Understanding Hotspots and Synergies

Hotspot analysis and synergy detection are two halves of one question: which parts of the peptide drive the property you care about? Hotspots answer it one position at a time. Synergies answer it for positions acting together, and they are built directly on top of the single-position numbers hotspots produce — which is why the two are explained here as one piece.

This page explains how both are computed and, more importantly, how to read the numbers without over-trusting them. For the step-by-step settings, see How to Analyze Hotspots and Synergies.

Hotspots: a position's impact is relative, always

How impact is computed

At each position, Ideation groups the aligned peptides by the monomer they carry there — or None where the peptide has a gap. A pendant chemical object or a bridge does not enter its residue's grouping: each is scored separately, by its own variants (see below). Every variant is reduced to one representative value (the mean or median of the peptides carrying it, per your Baseline Aggregation Method).

The position's raw score is the range of those representatives: the highest minus the lowest. A position where every variant behaves the same has a range near zero; one where swapping the residue swings the endpoint has a large range.

That raw range is then divided by the largest range of any position currently in scope, and the result is what the impact bar shows as a percentage.

Two consequences follow from the same normalisation:

• Filtering changes every number. Hotspot analysis runs on the peptides currently passing your filters, so applying or removing a filter re-normalises the whole set and can flip an element from Hot to Cold without any data changing.
• Positions, chemical objects, bridges and regions are not comparable. Each is normalised against the strongest of its own kind — a bridge against the strongest bridge, a chemical object against the strongest chemical object. A 70% region and a 70% position do not mean the same thing.

Hot, Warm and Cold

The percentage is banded into a badge using two cut-offs you control, defaulting to Hot at 70% and Warm at 50%:

Since the thresholds sit on a relative scale, moving them changes how much of the peptide is highlighted, not how strong the underlying effect is.

The reference-peptide baseline measures something different

Switching Baseline Aggregation Method to Reference peptide value does not merely change the number impact is measured from — it changes the metric. Instead of the spread between variants, the score becomes the largest absolute deviation of any variant from the reference peptide's value.

That is a one-sided measure, and it has a counter-intuitive consequence: a position where all variants sit equally far from the reference — that is, a position with no variation among its variants at all — scores high, because every variant deviates. Under the average or median baseline, that same position would score near zero.

Use Reference peptide value when your question is "how far does this series stray from my lead compound?" Use Average or Median when the question is "where does this series vary?"

Other things worth knowing

• A gap is a variant called None, so "carries a residue" versus "carries nothing" is scored as if it were a substitution.
• A pendant chemical object is scored on its own node, by grouping peptides on the chemical carried at that attachment (None when absent) — its residue keeps its monomer-only score.
• A bridge is scored by its own flavours — each chemical linker, each direct bridge type, and the unbridged peptides — the same groups its gallery shows.
• If the reference peptide has no value for the chosen endpoint, reference mode zeroes every position to Cold rather than falling back to another baseline.
• Peptides missing a value for the endpoint are silently excluded from that position's groups.

Synergies: departures from an additive model

The model

Synergy detection asks whether two or three positions do something together that their individual effects do not predict.

First, each variant at each position gets a main effect — the mean of the peptides carrying that variant, minus the overall mean of the dataset. This is the additive part: the assumption that a peptide's value is the baseline plus the sum of its independent per-position contributions.

For a given combination of variants across a position set, Ideation compares:

• the observed value: the mean of peptides that actually carry that exact combination;
• the predicted value: the baseline plus the main effects of each variant in the combination.

The difference — observed minus predicted — is the residual. A residual near zero means the positions behave independently. A large residual means they interact.

How a position set is scored

A set's score is the support-weighted root-mean-square of its combinations' residuals: each residual is squared, weighted by how many peptides carry that combination, averaged, then square-rooted. Weighting by support means a large residual seen in two peptides counts for far less than a moderate one seen in twenty.

The direction — synergistic or antagonistic — comes from the set's dominant combination, chosen by support-weighted deviation rather than by raw residual size, for the same reason. Residuals within a narrow neutral band (5% of the endpoint's standard deviation) are treated as noise rather than as a signal in either direction.

Two guards against false positives

Synergy scores are normalised the same relative way hotspots are, and the code adds two protections hotspots lack:

• An absolute floor. A set must explain a non-trivial share of the endpoint's spread — at least 10% of its standard deviation — before it is reported at all. Without this, relative normalisation would confidently flag the strongest near-additive set on perfectly additive data.
• Per-order normalisation. Pairs are scaled against the strongest pair, triples against the strongest triple, regions against the strongest region. This means enabling triples only adds triple results; it never rescales the pairs you were already looking at. It also means a 100% pair and a 100% triple are each strongest-of-their-kind and not comparable to each other.

Confounding: why some combinations show N/A

This is the single most important thing to understand about synergy results.

The additive model can only separate an interaction from the single-position effects if the data contains a contrast — peptides that carry one variant without the other. When a variant appears only inside one combination and never on its own, there is no such contrast. The model would then attribute the variant's entire effect twice, once as a main effect and once as an interaction, and the resulting residual would be an artefact of the fit rather than a real interaction.

Ideation detects this and marks the combination not scored (N/A). Confounded combinations are excluded from the set's score, from its direction, and from the choice of dominant combination.

A knock-on effect: if fewer than two combinations in a set survive the confounding check, the set is dropped entirely and never appears in the list. You see nothing, rather than an explanation of why.

Strength is not activity

A synergy score measures interaction strength — how far the combination departs from additivity. It says nothing about whether the departure is good.

An antagonistic pair that destroys activity and a synergistic pair that boosts it can carry identical strength scores. Activity is tracked as a separate axis, oriented by whether higher or lower values count as active for your endpoint, and exposed through the Filter by activity impact control and the "best opportunity" a set offers. Read the two together: strength tells you the positions interact; activity tells you whether you want that interaction.

Which combinations get examined

Enumerating every pair and triple across a long peptide is not affordable in a browser, so Ideation bounds the search:

• Only positions with at least two distinct variants are considered — an invariant position cannot interact with anything.
• Pairs are enumerated over at most 24 positions. If more qualify, the least individually-variable ones are dropped.
• Triples are enumerated over at most 8 positions, ranked by how strongly each position participates in the pairs already found — not by its own single-position variability. This is deliberate: it keeps positions that matter only in combination, which a variability ranking would discard.
• In regions mode, only windows whose first and last positions both vary are kept, so a region is never just a shorter one padded with invariant positions.

Two further behaviours that change results rather than just the view:

• Ignore gap combinations (on by default) does not merely hide rows: the additive model is refitted on the subset of peptides having a residue at every position in the set. Baselines, main effects and therefore every residual change when you toggle it.
• The minimum peptides per combination threshold applies per combination, not per set. Combinations below it drop out of the score while the set itself can still be reported.

Where to go next

• How to Analyze Hotspots and Synergies — the settings, controls and reading the canvas
• Understanding Peptide Clustering — the other analysis built on the same structural data
• Understanding Peptides in Ideation — the wider concepts