Small molecule assessment

What it doesCopy Anchor Link

Judges the structure the molecule tab is drawing: is it inside the property windows medicinal chemists use to call a compound drug-like, lead-like or fragment-like, and does it carry a substructure known to mislead assays or to be reactive. Every verdict names the bound it turned on and this molecule's value, and every alert can be drawn on the structure.

When you'd use thisCopy Anchor Link

When a hit list, a vendor SDF or a SMILES from a supplementary table needs a first pass before anyone orders it: a compound that breaks Lipinski on mass alone, or matches a PAINS pattern, is worth knowing about before it costs an assay. It is also the quickest way to see why two compounds of similar size are treated differently: one fails Veber on rotatable bonds, the other on polar surface area.

What this computesCopy Anchor Link

Drug-likeness rules — twenty published property windows, each shown as holding or failing, and a failing one lists every bound it broke with this molecule's value beside it: Lipinski's rule of five, beyond rule of five (Doak), rule of four for protein–protein interaction inhibitors, rule of three and extended rule of three for fragments, rule of two for building blocks, the Ghose, Veber and Egan filters, REOS, ChemAxon drug-likeness, Pfizer 3/75, GSK 4/400, Oprea lead-likeness, Xu's drug-like index, CNS penetration and respiratory (inhaled) drug space. The rules are reproduced as the medchem library implements them, and checked against it molecule for molecule.

The papers' own readings of Lipinski, Ghose and REOS — three more windows, marked (paper), because the common implementation departs from what each paper wrote. Lipinski's "molecular weight" is the average mass, not the monoisotopic one; Ghose's "total number of atoms" counts every hydrogen; REOS bounds the net formal charge to −2…+2. Each is shown beside the version it re-reads and says whether the two agree on this molecule. On the first 5,000 NCI compounds the two Ghose readings disagree on 2,089, REOS on 18, Lipinski on none (the mass difference reaches a verdict only for a heavily halogenated compound near 500 Da).

Structural alerts — thirteen catalogues, each a switch: PAINS A, B and C (pan-assay interference compounds, Baell & Holloway 2010), Brenk (reactive, unstable or toxic groups, Brenk 2008), NIH (problematic functional groups, Jadhav 2009 and Doveston 2014), ZINC (unwanted groups and size limits from the ZINC database's filtering), and the seven rule sets ChEMBL curated from industry and screening-centre exclusion lists: Glaxo, Dundee, BMS, MLSMR, Inpharmatica, LINT and SureChEMBL. They are matched the way RDKit's FilterCatalog matches them, and a pattern allowed a number of occurrences (ZINC's "more than 40 heavy atoms", for instance) fires only past that count. The catalogues overlap: the NIH set as RDKit ships it carries the same 180 patterns as ChEMBL's BMS set, so a molecule that trips one trips the other. Each alert names its pattern, how many atoms it covers, and can be drawn on the structure. An absent alert is not a clean bill of health: these catalogues cover known nuisance patterns only.

Quantitative estimate of drug-likeness (QED) (Bickerton et al. 2012): eight properties (average molecular weight, Crippen logP, hydrogen-bond acceptors by QED's own eleven patterns, hydrogen-bond donors, polar surface area, rotatable bonds under the strict definition, aromatic rings as QED counts them, and unwanted functional groups from QED's list of 116), each turned into a desirability between 0 and 1 by an asymmetric double-sigmoid curve fitted to 771 orally dosed approved drugs, and combined as a weighted geometric mean. All three of the paper's weightings are shown: the mean weights (the mean of its 1,000 most informative weight sets), the maximum-information weights and unweighted. Behind the score: every input's value, desirability and weight, the input dragging the score down most, and the unwanted groups that matched. Checked against RDKit's QED on every molecule of the NCI and ChEMBL approved-drug sets; the paper's own logP came from a different program, so its published scores differ slightly.

Formula and isotope pattern: the molecular formula as RDKit writes it (every hydrogen counted, the net charge as a suffix), the monoisotopic and average masses, and the natural-abundance isotope envelope: the share of molecules at the monoisotopic mass (M) and at each nominal mass above it (M+1, M+2 …), the pattern a mass spectrometer that does not resolve fine structure records. The envelope uses the same coarse isotope model as the peptide mass tools, and is withheld for an isotope-labelled structure, whose spectrum natural abundance does not describe. The formula and monoisotopic mass are checked against PubChem's published values for the bundled compounds, and the envelope against OpenMS on every molecule of the NCI and ChEMBL approved-drug sets.

Synthetic accessibility (Ertl & Schuffenhauer 2009): a score from 1 (easy to make) to 10 (very hard), built from how common each of the molecule's fragments is among one million representative PubChem molecules, less a complexity penalty for size, stereocentres, spiro atoms, bridgeheads and macrocycles, plus a correction for symmetric molecules; every term is shown, with how many fragments the table did not know. Natural-product likeness (Ertl, Roggo & Schuffenhauer 2008): a Bayesian score from about -5 (like synthetic molecules) to 5 (like natural products), with the coverage (the share of the molecule's fragments the model has seen), and a warning when that falls below half. Both are checked against RDKit's reference scripts on every molecule of the NCI and ChEMBL approved-drug sets.

Scaffold (Bemis & Murcko 1996): the molecule's ring systems and the linkers joining them, every side chain cut away (an atom double-bonded to the scaffold, such as a ring carbonyl oxygen, stays), shown as SMILES and drawn on the structure; and the generic framework, in which every atom is carbon and every bond single, so analogues differing only in heteroatoms or bond orders share it. A molecule without a ring has no scaffold, and one with an atom carrying more bonds than carbon can take has no generic framework. Checked against RDKit's Murcko scaffold on every molecule of the NCI and ChEMBL approved-drug sets.

The properties the rules read, each under the definition it uses: monoisotopic mass, cLogP and molar refractivity by the Crippen atom-contribution model, TPSA counting nitrogen and oxygen only, hydrogen-bond donors and acceptors in both RDKit's and Lipinski's simpler definition, rotatable bonds, rings, heavy atoms and atoms in the graph. Behind them, every RDKit descriptor under RDKit's own name.

BOILED-Egg (Daina & Zoete 2016): the molecule plotted on WLOGP against tPSA with the paper's two published ellipses, the white for passive human gastrointestinal absorption (HIA) and the yolk for blood–brain barrier (BBB) penetration, and a verdict for each. tPSA here counts sulfur and phosphorus, as the paper requires, and logP is RDKit's Wildman–Crippen logP, standing in for the paper's own WLOGP implementation, which was not published. Behind the verdicts: each ellipse's summed distance to its two foci in the paper's normalised space against its major axis, and the scaling that maps tPSA and logP into that space. The ellipses are the published ones, checked against the paper's own ellipse traces point for point and against its classification of every one of its 920 training molecules.

Atom contributions: the depiction shaded by a per-atom quantity, each the descriptor of the same name broken down by atom, as in Riniker & Landrum's similarity maps: Wildman–Crippen logP per atom, Ertl polar surface per atom with nitrogen and oxygen only or with sulfur and phosphorus too, and Gasteiger–Marsili partial charge. The panel gives the largest per-atom magnitude and the sum over the heavy atoms; circle size is the atom's share of the largest, colour its sign.

Bioavailability radar: SwissADME's six properties for oral bioavailability (lipophilicity, size, polarity, solubility, saturation and flexibility), each drawn against its published window. A molecule is drug-like by the radar only when every axis is inside. The lipophilicity upper bound is a choice between the two values the SwissADME paper gives (5.0 in its figure, 6.0 in its methods). The same radar can draw any rule's own property windows instead, so you can see which bound a failing rule broke and by how much.

Aqueous solubility (ESOL): Delaney's estimate of the log molar solubility (log S) from logP, molecular weight, rotatable bonds and the aromatic proportion of heavy atoms, under two coefficient sets: Delaney's own 2004 coefficients (fitted to Daylight CLOGP, read here with Wildman–Crippen logP) and a refit of the same four terms to RDKit's descriptors. Each is given as log S, mol/L and mg/mL, with the four inputs and both equations. On Delaney's 1,128 measured compounds the paper's own predictions are off by 0.91 log units (root mean square), its coefficients with Crippen logP by 1.10, and the refit by 1.01 on the compounds it was fitted to.

Protonation states at a pH (Dimorphite-DL, Ropp et al. 2019): every ionisable group the method recognises (41 substructures, from carboxylic acids, phenols and sulfonamides to amines, amidines and aromatic nitrogens), each with a pKa mean and spread fitted to measured pKa values. Set a pH window (default 6.4–8.4, physiological 7.4 ± 1) and a precision (how many standard deviations widen each group's pKa window): a group whose window overlaps the pH range is kept in both forms, one above it stays protonated, one below it deprotonated, and every combination is listed as a SMILES. Each group is a chip that draws its ionisable atom on the structure, with its pKa, its state and how many times the pattern matched that atom. The list stops at a cap on the number of states (default 128), keeping the most probable states at the middle of the pH window, and says so when it does. The results reproduce Dimorphite-DL 2.1.0 state for state.

Tautomers: every tautomer the structure has under RDKit's 37 hydrogen-shift transforms ((thio)keto–enol 1,3 and 1,5, imines, heteroatom shifts from 1,3 to 1,11 across aromatic rings, furanones, ketene–ynol, nitro–aci-nitro, oxime–nitroso, cyano–isocyanic acid, isocyanides and phosphonic acids, after MolVS and Sitzmann et al. 2010), and the canonical tautomer RDKit's scoring picks: aromatic rings first (100 each, 250 for an all-carbon ring), then twelve weighted groups (benzoquinone, oxime, C=O, N=O, P=O, C=heteroatom, guanidine, aci-nitro, methyl and others), then fewer hydrogens on phosphorus, sulfur, selenium and tellurium; a tie goes to the smaller SMILES. The reading says whether the structure as drawn is already the canonical one, lists every tautomer, breaks the winner's score into its parts, and draws the atoms that trade hydrogens on the structure. Two caps stop a large molecule (1,000 tautomers or 1,000 transform applications, RDKit's defaults), and the reading says when one did. The canonical tautomer and every tautomer equal RDKit's own on thousands of drugs and screening compounds.

Similarity to another structure: paste a second molecule and see how alike the two are under six fingerprints (Morgan / ECFP4, atom pairs, topological torsions, RDKit path, pattern and MACCS keys) as a Tanimoto coefficient each, with the bits set in each and in common. On the fingerprint you choose, every metric RDKit's DataStructs offers: Tanimoto (Jaccard), Dice (Sørensen), cosine (Ochiai), Sokal–Sneath, Russell–Rao, Rogot–Goldberg, all-bit, Kulczynski, McConnaughey, asymmetric, Braun-Blanquet, on-bit and Tversky with both of its weights adjustable. Also count Tanimoto, Dice and Tversky on Morgan environment counts, where a repeated group counts each time it appears. Every value equals RDKit's own on thousands of molecules. The maximum common substructure of the pair (RDKit FindMCS: shared atoms, shared bonds and its SMARTS) can be drawn on the structure on stage, and so can each Morgan environment of it: every ECFP4 identifier with its radius, centre atom and how often each molecule carries it, drawn as RDKit's DrawMorganBit draws it.

Substructure search: which records of the open file contain a SMARTS pattern, found with RDKit's substructure library (a pattern-fingerprint screen, then atom-by-atom matching). The pattern is the same one the molecule tab highlights, so the record on screen shows its matches.

Clustering and diversity picking: every record of the open file grouped by 1 − Tanimoto on Morgan fingerprints, by Butina sphere exclusion (a distance threshold, optionally reordering after each cluster), MaxMin diversity picking (a number to pick and the seed of the first pick), or average linkage (UPGMA) cut into a chosen number of clusters. Each cluster is titled by its centroid and lists its members; a chart gives the cluster sizes and the reading counts singletons. Butina and MaxMin equal RDKit's own for the same settings and seed; average linkage equals scipy's wherever no two distances tie.

Running a reaction over the file: a reaction template applied to one list of structures per reactant, every combination, by RDKit's reaction engine. Pick one of 94 library reactions: the 58 robust organic reactions of Hartenfeller et al. (2011) (amide couplings such as Schotten–Baumann, Suzuki and Stille couplings, Mitsunobu reactions, heterocycle syntheses such as benzimidazoles, pyrazoles, Fischer indoles and Pictet–Spengler) and 36 click-chemistry reactions (azide–alkyne cycloadditions, epoxide openings, isocyanate couplings and more). Or type your own reaction SMARTS. Each reactant comes from every structure in the file, the structure on screen, or structures you type, and the run stops at a combination cap (default 2,500), and says so when it does. The reading counts the products, the combinations that matched no template (a structure without the functional group it needed), the product sets dropped because a product could not be sanitised, and any reactant that could not be read. Each product is listed with the reactants that made it, and Save the products and draw them as a grid writes them into the project and opens them as a molecule tab in its grid of structures, each named by those reactants. The reaction itself is drawn coloured by reactant: each reactant's atoms in its own colour, and each product atom carrying an atom map in the colour of the reactant it comes from. Every product equals RDKit's own on all 94 reactions run over hundreds of thousands of ZINC reactants. The bundled 6-apa-acylation.smi is the semi-synthetic penicillins' chemistry: run the default amide coupling on it and 6-aminopenicillanic acid acylated by each side-chain acid gives benzylpenicillin, phenoxymethylpenicillin, ampicillin, amoxicillin, methicillin, oxacillin, cloxacillin, dicloxacillin and nafcillin, each identical to PubChem's structure.

Mapping the atoms of a reaction: a reaction drawn without atom maps (a Rhea or RXN file, or one you type) gets them from RXNMapper (Schwaller et al. 2021), a transformer trained without supervision whose attention says which reactant atom each product atom came from. Map atoms writes the maps onto the drawing and turns on colouring by reactant; a reaction that already carries maps keeps them. Each mapping carries RXNMapper's confidence. The maps equal RXNMapper's own on all 1,851 reactions of the curated Golden set (Lin et al. 2022); in the bundled Rhea ATP hydrolysis, water's oxygen lands on the released phosphate. Once a reaction carries maps, Export writes it as an atom-mapped reaction SMILES and as an atom-mapped MDL rxnfile, each atom's map in its mapping column and the reaction laid out afresh in 2D; before then it writes the plain reaction SMILES. Every map RXNMapper assigns on the 1,851 Golden-set reactions survives into the rxnfile.

Conformers and MMFF94: 3D conformers of the structure on screen, built from the torsion angles observed in crystal structures (OpenChemLib's generator), each scored by the MMFF94 force field (Halgren 1996) and, if you ask, energy-minimised. Choose how many, the torsion strategy (likely torsions at random, adaptive random, likely torsions in order, or purely random), the random seed, how many torsion sets to try, 60° torsion steps instead of the library's angles, the force-field variant (MMFF94, MMFF94s, Halgren's 1999 variant with planar delocalised nitrogens, or MMFF94s+), and the minimiser's iteration cap and gradient and energy tolerances. The conformers are listed lowest energy first with how far each lies above the lowest (kcal/mol), and for a file that carries its own 3D coordinates the MMFF94 energy of that geometry is given beside them. Pressing a conformer draws it in the 3D view, so a SMILES or a 2D drawing gets a 3D structure too. The force field equals Halgren's own reference program (OPTIMOL) to 0.01 kcal/mol on every one of the 761 structures of the MMFF94 validation suite. A conformer generator has no single right answer, so its quality is measured instead: the best of 10 conformers comes within 1 Å heavy-atom RMSD of the crystal structure for 561 of the suite's 759 molecules it can build (median 0.42 Å).

A docked pose in its pocket: a docking pose (an AutoDock Vina PDBQT) is drawn in 3D with the receptor it was docked into, picked from the project's structure files: the protein's fold as a cartoon and its binding pocket, every residue with a heavy atom within an adjustable pocket distance of the ligand, as sticks and listed by name. The bundled imatinib poses in Abl kinase (PDB 2HYY) show the gatekeeper Thr315 and the hinge Met318 at 5 Å. The pocket is checked residue for residue against an independent reading of the PDB files on every ligand of the bundled crystal structures.

Draw and edit: a structure editor opens on the current structure (a 3D record is laid out flat first). Every change is read back by the same chemistry engine the rest of the panel uses, which shows the canonical SMILES of exactly what will be saved and says why when a drawing is not a valid molecule (a carbon with five bonds, say). Saving adds the drawing to the project as a molfile and opens it, so every readout on this page applies to it. A structure passes through the editor unchanged, stereocentres included, for every structure of the bundled chemistry files.

Peptides from HELM: a HELM file (one peptide per line, the notation ChEMBL and Pistoia use for biologics) opens as one structure per peptide, built residue by residue the way RDKit builds it: natural and D-amino acids, common non-natural residues (Nle, Sar, Orn, Glp and others), N-terminal acetyl and C-terminal amide caps, disulfide bridges and head-to-tail cyclisation. Each structure then gets every readout above. A peptide using a monomer outside that set, or longer than 100 residues (an antibody chain), is kept with the reason it was not drawn. Construction matches RDKit atom for atom on 503 HELM strings, the 34 approved peptide drugs of the bundled ChEMBL file among them (20 drawn: oxytocin, vasopressin, exenatide and lanreotide among them).

Library grid: every record of a file screened at once, one row each, by the rules and alert sets chosen above, with the properties those rules read. A record RDKit cannot read stays as a row saying so, so nothing shifts. Each row equals what the panel reports for that record alone. The structure you had open, the grid's sort and filter, and the data column chosen as the activity come back when the project is reopened.

How the records are distributed: the file's records binned by one rule property (mass, cLogP, TPSA, donors, acceptors, rotatable bonds and the rest), each bar split at the bound of a rule you pick (Lipinski's 500 Da, the rule of three's 300 Da, Veber's 140 Ų and so on), with how many records fall within it. The bins count the value the rule itself judged, so a split agrees with the grid's pass and fail. Beside it, molecular weight against cLogP with the Lipinski, rule of three and beyond rule of five windows drawn, and how many records each holds.

Bioactivity of a protein's ligands: a protein structure (PDB or mmCIF) names the UniProt accession of each chain from its own database references. From there the panel finds the chain's ChEMBL target and reads its activities: every measured compound with its pChEMBL value, assay type and source document, written into the project as a structure file that every readout above can then screen. The references equal Biopython's on 23 entries.

Map this file (chemical space): every structure of the file placed by principal component analysis, computed exactly as R's prcomp computes it, on either the scaled property table or Morgan radius-2 fingerprint bits. The axes carry each component's share of the variance, the loadings say what each axis is made of, and a property that is constant across the file is named and left out rather than dividing by zero. Beside it, a similarity heat map: Tanimoto on Morgan bits for every pair of the first 60 structures, ordered by the leaves of the average-linkage tree (the structures beyond 60 are counted, not silently dropped), and that tree can be opened as a Newick tree. Pressing a point or a cell opens its structure. The components equal R's on every molecule of 4,991 NCI compounds and 3,417 approved drugs. t-SNE and UMAP are not offered: both are stochastic and have no reference answer to check against.

Series in this file: the structure–activity view of a compound series.

• Scaffold network (Kruger et al. 2020), computed as RDKit's CreateScaffoldNetwork computes it, under either RDKit's default ring-side cuts or the 46 BRICS bond types (Degen et al. 2008). The ring scaffolds (no attachment point) are listed by how many structures contain each and how many times the network made it, with the count of every edge type (initialise, fragment, generic, remove attachment, generic bond). Pressing a scaffold draws what it breaks into as a levelled graph. A structure with more cuttable bonds than the limit (20 ring-side, 15 BRICS) is named and left out, because its network doubles with every bond. The networks equal RDKit's node for node and edge for edge on all 4,991 NCI compounds and 3,415 of 3,417 approved drugs under ring-side cuts, and on 4,989 and 3,374 under BRICS; every molecule left out is one over the limit.
• R-group decomposition and SAR matrix: the file decomposed on a core (a pressed scaffold, or SMARTS you type) by RDKit's RGroupDecomposition, reporting how many structures carry the core and its substitution sites R1, R2 …. The SAR matrix crosses any two sites, each cell holding the number of structures carrying both substituents or their mean activity. The core can be drawn on the structure on stage.
• Matched molecular pairs (Hussain & Rea 2010), as mmpdb finds them: pairs of structures that differ by one change no larger than a heavy-atom limit, grouped by the change (as a reaction SMIRKS), most frequent first.
• Activity cliffs (SALI) (Guha & Van Drie 2008): on a numeric data field you choose, as pIC50 from a µM IC50 or as given, every pair scored by the structure–activity landscape index, the activity difference over one minus the Tanimoto similarity on RDKit path fingerprints. Two structures with identical fingerprints and different activities score infinity, as the equation says. The pairs are plotted as similarity against activity difference and ranked steepest first. The bundled moonshot-mpro-ic50.sdf (COVID Moonshot SARS-CoV-2 main protease series, CC0) carries the measured IC50 values to run it on.

Every row, node, cell and point opens its structure on the stage. A pair (a matched pair, or an activity cliff's row or point) opens both: the first on the stage and the second in Similarity to another structure, drawn there with its Tanimoto scores and the maximum common substructure the two share.

One reading, generated from the run: how many rules hold, which fail, and which alert sets matched, with the chain it rests on.

ControlsCopy Anchor Link

Each section's controls, by the names the panel gives them.

One structure

A file of structures

How to read the resultCopy Anchor Link

• Readings are taken with hydrogens implicit, whatever the file wrote and however the tab is drawing it. A database record writes its hydrogens out, and counted that way the rotatable-bond and donor numbers come out higher than the ones the rules' own papers were derived on. A molfile and the same compound's SMILES give the same answer here.
• Mass is monoisotopic in every rule, as medchem uses it, including Lipinski's, whose paper used average molecular weight. A compound within a dalton or two of 500 can read differently in a tool that uses average mass.
• Pressing an alert draws its matched atoms and bonds on the structure. Press it again to clear it. When the tab draws hydrogens as atoms, the highlight is carried onto the same heavy atoms.
• On the egg, inside the white means well absorbed from the gut, inside the yolk means able to reach the brain, both by passive diffusion. The two ellipses were fitted separately, and a corner of the yolk lies outside the white; a molecule there is flagged rather than read as both.
• The egg places the molecule as drawn. The paper places the neutral, unsalted parent, so a record with a counter-ion or a net charge is flagged next to the plot.
• Shading the structure replaces an alert drawn on it, and the other way round: one overlay at a time.
• Pressing a spoke of the radar, or the molecule's point on the egg, finds the value it was drawn from: the descriptor, the solubility estimate or the egg's tPSA and logP, marked where the panel lists it.
• The radar's properties are computed here, not by SwissADME. Lipophilicity is Wildman–Crippen logP rather than XLOGP3, so a molecule near a window's edge can read differently there.
• ESOL is intrinsic solubility from structure alone. pH, salt form and crystal packing are not in it, so a drug given as a salt can be far more soluble than its estimate.
• Similarity depends on the fingerprint. Two molecules can score 0.06 under torsions and 0.32 under pattern bits; a cut-off such as "0.7 is similar" belongs to one fingerprint and one metric. The structure on screen is the query, which matters only for Tversky: its weight a counts what only this structure has, b what only the other has.
• Clustering runs when you press it, because it compares every record with every other. Change a setting or open another file and the panel says the result is stale. Average linkage stops at 500 structures; Butina and MaxMin take the whole file.
• The rule and alert-set lists are knobs. Switching a set off means it was not screened, and the reading says so rather than reporting no alerts.

Where it livesCopy Anchor Link

In the rail beside a molecule tab, and through the assistant, the workflow builder, the HTTP API and MCP as assess_small_molecule, which takes a SMILES string or a molfile and returns the same reading, rules, alerts, descriptors, BOILED-Egg placement and per-atom contributions; compare_molecules scores a pair, enumerate_protonation_states lists a molecule's protonation states at a pH, search_molecule_library searches a list of structures, cluster_molecules clusters or picks from one, enumerate_reaction runs a reaction over lists of reactants and write_structure_library writes a list of structures as SDF or SMILES, and canonical_tautomer gives a molecule's tautomers and the canonical one.

Pitfalls worth knowingCopy Anchor Link

• A rule is a rule of thumb. Many approved drugs fail Lipinski (antibiotics and natural products above all), and kanamycin in the bundled examples holds only one of the twenty windows.

• An alert is not a verdict on activity or toxicity. PAINS patterns were drawn from one screening library's frequent hitters; a match says "check this hit twice", not "discard".

• TPSA will not match a database that counts sulfur and phosphorus. The rules here use the nitrogen-and-oxygen definition they were published with; the BOILED-Egg uses the one that counts both, so ampicillin's two readings differ by its sulfur.

• The egg's logP is not the paper's WLOGP. Both follow Wildman and Crippen's fragments, but the two implementations disagree on most of the paper's own molecules, and on its brain-penetration set the substitution changes the verdict for about one molecule in twenty. A point near an ellipse's edge deserves that caution.

• The paper's P-gp substrate prediction is a separate model and is not drawn on the egg.

• Protonation states are a pattern method, not a pKa calculation. Each group carries one pKa distribution whatever surrounds it, so a nitrogen the patterns read as an amine is treated as basic even where its neighbours make it an amide in all but name (ampicillin's β-lactam nitrogen is one). Charged groups are neutralised before the groups are matched, and a counter-ion or solvent beside a neutralised group is dropped from the result, as the reference program drops it.

• A reaction template is not a reaction prediction. It forms the bond wherever its patterns match, protecting groups and selectivity aside. Run over 6-APA and its side-chain acids, the amide coupling also acylates the amino acids' own amines with 6-APA's carboxylic acid, which is why the real synthesis protects them; those 21 extra products are listed beside the nine drugs.

• Atom mapping is not offered. The one mapper that runs in the browser (Indigo's) maps fewer than four reactions in ten correctly on a manually curated benchmark of 1,851 reactions (Lin et al., 2022), and gets a simple esterification wrong.

• The canonical tautomer is a convention, not the form in water. RDKit's score prefers aromatic rings and carbonyls by rule, which makes the same molecule always come out the same way for comparison and registration; which tautomer dominates in solution depends on solvent and pH, which the score does not see.

RelatedCopy Anchor Link

Draws on the molecule viewer, which shows the structure and its computed properties; this panel adds the judgement.