Skip to content
LumenaryAcademy
Life & Earth 70 min

Biochemistry

The chemistry inside living things — proteins and enzymes, sugars and fats, DNA and RNA, the energy currency ATP, and what AlphaFold changed.

Code tested · 6AI-draftedAwaiting review by a chemist

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

MotivationH

A flask at 37 °C

Inside every cell of your body, thousands of different chemical reactions run at the same time, in water, near pH 7.4, at about 37 °C, without any flame, strong acid or high pressure. Sugars are taken apart and their energy is saved for later. Proteins are built from a text written in a four-letter alphabet. Membranes assemble themselves out of molecules that simply dislike water. Each reaction runs at the right speed because a specific catalyst, an enzyme, makes it run; without enzymes most of them would take years.

Biochemistry is the chemistry of these processes, and it uses everything earlier in this course: hydrogen bonds, free energy, pKa values, rate laws and handedness. Four families of molecules do most of the work: proteins, carbohydrates, lipids and nucleic acids. This lesson takes each in turn, follows glucose as a cell turns it into usable energy, and ends with the problem of predicting a protein's shape, which a neural network largely solved between 2020 and 2024.

IntuitionH

Amino acids and the four levels of protein structure

A protein is a chain of amino acids. Every amino acid in a protein has the same backbone, a carbon atom (the α-carbon) carrying an amino group, a carboxylic acid group and a hydrogen atom, plus one side chain, written R, that differs from one amino acid to the next. Twenty side chains are written into the genetic code. Some are oily hydrocarbons (leucine, valine, phenylalanine), some carry an alcohol or amide (serine, asparagine), some are acids (aspartate, glutamate) and some are bases (lysine, arginine, histidine). Glycine's side chain is just a hydrogen atom, and proline's loops back to bond to its own backbone nitrogen.

Every amino acid except glycine is chiral, and living things use almost exclusively the L enantiomer: (S) in the Cahn–Ingold–Prelog system of Stereochemistry for all except cysteine, which is (R) only because its sulfur atom changes the priority order.

Two amino acids join when the carboxylic acid of one condenses with the amino group of the next, releasing water and forming an amide, which biochemists call the peptide bond. Repeating this gives a chain with a free amino group at one end (the N-terminus) and a free carboxylate at the other (the C-terminus).

H₃N⁺CαR₁HCONHCαR₂HCOO⁻N-terminusC-terminuspeptide bondplanar amide group
A dipeptide. The peptide bond (highlighted) is an amide: because the nitrogen's lone pair is shared with the C=O group, the bond has partial double-bond character, cannot rotate freely, and holds six atoms (shaded) in one plane. The chain can twist only about the bonds to the α-carbons, which is why a protein backbone behaves like a string of flat tiles joined at hinges.

Biochemists describe the shape of a protein at four levels.

  1. Primary structure is the sequence of amino acids, read from the N-terminus. Human insulin's B chain begins Phe-Val-Asn-Gln-His-Leu, written FVNQHL in the one-letter code.
  2. Secondary structure is the local folding of the backbone, held by hydrogen bonds between backbone N–H and C=O groups. In an α-helix each C=O bonds to the N–H four residues further on, giving 3.6 residues per turn and a rise of 0.15 nm per residue. In a β-sheet, extended strands lie side by side, parallel or antiparallel, hydrogen-bonded across. Linus Pauling, Robert Corey and Herman Branson predicted the α-helix in 1951 from the geometry of the planar peptide bond, before any protein structure had been seen.
  3. Tertiary structure is the three-dimensional fold of the whole chain. Oily side chains bury themselves in the core, away from water (the hydrophobic effect, an entropy effect of the surrounding water, as in Entropy and free energy), while charged and polar ones stay on the surface. Hydrogen bonds, salt bridges between opposite charges, and covalent disulfide bridges between cysteines add stability.
  4. Quaternary structure is the assembly of several folded chains. Haemoglobin is four chains, two α and two β, whose cooperation lets it load oxygen in the lungs and release it in the tissues.

Around 1960 Christian Anfinsen unfolded the enzyme ribonuclease A, removed the unfolding agents, and watched it refold into its active shape unaided. His conclusion, that the sequence determines the structure, earned a share of the 1972 Nobel Prize in Chemistry and set a great problem: given a sequence, predict the fold.

Side chains with acidic or basic groups gain or lose protons as the pH changes, so a protein's charge depends on pH. At its isoelectric point (pI) the net charge is zero; proteins are least soluble there, which is used to separate them. The block below adds up the Henderson–Hasselbalch fractions for every ionisable group of a peptide and finds the pI by bisection.

python
import numpy as np

# Approximate pKa values of the ionisable groups in a peptide. Real values
# shift with neighbouring groups and solvent, and tables differ by a few tenths.
PKA = {"N-term": 9.0, "C-term": 2.2, "D": 3.9, "E": 4.1, "H": 6.0,
       "C": 8.3, "Y": 10.1, "K": 10.5, "R": 12.5}
BASIC = {"N-term", "H", "K", "R"}            # these carry +1 when protonated

def net_charge(seq, pH):
    groups = ["N-term", "C-term"] + [aa for aa in seq if aa in PKA]
    pka = np.array([PKA[g] for g in groups])
    protonated = 1.0 / (1.0 + 10.0 ** (pH - pka))   # Henderson–Hasselbalch
    basic = np.array([g in BASIC for g in groups])
    return float(np.sum(np.where(basic, protonated, protonated - 1.0)))

def isoelectric_point(seq, lo=0.0, hi=14.0):
    for _ in range(50):                      # bisection: charge falls as pH rises
        mid = (lo + hi) / 2
        lo, hi = (mid, hi) if net_charge(seq, mid) > 0 else (lo, mid)
    return (lo + hi) / 2

peptides = {"Gly-Gly": "GG", "Lys-Lys-Gly-Gly": "KKGG", "Asp-Asp-Gly-Ala": "DDGA",
            "glucagon (29 residues)": "HSQGTFTSDYSKYLDSRRAQDFVQWLMNT"}
print(f"{'peptide':24s} {'z, pH 2':>8s} {'z, pH 7.4':>10s} {'z, pH 12':>9s} {'pI':>6s}")
for name, seq in peptides.items():
    z = [net_charge(seq, p) for p in (2.0, 7.4, 12.0)]
    print(f"{name:24s} {z[0]:8.2f} {z[1]:10.2f} {z[2]:9.2f} {isoelectric_point(seq):6.2f}")
Runs entirely in your browser, in an isolated worker. Nothing is sent to a server.

Gly-Gly has no ionisable side chains, so its pI, 5.60, sits midway between the pKa values of its two ends. Two lysines push the pI up to 10.53; two aspartates pull it down to 2.88; glucagon, the hormone that raises blood sugar, comes out at 7.47. These are estimates: inside a folded protein a buried group's pKa can shift by several units, which is how many enzymes tune a catalytic residue.

DefinitionsH

Enzymes and Michaelis–Menten kinetics

An enzyme is a biological catalyst, almost always a protein (a few are RNA). Like any catalyst it lowers the activation free energy and speeds the forward and reverse reactions by the same factor, so it does not change the equilibrium constant. The size of the effect is remarkable: orotidine 5′-phosphate decarboxylase, which makes a building block of RNA, speeds its reaction by a factor of about 101710^{17}, from a half-life of millions of years to milliseconds.

The reaction happens in the active site, a pocket lined with side chains that binds the reactant (the substrate) and stabilises the transition state even more. Emil Fischer's 1894 picture of a lock and key has been refined into induced fit: the pocket changes shape as the substrate binds.

In 1913 Leonor Michaelis and Maud Menten measured how fast the enzyme invertase split sucrose at different sucrose concentrations and found a law that fits most enzymes: the rate rises in proportion to substrate at first, then levels off. Their model is the simplest mechanism that does this.

E+S⇌k1k−1ES→  k2  E+P\mathrm{E} + \mathrm{S} \underset{k_{-1}}{\overset{k_1}{\rightleftharpoons}} \mathrm{ES} \xrightarrow{\;k_2\;} \mathrm{E} + \mathrm{P}

Theorem— The Michaelis–Menten equation

Suppose the total enzyme concentration [E]T[\mathrm{E}]_\text{T} is much smaller than the substrate concentration, and that after a brief start the concentration of the complex ES stays constant (the steady-state assumption of George Briggs and J. B. S. Haldane, 1925). Then the initial rate of product formation is

v=Vmax[S]KM+[S],Vmax=k2[E]T,KM=k−1+k2k1.v = \frac{V_\text{max}[\mathrm{S}]}{K_\text{M} + [\mathrm{S}]}, \qquad V_\text{max} = k_2[\mathrm{E}]_\text{T}, \qquad K_\text{M} = \frac{k_{-1} + k_2}{k_1}.

Proof

The complex forms at rate k1[E][S]k_1[\mathrm{E}][\mathrm{S}] and breaks down at rate (k−1+k2)[ES](k_{-1} + k_2)[\mathrm{ES}]. At steady state these are equal. The free enzyme is whatever is not in the complex, [E]=[E]T−[ES][\mathrm{E}] = [\mathrm{E}]_\text{T} - [\mathrm{ES}], so

k1([E]T−[ES])[S]=(k−1+k2)[ES]  ⟹  [ES]=[E]T[S]KM+[S].k_1([\mathrm{E}]_\text{T} - [\mathrm{ES}])[\mathrm{S}] = (k_{-1} + k_2)[\mathrm{ES}] \;\Longrightarrow\; [\mathrm{ES}] = \frac{[\mathrm{E}]_\text{T}[\mathrm{S}]}{K_\text{M} + [\mathrm{S}]}.

Because [E]T≪[S][\mathrm{E}]_\text{T} \ll [\mathrm{S}], binding uses up a negligible part of the substrate, so [S][\mathrm{S}] here is the substrate concentration added. The product forms at v=k2[ES]v = k_2[\mathrm{ES}], which gives the equation.

∎

Definition— Vmax, Km and kcat

VmaxV_\text{max} is the limiting rate when the enzyme is saturated. KMK_\text{M}, the Michaelis constant, is the substrate concentration at which v=Vmax/2v = V_\text{max}/2. The turnover number kcat=Vmax/[E]Tk_\text{cat} = V_\text{max}/[\mathrm{E}]_\text{T} (equal to k2k_2 in the simple mechanism) is the number of substrate molecules one active site converts per second when saturated. The ratio kcat/KMk_\text{cat}/K_\text{M}, the specificity constant, is the effective second-order rate constant at low substrate concentration.

VmaxVmax/2Km[S]vfirst order: v ≈ (Vmax/Km)[S]zero order: v → Vmax
The Michaelis–Menten curve for Vmax = 40 µM/min and Km = 2.5 mM. At low [S] the rate is nearly proportional to [S] (first order); at high [S] it approaches Vmax (zero order: the enzyme is saturated). Km is read off where the rate is half of Vmax.

Carbonic anhydrase, which interconverts CO₂ and bicarbonate in your blood, turns over about a million molecules per second per active site; RuBisCO, which fixes CO₂ in plants (The chemistry of energy), manages a few. No enzyme can have kcat/KMk_\text{cat}/K_\text{M} much above 10810^8 to 109 M−1 s−110^9\ \text{M}^{-1}\,\text{s}^{-1}, the rate at which substrate diffuses into the active site.

Inhibitors slow enzymes, and many medicines are inhibitors. A competitive inhibitor resembles the substrate and competes for the active site. It raises the apparent KMK_\text{M} (more substrate is needed to reach half speed) but leaves VmaxV_\text{max} unchanged, because enough substrate always outcompetes it. The next lesson builds on this.

ComputationC

Fitting Vmax and Km to data

In the laboratory one measures initial rates viv_i at several substrate concentrations [S]i[\mathrm{S}]_i and must estimate VmaxV_\text{max} and KMK_\text{M}. For most of the twentieth century this was done with the Lineweaver–Burk plot: taking reciprocals of the Michaelis–Menten equation gives a straight line,

1v=KMVmax⋅1[S]+1Vmax,\frac{1}{v} = \frac{K_\text{M}}{V_\text{max}}\cdot\frac{1}{[\mathrm{S}]} + \frac{1}{V_\text{max}},

whose slope and intercept a ruler could find. The trouble is what the reciprocals do to the errors. The slowest rates, at low substrate, carry the largest relative error, and taking 1/v1/v magnifies exactly those points and places them far out on the line, where they control its slope.

The honest method is nonlinear least squares: choose VmaxV_\text{max} and KMK_\text{M} to minimise ∑i(vi−Vmax[S]i/(KM+[S]i))2\sum_i \bigl(v_i - V_\text{max}[\mathrm{S}]_i/(K_\text{M} + [\mathrm{S}]_i)\bigr)^2 (Optimisation). There is a neat shortcut: for a fixed KMK_\text{M} the model is linear in VmaxV_\text{max}, so the best VmaxV_\text{max} has a closed form, Vmax=∑fivi/∑fi2V_\text{max} = \sum f_i v_i / \sum f_i^2 with fi=[S]i/(KM+[S]i)f_i = [\mathrm{S}]_i/(K_\text{M} + [\mathrm{S}]_i). That leaves a search in one variable. The block fits one simulated experiment both ways, then repeats the experiment 2000 times to compare the methods fairly.

python
import numpy as np

S = np.array([0.25, 0.5, 1.0, 2.0, 4.0, 8.0, 16.0, 32.0])   # substrate, mM
Vmax_true, Km_true = 40.0, 2.5                               # µM/min, mM
rng = np.random.default_rng(2024)

def experiment():
    """Initial rates with a constant measurement error of 0.8 µM/min."""
    return Vmax_true * S / (Km_true + S) + 0.8 * rng.standard_normal(S.size)

Km_grid = np.geomspace(0.05, 50, 6000)
F = S / (Km_grid[:, None] + S)            # model shape for every trial Km
FF = np.sum(F * F, axis=1)

def fit_least_squares(v):
    """For fixed Km the model is linear in Vmax, so Vmax has a closed form;
    pick the Km on the grid with the smallest sum of squared residuals."""
    Fv = F @ v
    sse = v @ v - Fv ** 2 / FF
    i = np.argmin(sse)
    return Fv[i] / FF[i], Km_grid[i]

def fit_lineweaver_burk(v):
    slope, intercept = np.polyfit(1 / S, 1 / v, 1)
    return 1 / intercept, slope / intercept

v = experiment()
print(" [S]/mM :", " ".join(f"{s:6.2f}" for s in S))
print(" v      :", " ".join(f"{x:6.2f}" for x in v))
for name, fit in [("least squares", fit_least_squares), ("Lineweaver–Burk", fit_lineweaver_burk)]:
    Vmax, Km = fit(v)
    print(f"{name:16s} Vmax = {Vmax:5.1f} µM/min, Km = {Km:4.2f} mM")

# Repeat the whole experiment 2000 times to compare the two methods fairly.
results = {"least squares": [], "Lineweaver–Burk": []}
for _ in range(2000):
    v = experiment()
    results["least squares"].append(fit_least_squares(v)[1])
    results["Lineweaver–Burk"].append(fit_lineweaver_burk(v)[1])
print("\nKm over 2000 simulated experiments (true value 2.5 mM):")
for name, kms in results.items():
    kms = np.array(kms)
    lo, med, hi = np.percentile(kms, [5, 50, 95])
    print(f"{name:16s} median {med:4.2f} mM, 90 % of fits between {lo:4.2f} and {hi:4.2f}")
Runs entirely in your browser, in an isolated worker. Nothing is sent to a server.

On the first data set, least squares recovers Vmax=40.2V_\text{max} = 40.2 µM/min and KM=2.52K_\text{M} = 2.52 mM, close to the true 40 and 2.5; the double-reciprocal line gives 35.1 and 1.72, pulled off by the noise on the two slowest points.

Professional fitting software does the same with a general optimiser and also reports confidence intervals. The double-reciprocal plot survives as a way to display inhibition patterns, not to estimate constants.

ExplanationH

Carbohydrates and lipids

Carbohydrates have the rough formula (CH2O)n(\mathrm{CH_2O})_n. Glucose, C₆H₁₂O₆, is an aldehyde with five hydroxyl groups, but in water almost none of it stays as an open chain: the hydroxyl on carbon 5 attacks the aldehyde to close a six-membered ring (a cyclic hemiacetal, the kind of addition met in Carbonyls and aromatics). Ring closing creates a new stereocentre at carbon 1, so there are two forms, α and β, which interconvert in solution and settle at roughly one third α to two thirds β.

Sugars join through glycosidic bonds, and which carbon links to which, with which stereochemistry, decides everything. Sucrose is glucose joined to fructose; lactose is galactose joined to glucose. Starch and cellulose are both long chains of glucose. In starch the links are α-1,4 (with α-1,6 branches in amylopectin and in the animal store glycogen), and the chain coils into a helix that our digestive enzymes can attack. In cellulose they are β-1,4, which makes straight chains that lie side by side, hydrogen-bonded into tough fibres: wood, cotton and paper. Human enzymes cannot cut β-1,4 links, so a potato feeds you and a sheet of paper does not, though both are made of glucose.

Lipids dissolve poorly in water and well in oily solvents. A fatty acid is a long hydrocarbon chain ending in a carboxylic acid; three of them esterified to glycerol make a triglyceride, the fat and oil of food. A saturated chain, such as stearic acid (C18, melting point 69 °C), is straight and packs neatly into a solid. A cis double bond, as in oleic acid (C18, one double bond, melting point about 13 °C), kinks the chain, the molecules pack badly, and the fat is an oil at room temperature: olive oil against butter. Fats store about 37 kJ per gram, more than twice the figure for carbohydrate, because their chains are barely oxidised already (Thermochemistry).

Phospholipids replace one fatty acid with a charged phosphate group, giving a molecule with a water-loving head and two oily tails. In water they assemble spontaneously into a bilayer about 4 to 5 nm thick, tails inwards: the basic structure of every cell membrane. Nothing builds it; the hydrophobic effect does.

Note

A kitchen-safe way to see protein chemistry, with an adult: gelatine is the protein collagen, partly broken down. Fresh pineapple and kiwi fruit contain enzymes that cut proteins (bromelain and actinidin), so a jelly made with fresh pineapple pieces fails to set, while one made with tinned pineapple, which was heated during canning, sets normally. Heat unfolds (denatures) the enzyme, and an unfolded enzyme no longer works. Cooking an egg white, which turns from clear to white as its proteins unfold and tangle, shows denaturation too.

Worked examplesH

Nucleic acids and base pairing

A nucleotide has three parts: a phosphate, a five-carbon sugar and a flat, nitrogen-containing base. In DNA the sugar is deoxyribose and the bases are adenine (A), guanine (G), cytosine (C) and thymine (T); in RNA the sugar is ribose and uracil (U) replaces thymine. Phosphodiester bonds link the 3′ carbon of one sugar to the 5′ carbon of the next, so a strand has a direction, written from its 5′ end to its 3′ end.

In 1953 James Watson and Francis Crick, using X-ray photographs taken by Rosalind Franklin and Raymond Gosling and data from Maurice Wilkins, proposed the double helix: two strands running in opposite directions (antiparallel), wound around each other, with the bases stacked inside and paired by hydrogen bonds. A pairs with T through two hydrogen bonds; G pairs with C through three. The pairs have the same width, so any sequence fits the same helix, and each strand fully determines the other: that is how DNA is copied. Erwin Chargaff had already found that in any DNA the amount of A equals that of T, and G equals C, which pairing explains. In the common B form there are about 10.5 base pairs per turn, 0.34 nm apart.

5′3′5′3′ATGCCGTAAT
Base pairing in a short stretch of DNA. The two sugar–phosphate backbones run in opposite directions. A–T pairs are held by two hydrogen bonds (dashed), G–C pairs by three, so G–C-rich DNA needs more energy to separate.

Genes are read in two steps. In transcription, an enzyme copies a stretch of DNA into messenger RNA; in translation, the ribosome reads the RNA three bases (one codon) at a time and joins the matching amino acids. There are 43=644^3 = 64 codons for 20 amino acids and a stop signal, so most amino acids have several. The block writes the partner strand of a short invented gene, translates it, and estimates the melting temperature of short DNA primers with a laboratory rule of thumb.

python
# The standard genetic code, packed in the conventional T, C, A, G order:
# codon (first, second, third base) -> index 16*i + 4*j + k.  '*' marks a stop.
BASES = "TCAG"
AMINO = "FFLLSSSSYY**CC*WLLLLPPPPHHQQRRRRIIIMTTTTNNKKSSRRVVVVAAAADDEEGGGG"
CODE = {a + b + c: AMINO[16 * i + 4 * j + k]
        for i, a in enumerate(BASES) for j, b in enumerate(BASES) for k, c in enumerate(BASES)}
PAIR = {"A": "T", "T": "A", "G": "C", "C": "G"}

def reverse_complement(dna):
    """The partner strand, read 5'->3' (the strands are antiparallel)."""
    return "".join(PAIR[b] for b in reversed(dna))

def hydrogen_bonds(dna):
    """Watson–Crick pairs: A–T has two hydrogen bonds, G–C has three."""
    return sum(3 if b in "GC" else 2 for b in dna)

def translate(dna):
    protein = ""
    for i in range(0, len(dna) - 2, 3):
        aa = CODE[dna[i:i + 3]]
        if aa == "*":
            break
        protein += aa
    return protein

gene = "ATGGCTTGGAAAGAGCTGCACGGTTGA"          # an invented 27-base open reading frame
print("coding strand  5'-" + gene + "-3'")
print("template, 5'-3'  " + reverse_complement(gene))
gc = sum(b in "GC" for b in gene) / len(gene)
print(f"length {len(gene)} bp, GC content {gc:.0%}, hydrogen bonds {hydrogen_bonds(gene)}")
print("protein:", translate(gene))
print("codons per amino acid:", {aa: AMINO.count(aa) for aa in "LSRMW*"})

# Wallace rule of thumb for short primers (14-20 bases): Tm ≈ 2(A+T) + 4(G+C) °C
for primer in ["ATGGCTTGGAAAGAGC", "ATATATTATAATTATA", "GCGGCCGCGCCGGCGC"]:
    at = sum(b in "AT" for b in primer)
    print(f"{primer}  Tm ≈ {2 * at + 4 * (len(primer) - at)} °C")
Runs entirely in your browser, in an isolated worker. Nothing is sent to a server.

The gene codes for the eight-residue peptide MAWKELHG and then stops at TGA. Leucine, serine and arginine each have six codons, methionine and tryptophan only one, and three codons mean "stop". The primer made only of A and T is estimated to separate at about 32 °C, the G–C-rich one at about 64 °C, which is why laboratories designing primers for the polymerase chain reaction aim for a GC content near 50 %.

The rule is crude, and the reason is instructive. The extra hydrogen bond is only part of why G–C pairs hold more tightly: much of a duplex's stability comes from the stacking of neighbouring flat bases on top of each other, and G–C steps stack more strongly too. Professional primer-design tools therefore sum measured free energies over neighbouring pairs of pairs (the nearest-neighbour model) rather than counting hydrogen bonds.

ExplanationH

Metabolism in outline: ATP and glycolysis

A cell must run reactions that are uphill in free energy: building proteins, pumping ions, contracting muscle. It does so by coupling each one to a reaction that is strongly downhill, and the universal downhill reaction is the hydrolysis of adenosine triphosphate, ATP:

ATP+H2O⟶ADP+Pi,ΔG∘′=−30.5 kJ/mol.\mathrm{ATP} + \mathrm{H_2O} \longrightarrow \mathrm{ADP} + \mathrm{P_i}, \qquad \Delta G^{\circ\prime} = -30.5\ \text{kJ/mol}.

The prime marks the biochemical standard state: pH 7 rather than 1 M H⁺, with the other species at 1 M. As in Chemical equilibrium, the free energy available in the cell depends on the actual concentrations, ΔG=ΔG∘′+RTln⁡Q\Delta G = \Delta G^{\circ\prime} + RT\ln Q, and cells keep ATP far above its equilibrium level.

python
import numpy as np

R, T = 8.314e-3, 310.15          # kJ/(mol K); body temperature, 37 °C

# 1. ATP hydrolysis: ATP + H2O -> ADP + Pi, ΔG°' = -30.5 kJ/mol (pH 7, Mg2+ present).
dG0 = -30.5
ATP, ADP, Pi = 4.0e-3, 0.4e-3, 4.0e-3          # mol/L, illustrative cell values
Q = ADP * Pi / ATP
dG = dG0 + R * T * np.log(Q)
print(f"Q = {Q:.1e}:  ΔG = {dG0} + RT ln Q = {dG:.1f} kJ/mol in the cell")
K = np.exp(-dG0 / (R * T))
print(f"equilibrium constant K' = {K:.2e}, about {K / Q:.0e} times larger than Q")

# 2. Coupling: phosphorylating glucose is uphill on its own, downhill with ATP.
step = {"glucose + Pi -> glucose-6-phosphate + H2O": 13.8,
        "ATP + H2O -> ADP + Pi": -30.5}
coupled = sum(step.values())
for name, g in step.items():
    print(f"{name:44s} ΔG°' = {g:+6.1f} kJ/mol, K' = {np.exp(-g / (R * T)):.2e}")
print(f"{'coupled: glucose + ATP -> G6P + ADP':44s} ΔG°' = {coupled:+6.1f} kJ/mol, K' = {np.exp(-coupled / (R * T)):.2e}")

# 3. How much of glucose's free energy ends up in ATP?
dG_glucose = -2870.0                        # C6H12O6 + 6 O2 -> 6 CO2 + 6 H2O, kJ/mol
for label, n_atp in [("glycolysis only (no O2)", 2), ("full aerobic respiration", 30)]:
    for per_atp in (-dG0, -dG):
        print(f"{label:26s} {n_atp:2d} ATP x {per_atp:4.1f} kJ = {n_atp * per_atp:6.0f} kJ "
              f"= {100 * n_atp * per_atp / -dG_glucose:4.1f} % of 2870 kJ")
Runs entirely in your browser, in an isolated worker. Nothing is sent to a server.

With these illustrative concentrations, ATP hydrolysis in the cell releases about 51 kJ/mol, not 30.5: the cell gets extra free energy by keeping the reaction far from equilibrium. Coupling is just addition of free energies. Phosphorylating glucose directly has an equilibrium constant of about 0.005, hopelessly uphill; coupled to ATP hydrolysis inside one enzyme, hexokinase, it has an equilibrium constant of about 650.

Glycolysis, found in almost every living thing, splits glucose into two molecules of pyruvate in ten enzyme-catalysed steps. The first steps spend two ATP to make the sugar reactive; it is then cut into two three-carbon pieces, and the later steps oxidise each piece and make two ATP each. The net yield is 2 ATP and 2 NADH (an electron carrier) per glucose, with no oxygen needed: sprinting muscles and bread yeast live on this.

With oxygen, pyruvate enters the mitochondria, the citric acid cycle strips it to CO₂, and the electrons collected in NADH flow through a chain of proteins to oxygen. That flow pumps protons across the inner mitochondrial membrane, and the protons flowing back turn ATP synthase, a rotary motor (Peter Mitchell's chemiosmotic theory of 1961, Nobel Prize 1978). The total is about 30 to 32 ATP per glucose; older books say 36 to 38, from earlier estimates of the proton cost of each ATP.

The last lines of the block do the accounting. Glycolysis alone captures about 2 % of glucose's free energy at the standard value per ATP (3.5 % at the cellular value); full respiration captures roughly 32 % to 53 %. The rest leaves as heat, the hundred-watt warmth of Thermochemistry.

ExplanationH

The professional view: AlphaFold and the structure revolution

For sixty years, the only way to know a protein's shape was to measure it: by X-ray crystallography (myoglobin and haemoglobin around 1958 to 1960, by John Kendrew and Max Perutz), by NMR spectroscopy (Spectroscopy) and, more recently, by cryo-electron microscopy. The Protein Data Bank, founded in 1971, now holds over 200 000 experimental structures. Each can take months or years of work, and some proteins resist every method.

Since 1994 the CASP competition has tested prediction blindly, on structures solved but not yet released. Progress was slow for two decades. At CASP14 in 2020, AlphaFold 2 from DeepMind produced models whose median accuracy on the competition's 0–100 GDT scale was about 92, comparable to experimental structures for most targets (Jumper et al., 2021). The network learns from the PDB and from multiple sequence alignments: when two positions in a protein family mutate together across evolution, they are probably in contact. Its core uses attention, the mechanism described in Lumenary CS's attention and transformers, to pass information between residues and between pairs of residues. In 2021 David Baker's group at the University of Washington published RoseTTAFold, with similar ideas and open code.

What changed:

  • Coverage. With EMBL-EBI, DeepMind released the AlphaFold Protein Structure Database, which by 2022 held predictions for over 200 million proteins, nearly every known sequence. A biologist studying a newly found protein now starts with a plausible model rather than nothing.
  • Experimental work sped up. Predicted models serve as starting points for solving crystal structures and for interpreting cryo-EM maps of huge complexes.
  • Design. Baker's group and others turned the problem round: generate a structure with a desired function, then find a sequence that folds into it. Designed proteins now include new binders, enzymes and vaccine components.
  • Complexes. AlphaFold 3 (Abramson et al., 2024) predicts how proteins sit with DNA, RNA, ions and small molecules, which matters for drug discovery. Its code was released for academic use in late 2024, and open re-implementations such as OpenFold and Boltz followed.

The 2024 Nobel Prize in Chemistry went half to David Baker, for computational protein design, and half jointly to Demis Hassabis and John Jumper, for protein structure prediction.

What did not change matters as much. AlphaFold predicts one static fold, not the motions a protein makes while it works, nor the folding pathway. Its per-residue confidence score (pLDDT) must be read: low-confidence regions are often genuinely disordered. Predictions for point mutations, for proteins unlike the training data, and for the exact poses of drug molecules are much less reliable, and important predictions are still checked by experiment. AI for chemistry returns to these methods.

Common misconceptionsH

Common misconceptions

Misconception

"ATP stores energy in a high-energy bond that releases energy when it breaks." Breaking any bond costs energy. ATP hydrolysis releases free energy because the products, ADP and phosphate in water, are much more stable than ATP and water: less charge repulsion, better resonance stabilisation and better solvation. The free energy belongs to the whole reaction, not to one bond.

Misconception

"Enzymes make reactions go that otherwise would not." An enzyme changes how fast equilibrium is reached, never where it lies. A reaction with a positive ΔG\Delta G runs in the cell only when it is coupled to a downhill one, such as ATP hydrolysis, so that the combined ΔG\Delta G is negative.

Misconception

"AlphaFold solved protein folding, so experiments are no longer needed." AlphaFold predicts the folded structure of many proteins accurately; it does not describe the folding pathway, the motions of the protein, or every complex and mutation reliably. Its predictions are hypotheses, and experimental structures remain the standard, and the training data, for the field.

Exercises

Exercises

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

1. Biochemistry

Warm-upHuman

An α-helix is held together by hydrogen bonds between the C=O group of one residue and the N–H group of the residue four places further along the same chain. Which level of protein structure is it?

Choose one

2. Biochemistry

CoreHuman

An enzyme follows Michaelis–Menten kinetics with KM=2K_\text{M} = 2 mM. At what substrate concentration, in mM, does it run at 80 % of VmaxV_\text{max}?

3. Biochemistry

Warm-upHuman

The DNA strand 5′-GATTACA-3′ is paired with its complementary strand. How many hydrogen bonds hold the two strands together (counting two per A–T pair and three per G–C pair)?

4. Biochemistry

CoreHuman

Write the strand complementary to 5′-GATTACA-3′, read in the usual 5′ → 3′ direction. Give just the letters.

5. Biochemistry

CoreHuman

A competitive inhibitor binds reversibly in an enzyme's active site, competing with the substrate. How do the measured VmaxV_\text{max} and KMK_\text{M} change when it is present?

Choose one

6. Biochemistry

Warm-upHuman

Starch and cellulose are both long chains made only of glucose. Why can humans digest starch but not cellulose?

Choose one

7. Biochemistry

CoreComputational

Write fit_michaelis_menten(S, v), which returns (Vmax, Km) minimising the sum of squared residuals ∑i(vi−VmaxSi/(KM+Si))2\sum_i (v_i - V_\text{max}S_i/(K_\text{M} + S_i))^2. Do not use a Lineweaver–Burk line: the tests include noisy data on which it gives the wrong answer. NumPy is available.

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

8. Biochemistry

CoreComputational

Write first_protein(dna), which finds the first ATG in a DNA string (upper or lower case), translates from there codon by codon with the standard genetic code (CODE is provided), and stops before the first in-frame stop codon (*). If there is no ATG, return ""; if no stop codon follows, translate every complete codon and ignore any leftover one or two bases.

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

Further reading and resources