Organic synthesis

Class 12 Chemistry, unit 14, book pages 227–230

Practise 25 MCQs Take the unit quiz Review 28 cards

The book's unit 14 is about artificial intelligence in organic synthesis: how AI tools process huge amounts of data, screen millions of molecules, generate new structures, optimise lead compounds and predict activity and toxicity, so that new medicines are found faster and more cheaply. The case of halicin, a diabetes drug that an MIT deep learning model identified as an antibiotic, shows the whole process. Two supplementary sections link the reactions of units 8 to 13 into synthetic routes and explain retrosynthesis, which FBISE examined in 2026.

Learning outcomes (1)

Describe the use of artificial intelligence tools in designing organic molecules which may have the potential to be used as medicine (halicin can be used as an example).

Unit 14 introductionArtificial intelligence in organic synthesis

Organic synthesis is the planned making of organic compounds from simpler ones. Its most valuable use is making new medicines (drugs). Finding a new drug the traditional way is slow and costly: chemists make and test thousands of compounds, and it often takes 10 to 15 years and more than a billion US dollars to bring one medicine to patients. Most candidate molecules fail along the way.

Artificial intelligence (AI) is the ability of a computer program to do tasks that normally need human intelligence, such as recognising patterns, learning from examples and making predictions. In this unit the book shows how AI tools help chemists decide which molecules are worth making and testing as medicines.

Artificial intelligence (AI): Computer systems that learn from data and make predictions or decisions that normally need human intelligence.
Machine learning (ML): A branch of AI in which a program learns patterns from a large set of examples (data) instead of being given fixed rules.
Deep learning: Machine learning that uses artificial neural networks with many layers; it can learn very complex patterns, for example between a molecule's structure and its activity.
Drug candidate: A molecule that is expected to have a useful effect on the body and is chosen for further testing.

How can a computer 'read' a molecule? The structure is turned into numbers. One way is a SMILES string, a line of text that describes the atoms and bonds (ethanol is $CCO$ ); another is a molecular graph, in which atoms are points and bonds are lines. The model learns links between these structures and measured properties such as antibacterial activity or toxicity, then predicts the same properties for molecules nobody has tested.

The main stages of drug discovery and development, and where AI tools help.

14.1Influence of artificial intelligence on pharmaceutical researchOften examined

AI has strongly influenced pharmaceutical research and development, especially the search for natural and synthetic compounds that could be used as medicines. The book lists five advantages of AI tools. Learn the five headings; each makes a good short answer.

1. Data-driven discovery

AI can process vast amounts of chemical, biological and pharmacological data, far more than a team of scientists could read, to identify possible drug candidates. It finds patterns that link structure to activity. Example: halicin was discovered by an AI model trained on a dataset of known medicinal compounds; the model predicted antibacterial properties of halicin that had not been noticed before.

2. Virtual screening

Virtual screening is the computer-based (online) testing of huge libraries of chemical compounds to pick out those most likely to interact with a particular biological target, such as an enzyme or a bacterial membrane. Only the top-ranked molecules are then made or bought and tested in the laboratory. This speeds up the early stages of drug development and narrows the pool of candidates from millions to a manageable number. For halicin, the AI examined over 100 million molecules in the ZINC15 database, a free online library of commercially available compounds.

Virtual screening does not replace laboratory tests; it decides which few molecules are worth testing.
	Traditional (high-throughput) screening	Virtual screening
How	each compound is physically tested in the lab	a computer model scores each compound
Number tested	thousands to about a million	hundreds of millions
Time and cost	months, expensive chemicals and robots	days, mainly computer time
Output	measured activity	a short ranked list, which must then be tested in the lab

3. Molecular generation

Instead of only searching known molecules, generative models invent new ones. They learn the patterns of real chemical structures and then create new molecules with desired properties that can be synthesised and tested. They are useful for exploring chemical structures outside the known compounds. Two types named in the book:

Variational autoencoder (VAE): an encoder turns each molecule into a short list of numbers (a point in a 'chemical space'), and a decoder turns numbers back into a molecule. Picking new points between or near good molecules gives new, similar molecules.
Generative adversarial network (GAN): two networks compete. The generator makes new molecules; the discriminator tries to tell them apart from real drug-like molecules. As they compete, the generator learns to make realistic, useful molecules.

4. Optimizing lead compounds

A lead compound is a molecule that already shows the desired biological activity but needs improvement (for example it is too weak, too toxic or breaks down too quickly in the body). Chemists change its structure step by step: add or swap functional groups, change chain length, and so on. AI speeds up this cycle of design and testing by predicting how each change in structure will affect performance (potency), stability and safety before the molecule is made, and it recommends the most promising changes.

5. Prediction of biological activity and toxicity

Machine learning models forecast how a substance will behave in the body and whether it is likely to be toxic. By comparing a new molecular structure with existing data on similar structures, the model flags potentially dangerous substances early, which reduces the need for extensive laboratory and animal testing. Machine learning also helps to understand how new medicines work (their mechanism) and to propose their molecular effects in living systems, which guides further improvement.

14.2Case study: halicinOften examined

Halicin is the best-known example of a medicine candidate found by AI. It was first studied as a possible diabetes drug but was not developed further. In 2020 researchers at the Massachusetts Institute of Technology (MIT), Cambridge, USA, found with a deep learning model that it is a powerful antibiotic. It is named after HAL 9000, the computer in the film 2001: A Space Odyssey.

halicin

C X_{5} H X_{3} N X_{5} O X_{2} S X_{3}

: a nitrothiazole ring joined through S to an aminothiadiazole ring

Key steps and the role of AI

The halicin story (book section 14.2, with standard details from the MIT study).
Step	What was done	Role of AI
1. Training	About 2300 molecules (known drugs and natural products) were tested for their ability to stop the growth of the bacterium E. coli.	A deep learning model learned the link between molecular structure and antibacterial activity.
2. Screening a drug library	The trained model scored about 6000 molecules already studied as drugs.	It ranked halicin highly, although its structure is unlike any known antibiotic (drug repurposing).
3. Screening ZINC15	Over 100 million molecules were scored in a few days.	Virtual screening found further candidates; this would be impossible in a laboratory.
4. Predicting action and safety	Top candidates were checked for predicted toxicity.	Halicin was predicted to have low toxicity to human cells and to disrupt bacterial cell membranes.
5. Laboratory confirmation	Halicin was tested on bacteria and in mice.	The AI prediction was confirmed: it killed many antibiotic-resistant bacteria.

How it works: halicin disrupts the bacterial cell membrane. It destroys the difference in hydrogen-ion concentration (the electrochemical gradient) across the membrane that bacteria need to make ATP (energy). This is a different target from most antibiotics, which is why bacteria that resist older drugs are still killed, and why in the laboratory E. coli did not become resistant to halicin in 30 days while it became resistant to ciprofloxacin within days.

Results: laboratory tests showed that halicin works against many antibiotic-resistant bacteria, for example Acinetobacter baumannii, Clostridioides difficile and Mycobacterium tuberculosis. Halicin is still being studied; it has not yet been approved as a medicine for humans.

Why the case matters

It proved that AI can find a useful drug that human experts had missed, because its structure did not look like an antibiotic.
It is an example of drug repurposing: finding a new use for an existing compound, which saves time because some safety data already exist.
It shows how AI can help against antibiotic resistance, one of the most serious health threats.
The success of halicin underscores the power of AI to identify and optimize new drugs, and promises faster, more effective development of medicines in future.

14.3Advantages of AI in drug development

The book sums up the use of AI in drug development in three advantages:

Speed: AI greatly reduces the time needed to find new drug candidates: screening that would take years in a laboratory takes days on a computer.
Cost-effectiveness: By narrowing down the list of possible candidates early, AI reduces the costs of experimental testing: fewer chemicals, fewer failed experiments and fewer animal tests.
Innovation: AI can explore new chemical structures that traditional methods may not consider, leading to entirely new classes of drugs (halicin is structurally unlike older antibiotics).

Limitations (for a balanced long answer)

A model is only as good as its training data: too little or poor-quality data gives wrong predictions.
Many models are a 'black box': they do not explain why a molecule should work.
A predicted molecule may be very hard or expensive to synthesise.

Supplement (units 8 to 13)Synthetic routes and functional group interconversions

A synthetic route is the sequence of reactions used to make a target molecule from available starting materials. Most routes change one functional group into another (a functional group interconversion) and sometimes lengthen the carbon chain. The reactions below come from units 8 to 13 and from grade 11; together they form a 'map' for planning routes.

Key interconversions. Halogenoalkanes and alcohols are the 'crossroads' of organic synthesis.

Reagents and conditions to learn for route questions.
Change	Reagent and conditions	Type of reaction
alkane $\to$ halogenoalkane	$Cl X_{2}$ or $Br X_{2}$ , UV light	free radical substitution
alkene $\to$ alkane	$H X_{2}$ , Ni catalyst, $15 0^{\circ} C$	addition (hydrogenation)
alkene $\to$ halogenoalkane	$HBr$ (or $HCl$ ), room temperature	electrophilic addition
alkene $\to$ alcohol	steam, $H X_{3} PO X_{4}$ catalyst, $30 0^{\circ} C$ , $60 atm$	electrophilic addition (hydration)
halogenoalkane $\to$ alcohol	$NaOH (aq)$ or $KOH (aq)$ , heat under reflux	nucleophilic substitution
halogenoalkane $\to$ alkene	$KOH$ in ethanol, heat	elimination
halogenoalkane $\to$ nitrile	$KCN$ in ethanol, heat under reflux	nucleophilic substitution (adds 1 C)
halogenoalkane $\to$ amine	excess concentrated $NH X_{3}$ in ethanol, heat in a sealed tube	nucleophilic substitution
nitrile $\to$ carboxylic acid	dilute $HCl$ or $H X_{2} SO X_{4}$ , heat under reflux	hydrolysis
nitrile $\to$ amine	$LiAlH X_{4}$ in dry ether (or $H X_{2}$ , Ni)	reduction
alcohol $\to$ halogenoalkane	$PCl X_{5}$ , $SOCl X_{2}$ , or $HBr$ (from $NaBr + conc. H X_{2} SO X_{4}$ )	substitution
alcohol $\to$ alkene	concentrated $H X_{2} SO X_{4}$ , $17 0^{\circ} C$ (or $Al X_{2} O X_{3}$ , $35 0^{\circ} C$ )	elimination (dehydration)
1° alcohol $\to$ aldehyde	$K X_{2} Cr X_{2} O X_{7}$ / dilute $H X_{2} SO X_{4}$ , warm, distil the aldehyde off as it forms	oxidation
1° alcohol or aldehyde $\to$ carboxylic acid	excess $K X_{2} Cr X_{2} O X_{7}$ / dilute $H X_{2} SO X_{4}$ , heat under reflux	oxidation
2° alcohol $\to$ ketone	$K X_{2} Cr X_{2} O X_{7}$ / dilute $H X_{2} SO X_{4}$ , heat	oxidation
aldehyde or ketone $\to$ alcohol	$NaBH X_{4}$ (or $LiAlH X_{4}$ )	reduction
aldehyde or ketone $\to$ hydroxynitrile	$HCN$ with a little $KCN$	nucleophilic addition (adds 1 C)
carboxylic acid $\to$ ester	alcohol, a few drops of concentrated $H X_{2} SO X_{4}$ , heat	esterification (condensation)
carboxylic acid $\to$ acyl chloride	$PCl X_{5}$ or $SOCl X_{2}$	substitution
benzene $\to$ nitrobenzene	concentrated $HNO X_{3}$ + concentrated $H X_{2} SO X_{4}$ , below $5 5^{\circ} C$	electrophilic substitution
nitrobenzene $\to$ phenylamine	$Sn$ + concentrated $HCl$ , heat, then $NaOH$	reduction
phenylamine $\to$ benzenediazonium chloride	$NaNO X_{2} + HCl$ , below $1 0^{\circ} C$	diazotisation
benzenediazonium ion $\to$ phenol	warm with water	hydrolysis ( $N X_{2}$ evolved)
benzene $\to$ methylbenzene	$CH X_{3} Cl$ , anhydrous $AlCl X_{3}$	Friedel-Crafts alkylation

Making the carbon chain longer

Most interconversions keep the number of carbon atoms the same. When the target has one more carbon than the starting material, a $C - C$ bond must be made. The two reactions at this level that do this both use the cyanide ion:

CH X_{3} CH X_{2} Br + CN X^{-} ethanol reflux CH X_{3} CH X_{2} CN + Br X^{-}

halogenoalkane to nitrile: bromoethane (2 C) gives propanenitrile (3 C)

CH X_{3} CHO + HCN KCN CH X_{3} CH (OH) CN

aldehyde to hydroxynitrile: ethanal (2 C) gives 2-hydroxypropanenitrile (3 C)

The $- CN$ group is then hydrolysed to $- COOH$ or reduced to $- CH X_{2} NH X_{2}$ . Friedel-Crafts alkylation and acylation do the same job for benzene rings.

Choosing a good route

Fewer steps are better. Every step loses some product, so the overall yield falls quickly with the number of steps.
Each step should give one main product without unwanted isomers or side reactions.
Starting materials should be cheap and available; reagents and conditions should be safe (for example, $KCN$ and $HCN$ are very toxic and need great care).

overall yield = y_{1} \times y_{2} \times y_{3} \times \dots

$y_{1}, y_{2}, \dots$ : fractional yields of the separate steps

Overall yield of a three-step route

A route has three steps with yields of

80%

70%

and

90%

. What is the overall yield?

Multiply the fractional yields, because each step uses only the product of the step before.
$0.80 \times 0.70 \times 0.90 = 0.504$
Convert to a percentage.
$0.504 \times 100% = 50.4%$

overall yield \approx 50%

Supplement (section 9.6 and units 8 to 13)Retrosynthesis: working backwards from the target

Retrosynthesis (retrosynthetic analysis) is a way of planning a synthesis by starting from the target molecule and working backwards, one step at a time, to simple starting materials. At each step you ask: 'Which known reaction could make this compound, and from what?' The backward step is shown with an open arrow, $\Rightarrow$ , which means 'can be made from'. The idea was developed by E. J. Corey (Nobel Prize in Chemistry, 1990). The book introduces it for halogenoalkanes in section 9.6.

Target molecule: The compound that has to be made.
Disconnection: An imaginary breaking of a bond in the target to find simpler pieces that a known reaction could join.
Precursor: The compound one step before the target in a synthetic route.

Method

Identify the functional group(s) in the target and in the starting material given.
Count the carbon atoms. If the chain is longer by one, a $C - C$ bond must be formed (use $CN X^{-}$ ).
Work backwards: find a reaction that makes the target's functional group and write its precursor. Repeat until you reach the starting material.
Now write the route forwards, with reagents and conditions over each arrow, and check each step.

Retrosynthesis of propan-2-ol from a halogenoalkane

Given the alcohol product

CH X_{3} CH (OH) CH X_{3}

, deduce the halogenoalkane starting material using retrosynthesis.

The target is a secondary alcohol with 3 carbon atoms. An $- OH$ group is put on a carbon by nucleophilic substitution of a halogen by $OH X^{-}$ , so disconnect the $C - OH$ bond.
$CH X_{3} CH (OH) CH X_{3} \Rightarrow CH X_{3} CH (X) CH X_{3} + OH X^{-}$
The carbon that carried $OH$ must have carried the halogen. So the precursor is 2-bromopropane (2-chloropropane or 2-iodopropane would also work).
$CH X_{3} CH (OH) CH X_{3} \Rightarrow CH X_{3} CHBrCH X_{3}$
Write the forward reaction with conditions.
$CH X_{3} CHBrCH X_{3} + NaOH (aq) heat CH X_{3} CH (OH) CH X_{3} + NaBr$

2-bromopropane starting material

propan-2-ol target

starting material: CH X_{3} CHBrCH X_{3} (2-bromopropane)

Ethanol to propanoic acid (chain lengthened by one carbon)

Suggest a three-step synthesis of propanoic acid,

CH X_{3} CH X_{2} COOH

, from ethanol.

Analysis. The target has 3 C, ethanol has 2 C, so a $C - C$ bond must be made with $CN X^{-}$ . Working backwards:

CH X_{3} CH X_{2} COOH \Rightarrow CH X_{3} CH X_{2} CN \Rightarrow CH X_{3} CH X_{2} Br \Rightarrow CH X_{3} CH X_{2} OH

Forward route:

Step 1: convert the alcohol to a halogenoalkane (substitution).
$CH X_{3} CH X_{2} OH + HBr NaBr, conc. H2SO4 CH X_{3} CH X_{2} Br + H X_{2} O$
Step 2: add one carbon with cyanide (nucleophilic substitution).
$CH X_{3} CH X_{2} Br + KCN ethanol reflux CH X_{3} CH X_{2} CN + KBr$
Step 3: hydrolyse the nitrile with dilute acid.
$CH X_{3} CH X_{2} CN + 2 H X_{2} O + HCl reflux CH X_{3} CH X_{2} COOH + NH X_{4} Cl$

ethanol

bromoethane

propanenitrile

propanoic acid

CH X_{3} CH X_{2} OH CH X_{3} CH X_{2} Br CH X_{3} CH X_{2} CN CH X_{3} CH X_{2} COOH

Ethene to ethyl ethanoate

Using ethene as the only organic starting material, show how ethyl ethanoate,

CH X_{3} COOCH X_{2} CH X_{3}

, can be made.

Analysis. An ester is made from an acid and an alcohol: $CH X_{3} COOCH X_{2} CH X_{3} \Rightarrow CH X_{3} COOH + CH X_{3} CH X_{2} OH$ . Both pieces have 2 C, so both can come from ethene: ethanol by hydration, and ethanoic acid by oxidising part of that ethanol.

Hydration of ethene gives ethanol.
$CH X_{2} = CH X_{2} + H X_{2} O H3PO4 300 X^{\circ} C, 60 atm CH X_{3} CH X_{2} OH$
Oxidise part of the ethanol fully to ethanoic acid (excess oxidant, reflux).
$CH X_{3} CH X_{2} OH + 2 [O] K2Cr2O7/H+ reflux CH X_{3} COOH + H X_{2} O$
Esterify the acid with the rest of the ethanol.
$CH X_{3} COOH + CH X_{3} CH X_{2} OH conc. H2SO4 heat CH X_{3} COOCH X_{2} CH X_{3} + H X_{2} O$

ethene

ethanol

ethanoic acid

ethyl ethanoate

Benzene to phenol

Outline a synthesis of phenol from benzene, giving reagents and conditions.

Analysis. An $- OH$ cannot be put directly on a benzene ring by substitution. Phenol is made from a diazonium salt, which comes from phenylamine, which comes from nitrobenzene: $C X_{6} H X_{5} OH \Rightarrow C X_{6} H X_{5} N X_{2} X^{+} \Rightarrow C X_{6} H X_{5} NH X_{2} \Rightarrow C X_{6} H X_{5} NO X_{2} \Rightarrow C X_{6} H X_{6}$ .

Nitration (electrophilic substitution).
$C X_{6} H X_{6} + HNO X_{3} conc. H2SO4 55 X^{\circ} C C X_{6} H X_{5} NO X_{2} + H X_{2} O$
Reduction of the nitro group.
$C X_{6} H X_{5} NO X_{2} + 6 [H] Sn, conc. HCl heat C X_{6} H X_{5} NH X_{2} + 2 H X_{2} O$
Diazotisation in the cold (the diazonium salt decomposes above about $1 0^{\circ} C$ ).
$C X_{6} H X_{5} NH X_{2} + NaNO X_{2} + 2 HCl below 10 X^{\circ} C C X_{6} H X_{5} N X_{2} X^{+} Cl X^{-} + NaCl + 2 H X_{2} O$
Warm the diazonium solution with water; nitrogen gas is given off.
$C X_{6} H X_{5} N X_{2} X^{+} Cl X^{-} + H X_{2} O warm C X_{6} H X_{5} OH + N X_{2} + HCl$

benzene

nitrobenzene

phenylamine

phenol

Formula sheet

Halogenoalkane to alcohol

\ce{R-X + OH- ->[\text{aq},\ \text{heat}] R-OH + X-}

\ce{R}: alkyl group; \ce{X}: Cl, Br or I

Halogenoalkane to nitrile (adds one carbon)

\ce{R-X + CN- ->[\text{ethanol}][\text{reflux}] R-CN + X-}

the new \ce{C-C} bond lengthens the chain by one carbon

Hydrolysis of a nitrile

\ce{R-CN + 2H2O + H+ ->[\text{reflux}] R-COOH + NH4+}

dilute acid, heat under reflux

Esterification

\ce{RCOOH + R'OH <=>[\text{conc. H2SO4}][\text{heat}] RCOOR' + H2O}

\ce{R}, \ce{R'}: alkyl groups

Retrosynthetic step

\text{target} \Rightarrow \text{precursor}

the open arrow \Rightarrow means 'can be made from'

Overall yield of a multistep route

\text{overall yield} = y_1 \times y_2 \times y_3 \times \cdots

y_1, y_2, \ldots: fractional yields of the separate steps

Units: fraction or percentage

Key terms

Organic synthesis: The planned preparation of organic compounds, often medicines, from simpler starting materials.
Artificial intelligence (AI): Computer systems that learn from data and make predictions or decisions that normally need human intelligence.
Machine learning: A branch of AI in which a program learns patterns from examples (data) instead of following fixed rules.
Deep learning: Machine learning with artificial neural networks of many layers, able to learn complex links such as structure to activity.
Drug candidate: A molecule expected to have a useful effect on the body and chosen for further testing.
Data-driven discovery: Using AI to process vast chemical, biological and pharmacological data to identify possible drug candidates.
Virtual screening: Computer-based testing of very large libraries of compounds to find those most likely to interact with a biological target.
ZINC15: A free online database of over 100 million commercially available compounds; it was screened by AI in the discovery of halicin.
Molecular generation: Use of generative AI models to create new molecules with desired properties.
Variational autoencoder (VAE): A generative model that encodes molecules as numbers in a chemical space and decodes new points back into new molecules.
Generative adversarial network (GAN): A generative model in which a generator makes new molecules and a discriminator judges them, so the generator improves.
Lead compound: A molecule that shows the desired biological activity and is used as the starting point for structural improvement.
Lead optimisation: Improving a lead compound's potency, stability and safety by changing its structure; AI predicts the effect of each change.
Toxicity prediction: Use of machine learning to forecast whether a substance is likely to be harmful before it is tested.
Drug repurposing: Finding a new medical use for a compound originally developed for another purpose.
Halicin: A compound (\ce{C5H3N5O2S3}) first developed as a diabetes drug and identified by AI at MIT as an antibiotic effective against resistant bacteria.
Antibiotic resistance: The ability of bacteria to survive antibiotics that used to kill them.
Synthetic route: The sequence of reactions used to make a target molecule from available starting materials.
Functional group interconversion: A reaction that changes one functional group into another, such as an alcohol into an aldehyde.
Retrosynthesis: Planning a synthesis by working backwards from the target molecule to simpler starting materials.
Target molecule: The compound that has to be synthesised.
Disconnection: An imaginary breaking of a bond in the target to find simpler pieces that a known reaction could join.

Multiple choice questionsp. 228

What is the primary function of AI in drug discovery as exemplified by the discovery of Halicin?

AManufacturing drugs
BPredicting biological activity and toxicity
CPerforming virtual screening
DBoth B and CAnswer

Why

In the halicin study the AI model did both jobs: it predicted antibacterial activity (and low toxicity) from structure, and it virtually screened over 100 million molecules in the ZINC15 database. AI does not manufacture drugs, so A is wrong, and choosing only B or only C leaves out half of its role.

What was Halicin originally developed for before being repurposed as an antibiotic?

ACancer treatment
BDiabetesAnswer
CCardiovascular disease
DAlzheimer's disease

Why

Halicin was first studied as a possible diabetes drug. AI later predicted its antibacterial activity, so it was repurposed as an antibiotic.

iii

Which database was used to screen molecules when discovering Halicin?

APubChem
BDrugBank
CZINC15Answer
DChemSpider

Why

The AI model examined over 100 million molecules in the ZINC15 database, a free online library of commercially available compounds. PubChem, DrugBank and ChemSpider are other chemical databases but were not the one named in this case.

What advantage does AI offer in the early stages of drug discovery?

AIncreased manufacturing speed
BReduction in the need for clinical trials
CCost-effective narrowing down of potential candidatesAnswer
DEnsuring drug approval from regulatory bodies

Why

In the early stages AI screens huge libraries and narrows millions of molecules down to a few worth testing, which saves time and money. It does not affect manufacturing, cannot replace clinical trials, and cannot guarantee approval.

What was the original purpose of the compound Halicin before its antibiotic properties were discovered?

AAn anti-inflammatory agent
BA diabetes drugAnswer
CA painkiller
DA cholesterol-lowering drug

Why

Halicin was originally developed as a diabetes drug (the same fact as MCQ ii). Its antibacterial activity was found later by an AI model at MIT.

Short answer questionsp. 228

i3 marks

What is the significance of AI in predicting the biological activity and toxicity of compounds?

Solution

Machine learning models learn from data on thousands of compounds whose activity and toxicity are already known. By comparing a new molecular structure with this data, they forecast how the substance will behave in the body (its biological activity) and whether it is likely to be toxic, before it is made.

Significance: (1) dangerous substances are identified at an early stage, so time and money are not wasted on them; (2) the need for extensive laboratory and animal testing is reduced; (3) the models help explain how new medicines work and suggest improvements. Example: the MIT model predicted that halicin would kill bacteria and have low toxicity to human cells, and lab tests confirmed it.

ii3 marks

Explain how AI models help in the optimization of lead compounds during drug discovery.

Solution

A lead compound is a molecule that already shows the desired biological activity but needs improvement, for example because it is weak, toxic or unstable in the body. Chemists improve it by changing its structure (adding, removing or swapping groups).

AI models predict how each change in structure will affect the molecule's performance (potency), stability and safety before it is synthesised. They recommend the most promising changes, so fewer molecules have to be made and tested. This speeds up the repeated cycle of design, synthesis and testing, and gives a better drug candidate in less time and at lower cost.

iii3 marks

Describe the process of virtual screening and its importance in the context of AI-driven drug discovery.

Solution

Virtual screening is the computer-based testing of very large libraries of chemical compounds to find those most likely to interact with a particular biological target (for example an enzyme or a bacterial membrane).

Process: (1) an AI model is trained on known active and inactive molecules; (2) the structures of millions of molecules from a database are fed into the model; (3) the model scores and ranks every molecule; (4) only the top-ranked few are bought or synthesised and tested in the laboratory.

Importance: it narrows the pool from millions to a manageable number, which speeds up the early stages of drug development and lowers costs. For halicin, over 100 million molecules of the ZINC15 database were screened in days.

iv3 marks

What role did the deep learning model play in the discovery of Halicin as an antibiotic?

Solution

MIT researchers trained a deep learning model on the molecular structures and measured antibacterial activity of about 2300 compounds. The model learned which structural features stop the growth of bacteria.

The trained model then screened thousands of known drug molecules and ranked halicin, originally a diabetes drug, as a likely antibiotic, although its structure is unlike known antibiotics. It also predicted low toxicity and an ability to disrupt bacterial cell membranes. Laboratory tests confirmed that halicin kills many antibiotic-resistant bacteria. So the model identified the candidate and predicted its activity; experiment confirmed it.

v3 marks

Why is MIT notable in the field of AI and drug discovery?

Solution

The Massachusetts Institute of Technology (MIT), Cambridge, USA, is notable because its researchers made the first well-known discovery of an antibiotic by deep learning. In 2020 they trained a model on the structures and bioactivity of thousands of compounds and used it to identify halicin, an old diabetes drug, as a powerful new antibiotic.

Their model screened over 100 million molecules, and laboratory tests confirmed that halicin kills many antibiotic-resistant bacteria. This showed that AI can find drugs that human experts miss and can speed up the fight against antibiotic resistance. MIT groups have since used the same approach to find other candidate antibiotics.

Long answer questionsp. 228

i7 marks

Discuss the various stages of drug discovery where AI tools are utilized, providing examples for each stage. How does AI improve the efficiency and effectiveness of these stages?

Solution

Introduction

Drug discovery is the long process of finding a new medicine and proving it safe and effective. Traditionally it takes 10 to 15 years and costs over a billion US dollars, and most candidates fail. Artificial intelligence (AI), especially machine learning, now helps at almost every stage.

Stages of drug discovery and the role of AI.

Stages where AI is used (with examples)

Target identification: AI analyses vast biological data to find the protein or process a drug should act on. Example: AlphaFold predicts the 3D shapes of proteins, so drugs can be designed to fit them.
Hit discovery (data-driven discovery and virtual screening): AI models trained on known compounds scan huge libraries and pick molecules likely to be active. Example: the MIT model screened over 100 million molecules in the ZINC15 database and found halicin.
Molecular generation: generative models such as variational autoencoders (VAE) and generative adversarial networks (GAN) design completely new molecules with the desired properties.
Lead optimisation: AI predicts how changes in a lead compound's structure will affect its potency, stability and safety, and recommends the best changes.
Preclinical testing: machine learning predicts biological activity and toxicity, so harmful molecules are dropped early and fewer animal tests are needed. Example: halicin was predicted to have low toxicity to human cells.
Clinical trials and after approval: AI helps select suitable patients, design trials and monitor reports of side effects.

How AI improves efficiency and effectiveness

Speed: screening that would take years in a laboratory takes days on a computer.
Cost-effectiveness: narrowing millions of candidates to a few reduces the cost of experimental testing.
Innovation: AI explores new chemical structures that chemists might not consider, leading to new classes of drugs (halicin is unlike any older antibiotic).
Better success rate: predicting toxicity and activity early means fewer late, expensive failures.

ii7 marks

Examine the case study of Halicin to illustrate the broader implications of AI in drug repurposing. What were the key steps involved, and how did AI contribute to each step? What does this case study suggest about the future of AI in medicinal chemistry?

Solution

Background

Drug repurposing means finding a new medical use for a compound that was developed for something else. Halicin was originally developed as a diabetes drug. In 2020 researchers at MIT used artificial intelligence to show that it is a powerful antibiotic.

halicin \ce{C5H3N5O2S3}

Key steps and the contribution of AI

Step 1, collecting data and training: about 2300 molecules were tested for their ability to stop the growth of E. coli. A deep learning model was trained on their structures and activity, so it learned which structural features give antibacterial action.
Step 2, screening known drugs: the trained model scored thousands of molecules already studied as drugs. It ranked halicin highly even though its structure does not resemble any known antibiotic, something a human chemist would probably have missed.
Step 3, virtual screening on a large scale: the model examined over 100 million molecules in the ZINC15 database within days, showing how AI can search chemical space far beyond laboratory capacity.
Step 4, predicting action and safety: the AI predicted low toxicity to human cells and an ability to disrupt bacterial cell membranes (the electrochemical gradient the cell needs for energy).
Step 5, laboratory confirmation: tests confirmed that halicin kills many antibiotic-resistant bacteria, such as Acinetobacter baumannii and Clostridioides difficile, and that bacteria did not easily become resistant to it.

Broader implications for drug repurposing

Because repurposed compounds have already been made and partly studied, AI screening of existing drug libraries can quickly turn old molecules into new treatments, saving much of the time and cost of developing a new molecule from zero. It is especially valuable against antibiotic resistance, where new drugs are urgently needed.

Future of AI in medicinal chemistry

AI will be used routinely to identify and optimise drug candidates, making discovery faster, cheaper and more innovative.
Generative models will design entirely new classes of molecules, and AI tools will also plan how to synthesise them.
Scientists will still be needed: AI predictions must be confirmed by laboratory tests and clinical trials, and good results need large, high-quality data.

Exam focus

The five ways AI is used in pharmaceutical research (data-driven discovery, virtual screening, molecular generation, lead optimisation, prediction of activity and toxicity). (long)
Book section 14.1 in full; every short question of the exercise is one of these points.
Go to answer
Case study of halicin: key steps and the role of AI. (long)
Book long question (ii) and three MCQs come from this case.
Go to answer
Stages of drug discovery where AI is used, with examples. (long)
Book long question (i); the flow chart earns marks.
Go to answer
What is virtual screening? Why is it important? (short)
Definition plus the halicin example (ZINC15, over 100 million molecules) is a likely 3-mark question.
Go to answer
Advantages of AI in drug development: speed, cost-effectiveness, innovation. (short)
Book section 14.3 and MCQ iv; easy marks.
Go to answer
Role of AI in lead optimisation and in predicting toxicity. (short)
Book short questions (i) and (ii).
Go to answer
Retrosynthesis: deduce the halogenoalkane that gives a given alcohol. (short)
Asked twice in 2026 (Set B Q.2 (vii), Set D Q.2 (viii) OR; mapped to unit 9).
Go to answer
Reagents and conditions for key interconversions; lengthening a chain with cyanide. (short)
Route questions from units 8 to 13 need this reaction map.
Go to answer