predict stereochemistry of reaction product

Predicting Product Stereochemistry: Where Chemists and AI Go Wrong

How to predict stereochemistry of reaction products: seven common errors, from SN2 vs SN1 to chelation control and stereo lost in SMILES files.

ChemStitchNovember 22, 2026

Give a student (S)-2-bromopropanoate with silver oxide and most will draw (R)-lactic acid, by backside attack. The flask gives (S)-lactic acid, because the neighbouring carboxylate attacks first. When you predict the stereochemistry of a reaction product, most errors start in that gap between the mechanism you assumed and the one that ran. Here are seven, each with an example and a check.

Mistake 1: Assuming SN2 inversion when the mechanism is something else

The problem. Inversion is the reflex answer for substitution at sp3 carbon. It holds only for a clean bimolecular displacement.

The example. Tertiary and benzylic substrates in polar protic solvents ionize first (SN1). The planar cation gives a largely racemized product, often with slight excess inversion because the departing anion still shields one face.

In the bromopropanoate case, the carboxylate displaces bromide intramolecularly to form an α-lactone (inversion). Water or hydroxide then opens it (a second inversion), so the net result is retention. Concentrated hydroxide outruns the carboxylate and gives plain SN2 inversion.

The fix. Look for a tertiary, benzylic or allylic carbon in an ionizing solvent. Then look for a lone-pair donor (carboxylate, ester carbonyl, sulfide, amine, aryl ring) that can reach the back of the reacting carbon. If one can, count the inversions.

Mistake 2: Mixing up syn and anti addition

The problem. When an alkene gains two stereocentres, their relative configuration depends on whether both new bonds form on the same face.

The example. The 2-butenes show it cleanly:

ReagentModecis-2-butene givestrans-2-butene gives
OsO4 (cat.), NMOsynmeso-2,3-butanediolracemic 2,3-butanediol
mCPBAsyncis-2,3-dimethyloxirane (meso)trans-2,3-dimethyloxirane (racemic)
Br2 (via bromonium ion)antiracemic 2,3-dibromobutanemeso-2,3-dibromobutane

Hydroboration–oxidation is also syn, and the C–B to C–O step keeps configuration. So 1-methylcyclohexene gives trans-2-methylcyclohexanol, as the racemate.

The fix. Take the mode from the mechanism: a concerted or four-centre step is syn, and a bridged cation opened from the back is anti. Then check for meso: adding two identical groups anti across a symmetric trans alkene, or syn across a symmetric cis alkene, must give one.

Mistake 3: Ignoring the geometry and E/Z ratio of the starting alkene

The problem. A stereospecific reaction copies the alkene geometry into the product. It also copies the ratio, so a 4:1 E/Z starting material gives a 4:1 mixture of diastereomers.

The example. mCPBA converts (E)-stilbene to trans-stilbene oxide and (Z)-stilbene to cis-stilbene oxide. Simmons–Smith cyclopropanation keeps geometry the same way. Wittig products and many trisubstituted alkenes arrive as E/Z mixtures.

A SMILES written CC=CC has no / or \ bond marks, so its geometry is undefined. Anyone, or any model, predicting from it is guessing.

The fix. Confirm the starting material's geometry and ratio from its 1H NMR: the vinyl 3J coupling for 1,2-disubstituted alkenes, NOE for trisubstituted ones. Write it into the structure: C/C=C/C is the E isomer.

Mistake 4: Drawing one enantiomer from achiral or racemic inputs

The problem. A new stereocentre formed with no chiral influence comes out racemic. A wedge on that product claims something the flask did not do.

The example. NaBH4 reduction of acetophenone gives racemic 1-phenylethanol. Enantioenrichment needs a chiral reagent, catalyst or auxiliary: a CBS reduction, a Noyori transfer hydrogenation, a ketoreductase. A racemic starting material with an achiral reagent also gives a racemic product, even when the relative configuration is fully defined.

The fix. Draw racemates with plain bonds, or label the wedges rac or rel. In a MOL V3000 file, enhanced stereo can record this. An AND group means a racemic mixture, an OR group means one enantiomer of unknown identity, and ABS means absolute. For the drawing side, see our guide to wedge and hash bond conventions.

Mistake 5: Treating a change in R/S as inversion, or no change as retention

The problem. R and S are labels from the Cahn–Ingold–Prelog priority rules. They follow priorities, not geometry.

The example. L-serine is (S) and L-cysteine is (R), yet both have the same spatial arrangement. CH2SH outranks CO2H; CH2OH does not. Convert serine's CH2OH to CH2SH and the label flips from S to R without the stereocentre being touched.

The reverse also happens. An SN2 inversion keeps its descriptor when the incoming group ranks differently from the group it replaced.

The fix. Track geometry by face and wedge through the scheme. Assign descriptors last, recomputed on the product.

Mistake 6: Applying Felkin–Anh when the substrate chelates

The problem. For nucleophilic addition to a carbonyl with an α-stereocentre, the Felkin–Anh model and the chelation model predict opposite diastereomers at the new carbinol carbon.

The example. Add a Grignard reagent to an α-benzyloxy ketone. Magnesium bridges the carbonyl and OBn oxygens in a five-membered chelate, and the nucleophile attacks its less hindered face. The result is the anti-Felkin (chelation) product.

Swap OBn for a bulky silyl ether such as TBDPS or TIPS and chelation fails. The Felkin model, in its Anh polar form with the C–O bond held perpendicular to the carbonyl, then sets the face.

The fix. Ask three questions. Is there an α-heteroatom that can bind a metal? Is the metal chelating (Mg, Zn, Ti) and the solvent weakly coordinating? Is the protecting group small (Bn, MOM, PMB) or a bulky silyl group?

Mistake 7: Losing the stereochemistry in the file, not the flask

The problem. A correct prediction saved in the wrong format loses its stereochemistry.

The example. C[C@H](O)c1ccccc1 is one enantiomer of 1-phenylethanol. CC(O)c1ccccc1 is the same connectivity with the stereocentre undefined.

A MOL V2000 file has one chiral flag for the whole molecule. It cannot say "this centre is absolute, that one is only relative", which V3000 enhanced stereo can. More detail is in what V3000 adds over V2000.

The fix. Paste the product SMILES into SMILES to Structure and check that the wedges you expect appear on the canvas. It computes canonical SMILES, InChI and InChIKey with RDKit, and warns before a V2000 export would lose stereochemistry, offering V3000 instead.

Tip In a standard InChIKey, the second block encodes stereo and isotope layers. If a product that should have a defined stereocentre has a key whose second block is UHFFFAOYSA, no stereochemistry made it into the string.

A five-question check to predict the stereochemistry of a reaction product

  1. Which atoms are new or changed stereocentres? Mark them first.
  2. How does each bond form: concerted on one face, by backside attack, through a planar intermediate, or through a bridged or chelated one?
  3. What stereochemical information goes in: alkene geometry, existing centres, a chiral catalyst? If none, the product is racemic.
  4. For additions next to an existing centre, which model sets the face: steric (Felkin–Anh) or chelation?
  5. Does the saved file still say what you drew? Check for isomeric SMILES, V3000 where needed, and the InChIKey second block.

Run the same check on AI output; ChemStitch labels its chat-based reaction prediction AI-suggested for this reason. A holistic benchmark of reaction-prediction models flagged stereoselectivity as one of the most sensitive areas for current models. Our post on where AI reaction prediction falls short covers the other failure modes.

Before a predicted structure goes into a notebook or registration system, paste it into SMILES to Structure and confirm the stereocentres survived. It runs in the browser and needs no account.

Try ChemStitch

AI-powered chemical structure editor. Free 14-day trial.

Start free trial →