Daniel Veber's group added in 2002 that molecules with a polar surface area up to 140 Ų and at most 10 rotatable bonds are more often absorbed in rats. The block applies both to nine approved medicines, using properties computed by PubChem.

python
# Computed properties from PubChem (https://pubchem.ncbi.nlm.nih.gov/), by compound ID:
# molar mass (g/mol), XLogP3, H-bond donors, H-bond acceptors, polar surface area (Ų),
# rotatable bonds.  Values as listed in PubChem in 2026.
drugs = {
    #  name           CID       M      logP  HBD HBA  TPSA  RotB
    "aspirin":      (2244,    180.16,  1.2,  1,  4,  63.6,  3),
    "paracetamol":  (1983,    151.16,  0.5,  2,  2,  49.3,  1),
    "ibuprofen":    (3672,    206.28,  3.5,  1,  2,  37.3,  4),
    "metformin":    (4091,    129.16, -1.3,  3,  1,  91.5,  2),
    "amoxicillin":  (33613,   365.4,  -2.0,  4,  7, 158.0,  4),
    "imatinib":     (5291,    493.6,   3.5,  2,  7,  86.3,  7),
    "atorvastatin": (60823,   558.6,   5.0,  4,  6, 112.0, 12),
    "azithromycin": (447043,  749.0,   4.0,  5, 14, 180.0,  7),
    "ciclosporin":  (5284373, 1202.6,  7.5,  5, 12, 279.0, 15),
}

def lipinski_violations(M, logP, hbd, hba):
    """Lipinski's rule of five: M <= 500, logP <= 5, donors <= 5, acceptors <= 10."""
    return (M > 500) + (logP > 5) + (hbd > 5) + (hba > 10)

def veber_ok(tpsa, rotb):
    """Veber et al. (2002): polar surface area <= 140 Ų and <= 10 rotatable bonds."""
    return tpsa <= 140 and rotb <= 10

print(f"{'medicine':13s} {'M':>7s} {'logP':>5s} {'HBD':>4s} {'HBA':>4s} {'TPSA':>6s} {'RotB':>5s}  Ro5 viol.  Veber")
for name, (cid, M, logP, hbd, hba, tpsa, rotb) in drugs.items():
    v = lipinski_violations(M, logP, hbd, hba)
    print(f"{name:13s} {M:7.1f} {logP:5.1f} {hbd:4d} {hba:4d} {tpsa:6.1f} {rotb:5d}  {v:5d}      "
          f"{'yes' if veber_ok(tpsa, rotb) else 'no'}")

# Charge matters too: the neutral form crosses membranes far more easily (Henderson-Hasselbalch)
def fraction_neutral(pKa, pH, acid):
    return 1 / (1 + 10 ** ((pH - pKa) if acid else (pKa - pH)))

print(f"\n{'fraction neutral':27s} {'stomach 1.5':>12s} {'intestine 6.5':>14s} {'blood 7.4':>10s}")
for label, pKa, acid in [("aspirin (acid, pKa 3.5)", 3.5, True), ("metformin (base, pKa 12.4)", 12.4, False)]:
    f = [fraction_neutral(pKa, pH, acid) for pH in (1.5, 6.5, 7.4)]
    print(f"{label:27s} {f[0]:12.2g} {f[1]:14.2g} {f[2]:10.2g}")
Runs entirely in your browser, in an isolated worker. Nothing is sent to a server.

The small classics, aspirin, paracetamol, ibuprofen and metformin, pass both tests easily. Imatinib, at 493.6 g/mol, just squeezes under the mass limit. Atorvastatin has one violation, which the rule still accepts, though its twelve rotatable bonds fail Veber's test. Then come the instructive failures: amoxicillin passes Lipinski but its polar surface area of 158 Ų fails Veber, yet it is well absorbed, because a transporter in the gut wall that normally takes up small peptides carries it across. Azithromycin (two violations) and ciclosporin (three) are large natural-product-derived molecules taken by mouth anyway.

The last two lines add charge, from Acids, bases and buffers: an ion crosses the oily interior of a membrane far less easily than a neutral molecule. Aspirin, an acid with pKa\mathrm{p}K_\text{a} 3.5, is 99 per cent neutral in the stomach but only 0.1 per cent in the small intestine; metformin, a strong base, is essentially never neutral, and it too is absorbed with the help of transporters. The neutral fraction is one factor, not a verdict: most aspirin is in fact absorbed in the small intestine, whose surface area dwarfs the stomach's, and as neutral molecules leave, the acid–base equilibrium replaces them.

Targeted protein degraders go further still. Instead of blocking a protein, a degrader hands it to the cell's disposal system. A PROTAC (proteolysis-targeting chimera) is a two-headed molecule: one end binds the target, the other an E3 ubiquitin ligase such as cereblon, the protein that thalidomide binds (Stereochemistry). Held together, the ligase tags the target with ubiquitin and the proteasome destroys it. The degrader is released to act again, and it can remove proteins that have no pocket worth inhibiting. PROTACs typically weigh 700 to 1100 g/mol, yet several are absorbed by mouth, and as of 2026 the first have reached late-stage clinical trials. The smaller molecular glues, including thalidomide's relatives lenalidomide and pomalidomide, reshape cereblon's surface so that it recognises new proteins. And many new medicines are not small molecules at all: antibodies, peptides such as the GLP-1 receptor agonists used in diabetes and obesity, and oligonucleotides that silence a gene's messenger RNA.

In industry these descriptors are computed with cheminformatics toolkits rather than typed in. With the open-source RDKit, for example, the calculation looks like this (not run here: RDKit is not available in the browser). Different programs compute log P by different methods, so their numbers differ slightly from PubChem's XLogP3.

python
# Not run here: needs RDKit (pip install rdkit)
from rdkit import Chem
from rdkit.Chem import Crippen, Descriptors, Lipinski, rdMolDescriptors

mol = Chem.MolFromSmiles("CC(=O)OC1=CC=CC=C1C(=O)O")      # aspirin
print(Descriptors.MolWt(mol), Crippen.MolLogP(mol),
      Lipinski.NumHDonors(mol), Lipinski.NumHAcceptors(mol),
      rdMolDescriptors.CalcTPSA(mol), Lipinski.NumRotatableBonds(mol))
ExplanationH

Antibiotics and resistance

In 1928 Alexander Fleming noticed that a mould contaminating one of his culture plates had killed the bacteria around it. A decade later Howard Florey, Ernst Chain, Norman Heatley and colleagues in Oxford purified the active substance, penicillin, and showed in 1940–41 that it cured infections in mice and then in people; Dorothy Hodgkin determined its structure by X-ray crystallography in 1945. Few medicines have saved more lives.

Penicillins, cephalosporins and carbapenems share a β-lactam, a four-membered cyclic amide. Bacteria build their cell walls from chains of sugars cross-linked by short peptides, and the cross-linking enzymes (transpeptidases, also called penicillin-binding proteins) attack a peptide bond with a serine. The β-lactam ring mimics that peptide, and because a four-membered ring is strained and its amide resonance is weak, it is unusually reactive: the serine opens it and is left acylated, permanently blocked. Without cross-links the wall fails, and the growing bacterium bursts. Human cells have no cell wall, which is why these antibiotics are selective.

Bacteria resist in a few chemical ways. They make β-lactamases, enzymes that hydrolyse the β-lactam ring before it reaches its target; they alter the target, as in MRSA (methicillin-resistant Staphylococcus aureus), whose extra penicillin-binding protein barely binds β-lactams; they pump the antibiotic out with efflux pumps; or they let less of it in. Chemists answer each move: β-lactamase inhibitors such as clavulanic acid, given together with amoxicillin, sacrifice themselves to the enzyme so that the antibiotic survives.

Resistance genes arise by mutation or are passed between bacteria, but they spread because of selection: an antibiotic kills the susceptible cells and leaves the resistant ones to multiply. The block is a deliberately simple model of this.

python
# A toy model of selection for resistance. Susceptible (S) and resistant (R)
# bacteria compete for the same space (carrying capacity K) and are washed out
# at rate d, as in a gut or a flowing culture. Resistance costs some growth
# rate. An antibiotic, given on days 0-6, adds a death rate for S only.
K = 1e9                       # cells
r_S, r_R = 2.0, 1.8           # growth rates, per day
d = 0.5                       # wash-out rate, per day
kill = 3.0                    # extra death rate of S while the drug is present, per day
dt = 0.001                    # time step, days

def step(S, R, drug):
    room = 1 - (S + R) / K
    dS = r_S * S * room - d * S - (kill * S if drug else 0.0)
    dR = r_R * R * room - d * R
    return S + dt * dS, R + dt * dR

S, R = 7.5e8, 750.0           # at equilibrium, with one resistant cell per million
report = {0, 2, 4, 6, 7, 14, 30, 60, 120, 240}
print(" day    susceptible     resistant   resistant fraction")
for day in range(241):
    if day in report:
        print(f"{day:4d}    {S:11.2e}   {R:11.2e}   {R / (S + R):10.4%}")
    for _ in range(int(round(1 / dt))):
        S, R = step(S, R, drug=day < 7)
Runs entirely in your browser, in an isolated worker. Nothing is sent to a server.

In this model a population with one resistant cell in a million is 99.7 % resistant after a week of treatment. When the drug stops, the 10 % growth cost of resistance lets susceptible cells win back slowly: still 91 % resistant after two months, 29 % after four, 0.1 % after eight. Real infections differ (the immune system clears bacteria, and treatment aims to eliminate an infection), and real resistance often costs less than 10 % because bacteria acquire compensating mutations, so resistance can persist for years. That is the logic of antimicrobial stewardship: antibiotics only when a qualified prescriber judges them necessary, and not as routine additives in farming.

A global analysis estimated that in 2019, 1.27 million deaths were directly attributable to bacterial resistance and 4.95 million associated with it (Antimicrobial Resistance Collaborators, 2022). New antibiotics are scientifically hard, because bacteria, especially Gram-negative ones with their double membranes, keep molecules out, and commercially hard, because a new antibiotic is held in reserve and sold little. Responses include new discovery methods, such as teixobactin from previously unculturable soil bacteria (2015) and halicin, found by a neural network trained on a few thousand measured molecules (Stokes et al., 2020), and new payment models, such as the NHS subscription in the UK that pays for access to an antibiotic rather than per pack sold.

ExplanationH

Regulators, ethics and access

Modern regulation was written in response to disasters. In 1937 more than a hundred people in the United States died after taking a sulfonamide syrup made with a toxic solvent that nobody had tested, and the law began to require proof of safety. Between 1957 and 1961 the sedative thalidomide, sold in dozens of countries and taken in pregnancy against morning sickness, caused severe birth defects in more than ten thousand babies (its stereochemistry is in Stereochemistry). The United States was largely spared because the FDA reviewer Frances Oldham Kelsey refused to approve it on incomplete safety data, and the 1962 Kefauver–Harris amendments then required proof of effectiveness as well as safety, through adequate and well-controlled trials.

Today regulators such as the FDA (United States), the European Medicines Agency, the PMDA (Japan), the NMPA (China), the CDSCO (India), ANVISA (Brazil) and SAHPRA (South Africa) review the evidence, and the African Medicines Agency, whose treaty came into force in 2021, coordinates across Africa. Through the ICH they share technical guidelines, so a company can file largely the same evidence everywhere. Clinical trials follow the World Medical Association's Declaration of Helsinki: informed consent, independent ethics review, and the principle that a participant's welfare comes before the interests of science. Registering every trial in a public registry before it starts makes it harder to hide unfavourable results.

Access is the hardest ethical question. A medicine that cannot be bought helps nobody. Patents reward discovery with a period of exclusive sales, after which generic manufacturers, many in India and China, make the same molecule for a fraction of the price. Around 2000, combination therapy for HIV cost over ten thousand US dollars per person per year; generic competition, public pressure and new funding cut the price by more than 95 % within a few years, and treatment reached millions. The WHO's Model List of Essential Medicines, first published in 1977, names the medicines every health system should be able to provide; the Medicines Patent Pool, founded in 2010, licenses patents so that generics can be made for low- and middle-income countries. Prices, rare diseases and neglected diseases of the poor are as much a part of medicinal chemistry as reaction mechanisms are.

Optional Christian reflection · Care for the stranger

A reflection from the Christian tradition that inspires Lumenary Academy. It is not part of the mathematics and you are welcome to skip it.

The parable of the Good Samaritan ends not with a definition of who counts as a neighbour but with an example: someone who stopped, treated a stranger's wounds and paid for his further care. Medicinal chemistry multiplies that kind of care: a molecule designed by a team in one country can relieve suffering for people its makers will never meet. Whether it reaches them depends on choices about prices, patents and priorities that every generation has to make again.

Common misconceptionsH

Common misconceptions

Misconception

"Natural remedies are safe; synthetic medicines are dangerous." Safety depends on the molecule and the amount, not on where it came from. Some of the most toxic substances known are natural, and many medicines (aspirin, penicillin, ciclosporin) are natural products or close relatives of them. Approved medicines are distinguished by having been tested.

Misconception

"A medicine acts on exactly one target." Most bind several proteins with different affinities, and many side effects come from those off-target interactions. Selectivity is a ratio of affinities, measured and optimised, never absolute.

Misconception

"Antibiotics work against colds and flu." Colds and flu are caused by viruses, which have no cell walls or bacterial ribosomes for antibiotics to attack. Using antibiotics against them helps nobody and selects for resistant bacteria.

Misconception

"Antibiotic resistance means a person's body has become resistant." It is the bacteria that become resistant, and resistant bacteria can pass from person to person. That is why resistance is a public-health problem and not only an individual one.

Exercises

Exercises

Two exercises ask for Python functions; run them in the editor before submitting.

1. Medicinal chemistry

Warm-upHuman

During lead optimisation, a compound's dissociation constant improves from Kd=1 μMK_\text{d} = 1\ \mu\text{M} to Kd=10K_\text{d} = 10 nM. Using ΔGbind∘=RTln⁡(Kd/c∘)\Delta G^\circ_\text{bind} = RT\ln(K_\text{d}/c^\circ) at body temperature (310.15 K, R=8.314R = 8.314 J mol⁻¹ K⁻¹), by how many kJ/mol has the binding free energy become more negative? Give the nearest whole number.

2. Medicinal chemistry

CoreHuman

A competitive inhibitor gives IC50=120\text{IC}_{50} = 120 nM in an enzyme assay run at a substrate concentration [S]=50 μ[\mathrm{S}] = 50\ \muM. The enzyme's Michaelis constant for this substrate is KM=10 μK_\text{M} = 10\ \muM. What is the inhibition constant KiK_\text{i} in nM?

3. Medicinal chemistry

StretchHuman

In an inhibitor series, the unsubstituted core has pIC50=5.7\text{pIC}_{50} = 5.7. Fitting a large SAR table additively (the Free–Wilson method) gives these contributions: bromine at position A adds 1.3 log units, and a methoxy group at position B adds 0.5. Assuming the effects are additive, what IC₅₀ in nM do you predict for the compound with bromine at A and methoxy at B? Give the nearest whole number.

4. Medicinal chemistry

Warm-upHuman

Using the 2011–2020 phase-transition rates in the lesson (Phase II → III: 28.9 %, Phase III → submission: 57.8 %, submission → approval: 90.6 %), how many of 200 programmes entering Phase II would you expect to reach approval? Give the nearest whole number.

5. Medicinal chemistry

CoreHuman

Why do β-lactam antibiotics such as penicillins kill bacteria while leaving human cells largely unaffected?

Choose one

6. Medicinal chemistry

CoreHuman

PubChem's computed polar surface area for amoxicillin is 158 Ų, above Veber's 140 Ų guideline, and its XLogP3 is −2.0. Yet amoxicillin is well absorbed when swallowed. What is the best explanation?

Choose one

7. Medicinal chemistry

CoreComputational

Write profile(mw, logp, hbd, hba, tpsa, rotb) that returns a dictionary with three entries:

  • "ro5_violations": how many of Lipinski's four limits are exceeded (molar mass > 500, log P > 5, donors > 5, acceptors > 10);
  • "ro5_ok": True when there is at most one violation;
  • "veber_ok": True when the polar surface area is at most 140 Ų and there are at most 10 rotatable bonds.

A value exactly at a limit (500, 5, 140, 10) does not violate it.

Runs entirely in your browser, in an isolated worker. Nothing is sent to a server.

8. Medicinal chemistry

StretchComputational

A research team models a hypothetical inhibitor. Write two functions.

  • free_needed(theta, kd) returns the free ligand concentration that gives fractional occupancy theta (between 0 and 1, exclusive) of a target with dissociation constant kd, from θ=[L]/(Kd+[L])\theta = [\mathrm{L}]/(K_\text{d} + [\mathrm{L}]).
  • total_needed(theta, kd, fu) returns the total plasma concentration needed when only the fraction fu of the compound is unbound (the rest is bound to plasma proteins and cannot reach the target).

Both results are in the same units as kd.

Runs entirely in your browser, in an isolated worker. Nothing is sent to a server.

Further reading and resources