Science

A binding constant is a concentration: the amount of compound needed to occupy a receptor. Lower means stronger. Because these concentrations span ten orders of magnitude — from picomolar to millimolar — they are compared on a logarithmic scale, which also turns the relationship the right way round, so that a larger number means tighter binding.

The measurements, and why they are not interchangeable

K_i — inhibition constant

The concentration at which a compound occupies half the receptors, corrected for the radioligand competing with it. Derived from IC₅₀ through the Cheng–Prusoff relation below. An equilibrium property of the compound and the receptor, largely independent of assay conditions — which is what makes two K_i values comparable across laboratories.

K_d — dissociation constant

The same equilibrium quantity measured by direct saturation binding rather than competition. K_d = k_off / k_on: the ratio of the rates at which the complex falls apart and forms.

IC₅₀ — half-maximal inhibitory concentration

The concentration producing half the maximum inhibition in that assay. It depends on how much radioligand was used and on that ligand’s own affinity, so it is not portable between experiments.

EC₅₀ — half-maximal effective concentration

A functional readout: the concentration producing half the maximum response, not half the maximum occupancy. Receptor reserve means a compound can produce a full response while occupying a small fraction of receptors, so EC₅₀ can sit far below K_i at the same target.

AC₅₀ — half-maximal activity concentration

The concentration producing half the maximum change in an assay’s readout, whichever direction that change runs — activation or inhibition. It is the quantity reported by high-throughput concentration–response screens, where a curve is fitted to every compound tested and the assay does not say in advance which way a compound will move it. Like IC₅₀ and EC₅₀ it belongs to its assay, not to the compound and the receptor, and it is never read as a binding constant.

Potency

ChEMBL’s term for concentrations from PubChem’s screening assays, where the value is the concentration at which the compound was active. It is a screening result, not a binding experiment designed for that target.

Physicochemical descriptors are computed here with RDKit from the structure PubChem returns, never from a stored local copy. They are marked R throughout, because they are calculated rather than measured.

Molecular weight

Sum of atomic weights, in g/mol. Computed from the molecular formula.

cLogP

Calculated partition coefficient between octanol and water, by the Crippen method: a sum of contributions from each atom in its bonding context. Positive means lipophilic. It is an estimate from structure, not a measurement — an experimental logP is a different number.

TPSA

Topological polar surface area, in Ų. The summed surface contribution of nitrogen and oxygen atoms and their attached hydrogens, from a fragment table. Used as a proxy for membrane permeability; values under roughly 90 Ų are associated with CNS penetration, though that is a correlation, not a rule.

H-bond donors / acceptors

Counts of N–H and O–H groups, and of nitrogen and oxygen atoms respectively, by Lipinski's definitions.

Rotatable bonds

Single, acyclic, non-terminal bonds between heavy atoms — a measure of conformational flexibility.

Fraction sp³

The proportion of carbon atoms that are sp³-hybridised: the count of sp³ carbons over all carbons. A measure of three-dimensionality; flat aromatic molecules score near zero.

Bemis–Murcko scaffolds

A compound’s scaffold is what remains after every side chain is stripped away: the ring systems, and the linkers joining them. Formally, delete all atoms that are not in a ring and not on a path between two rings, and repeat until nothing more can be removed. Compounds reducing to the same scaffold share a framework.

The Families page groups by this rather than by curated class names, because a label like “tryptamine” is an assertion somebody has to stand behind, while a scaffold is a derivation anyone can repeat from the structure. The scaffold SMILES is printed with each family so the grouping can be checked rather than trusted.

Bemis GW, Murcko MA. The properties of known drugs. 1. Molecular frameworks. J Med Chem. 1996;39(15):2887–93.

Binding and effect are separate questions. A compound’s affinity says how readily it occupies a receptor; its efficacy says what the receptor does once occupied. A high-affinity antagonist and a high-affinity agonist can have identical K_i values and opposite consequences, which is why the mechanism records on each dossier carry an action type and the binding data does not stand in for one.

Target classification

Assay target names arrive as free text from ChEMBL and mix three different kinds of thing: named molecular targets (5-HT_2A), whole organisms (Plasmodium falciparum), and cell lines (HepG2). They are sorted by an explicit pattern list, and each classified target carries the pattern that classified it, visible in the interface. A target matching no rule is shown as unclassified rather than assigned a class by guesswork.

Only molecular targets carry an affinity interpretation. A cytotoxicity value against a cell line is a real measurement of something else, and is never placed on a pX axis or drawn in the affinity rosette.

The anatomy plates shade a region by how strongly the gene for a compound’s target is transcribed there, from the Human Protein Atlas consensus dataset.

Transcript is not protein, and protein is not activation. mRNA abundance correlates imperfectly with receptor density, and a receptor present in a region is a receptor available to be bound — not evidence that a compound reaches it, occupies it, or changes what that region does. The plates therefore claim only that: this compound binds this target, and this target’s gene is transcribed here. Shading is never affinity, and never affinity multiplied by expression, because that product is not a quantity anyone measures.

Every quantity defined so far is an equilibrium: a state the system settles into given enough time. Nothing anyone experiences is at equilibrium. A dose arrives, receptors fill, something downstream happens, and hours later it stops — and the three stages run on clocks that differ by six orders of magnitude. This section defines each stage, and says which of them the archive can put a number on.

1 · Binding — what the affinity column leaves out

A compound and a receptor associate at a rate proportional to how much of each is present, and the complex falls apart at a rate that depends on nothing but itself. Affinity is the ratio of those two rates, and a ratio is a quotient of two numbers neither of which it reports.

The quantity that governs how long an occupied receptor stays occupied is the off-rate alone. It has its own half-life, and it is not the drug’s.

2 · Acting — occupancy is not response

An occupied receptor changes shape, couples a transducer, and sets a second messenger moving. Each step takes its own time — milliseconds for the conformational change, seconds to minutes for the messenger — and, more consequentially, each step multiplies. One receptor activates many transducers; one enzyme makes many molecules of product. By the time a response is measured, the signal has been amplified enough that a small occupied fraction can produce a nearly maximal effect.

A further complication the archive records but does not resolve: a receptor can couple more than one transducer, and an agonist can favor one over another. A compound biased toward G-protein coupling and one biased toward β-arrestin recruitment can share a K_i and drive different downstream courses. The mechanism records on each dossier carry an action type where a source states one; where none does, the dossier says so rather than inferring a direction from a binding constant.

3 · Behavioral consequences — why the clocks diverge

The compound is in plasma; the receptors are not. Between the two sits distribution across the blood–brain barrier, and the compartment where the effect is generated fills and empties more slowly than the blood does. The standard treatment adds a hypothetical compartment whose concentration lags plasma by a single rate constant.

Three mechanisms stack in the same direction, and together they are why subjective effects outlast the plasma curve so consistently: the effect compartment empties after plasma does; receptors stay occupied after free compound is cleared, at a rate set by k_offrather than by elimination; and amplification means a falling occupancy still drives a near-maximal response until it drops a long way. In some compounds a further term applies — an active metabolite with its own kinetics, which is why psilocybin and psilocin are filed as two records rather than one.

None of this is used to produce a number. The mechanisms above explain a direction, not a magnitude: no k_e0, no k_off and no free brain concentration is recorded for these compounds, so a duration cannot be calculated here even in principle. Every duration on this site is quoted from a trial that measured it in people, and where no trial has, the dossier shows an absence.

These are different quantities, they diverge sharply, and conflating them is the single easiest way to misread a dossier.

Duration of subjective effects is not derived from half-life. LSD is cleared with a plasma half-life of about 2.6 h while its subjective effects run 8–12 h — a factor of three or four. Effects persist through receptor occupancy and downstream signaling that outlast the plasma curve. Every duration on this site is quoted from a controlled human trial that measured it, never inferred from pharmacokinetics and never borrowed from a chemical relative. Why the two diverge so reliably is set out in section 5.

The analysis page carries two matrices. The first asks which targets a compound binds — one compound per row, one target per column. The second asks how alike two compounds’ whole profiles are — compounds on both axes. Both are built from the same evidence, either the archive’s measured affinities or DRIFT’s structural scores, and the two sources are never combined into one grid.

Both matrices are mostly empty: 95% of the measured grid and 92% of DRIFT’s. Any comparison between two compounds therefore has to decide what an empty cell means, and there are only two coherent answers. Leave it out, and the comparison rests on whatever the pair happens to share. Fill it with a zero, and the comparison spans everything — at the cost of asserting a value nobody measured. This site does both, in different plates, and says which is which.

1 · Leaving the gap out

The drug–target plate orders its axes this way. Only cells where both members carry a value contribute, and nothing is imputed.

The cost is coverage, and the plate reports it rather than hiding it. On DRIFT’s full grid only 12% of target pairs share three values, so most of that axis is ordered by tie-breaking; the page says so in place, next to the ordering it qualifies. The compound axis fares much better at 82%.

2 · Giving the gap a value, in the coordinate where zero means something

The compound–compound plate takes the other option, because a cosine is defined over a whole vector rather than over an intersection. Each compound becomes a vector across every target of its source — all 172 assayed, or all 489 DRIFT ranked — and every cell has to hold a number, including the empty ones. Which number is admissible depends entirely on the coordinate.

The component is 1/K_i, and that is what makes the zero admissible. A reciprocal binding constant tends to zero as K_i tends to infinity, so an empty cell holds the limit of the quantity itself at “does not bind”. Nothing is substituted for a missing value; the coordinate simply has a value there.

Measuring in pX instead — which is what this plate did first — puts a real and absurd number in the same place, because pX 0 is a binding constant of one molar. The consequence was not only cosmetic. Every component was then of the same order, so a target a pair had merely both been tested at contributed as much as one either of them binds, and the cosine largely tracked co-assay rather than pharmacology. In reciprocal space a target bound at 10 µM contributes about a hundred-thousandth of what one bound at 100 pM does. The effect is easy to see on real pairs: LSD against lisuride, two ergolines, scores 0.43 where an unrelated pair like LSD against atropine scores 0.00, against 0.75 and 0.08 under pX.

DRIFT’s score needs no reciprocal. It is already bounded at zero, and a target absent from a compound’s ranking is one the model found no structural resemblance to, so zero is already what absence means.

What the reciprocal does not fix. An empty cell still means “nobody ran this assay”, which is a different fact from “this compound does not bind here”. Zero is the right value for the second and says nothing about the first, and no method that requires a complete vector can tell them apart. That is the standing reason the drug–target plate above uses the other treatment, and the reason both are on the page.

The two plates can therefore disagree about which compounds belong together, and that disagreement is informative rather than a fault. Where they agree, the grouping survives both treatments of the missing data. Where they do not, the difference is usually telling you about the assay record rather than about the molecules.

3 · Building the tree

Both plates order their axes by average-linkage agglomeration: begin with every item alone, repeatedly join the two closest clusters, and read the leaves off the finished tree. Distances between merged clusters are updated by the Lance–Williams recurrence rather than recomputed from their members.

4 · Reading the tree, and the shaded groups

A join is drawn at the distance its two sides merged at, so depth reads as disagreement. Joins at the fallback distance — pairs too sparse to compare — are drawn as hairline dashes, because they are the algorithm running out of alternatives rather than a finding.

A group is shaded when its members joined below the tightness line: the closest quartile of the distances actually observed on that axis. Runs of one are never banded, so a singleton is not dressed up as a cluster. The threshold is derived from each matrix rather than fixed, so it means the same thing on a pX axis and on a cosine axis.

Two more obvious cuts were tried first and both misreport this data, which is worth recording because both are standard. Cutting at the fallback distance returns the entire axis as one group: average linkage blends real distances with the fallback as clusters grow, so even a tree whose pairs were mostly incomparable joins below it. Cutting into a fixed number of top-level groups fails the same way and for the same reason — with this much missing data the tree chains, absorbing one leaf at a time, so splitting the tallest join repeatedly peels off singletons and leaves one cluster holding 106 of 112 compounds. Cutting on tightness says the same thing whatever the shape of the rest of the tree.

5 · What DRIFT’s score is

Just over half of what DRIFT returns is not a molecular target at all — cell lines, organisms, tissue preparations, and the placeholders “No relevant target”, “Unchecked” and “ADMET”. All of it is excluded by the same rule that keeps such rows off the measured grid, so the DRIFT axes carry 489 molecular targets out of the 19,715 rows returned.

A dossier quotes trials and monographs that describe what a compound does to people, and those descriptions use clinical vocabulary in its clinical sense. Two of the terms are ordinary English words with narrower meanings here — affect is not a synonym for mood, and neither is a synonym for feeling — so the domains are defined once, below, rather than left to context.

These are the domains of a mental status examination: the headings a clinician describes a person under. None of them is a property of a molecule, and nothing in this glossary asserts that any compound produces any state.

Affect

The observable, moment-to-moment expression of emotion — described by its range and quality, as flat, blunted, restricted, labile or reactive.

/affective|(?:flat|blunted|restricted|constricted|labile|inappropriate|incongruent|pseudobulbar)\s+affect|affect\s+(?:lability|blunting)|emotional (?:blunting|lability|numbing)|tearful/i

Mood

The sustained, internally experienced emotional tone the person reports, as euthymic, dysphoric, euphoric, anxious or irritable.

/\bmood|euphor(?:ia|ic)|dysphor(?:ia|ic)|anxi(?:ety|ous|olytic|ogenic)|depress(?:ion|ed|ive)\b|irritab|panic|elation|apath/i

Perception

Sensory experience and its disturbances: illusions, in which a real stimulus is misinterpreted, and hallucinations, which occur without one.

/hallucinat|hallucinogen|illusion|perceptu|percept\b|visual distort|synaesthes|synesthes|depersonaliz|derealiz|sensory distort|altered (?:state|perception)/i

Cognition

Attention, concentration, orientation, learning and memory, and executive function — what a formal examination tests directly.

/cognitiv|memory|amnesi|attention|concentrat|confus|disorient|executive function|working memory/i

Thought process

The form and organization of thinking rather than its content: linear and goal-directed, or circumstantial, tangential or loosened.

/thought (?:process|disorder|form)|disorganiz|tangential|incoheren|flight of ideas|loosening of association/i

Thought content

What is being thought about — preoccupations, overvalued ideas, delusions, and thoughts of harm to self or others.

/delusion|paranoi|persecut|suicid|obsess|grandios/i

Psychomotor activity

The rate and quality of movement and speech, from retardation through to agitation.

/psychomotor|agitat|sedat|catatoni|restless|stupor|somnolen|hyperactiv|akathisi/i

Insight and judgment

Whether the person recognizes the experience as drug-induced or as externally real, and the soundness of decisions made while it lasts.

/\binsight|judg(?:e)?ment/i

The measure

Each compound page carries a web with one spoke per domain. The spoke is a count of records, not a rating of severity.

It measures the literature, not the compound. A long spoke means the subject has been written about; it says nothing about how strong the effect is. A spoke at the center means nothing recorded here discusses that domain, not that nothing happens — Thought process is currently zero for every compound in the archive, which is a fact about what these sources chose to report. Comparing two compounds compares how much has been published about each, and comparing spokes is not comparing effects.

A severity profile would be a different object and would need per-domain magnitudes from a validated instrument — 5D-ASC, ARCI, HRS or MEQ30 subscale scores from human trials. This archive holds none, and the previous build’s attempt at one is listed under held back for having had no source column at all. Counting sources is what the evidence supports, so counting sources is what the web does.

Two deliberate exclusions. Paper titles from the bibliography are not scanned: a title states a study’s subject, not its findings, and one well-titled paper would inflate a domain it never reported on. And the term lists are conservative about clinical homonyms — in this corpus the everyday verb “affects” outnumbered the clinical noun eight to one, so the Affect pattern requires a qualifier and bare “affects” matches nothing.

• PPubChem record — identity, structure, synonyms
• CChEMBL record — a measured bioactivity
• Rcomputed here with RDKit from an upstream structure
• Dderived here by arithmetic on a stored measurement
• —an explicit absence: no value exists. Never a zero, never a default