ΔΔG and Tm: why predictions cannot be converted to degrees

Check the endpoint and sign convention

ΔΔG describes a free-energy difference between a variant and WT. Publications and tools may use different definitions and signs. A number alone is therefore insufficient without the sign convention, units and source.

Foldiff uses “negative suggests stabilization” and kcal/mol. Melting temperature Tm describes protein behavior during heating and depends on assay conditions. For laboratory Tm data, the service calculates the difference between variant and WT means; it does not convert an energy prediction.

Why two predictors can disagree

Foldiff’s main selection uses a linear model of substitution and environment features. ThermoMPNN uses a neural network. Their protein representations and estimation methods differ. The beta shows values side by side without averaging them into a single “success probability.”

For S125D, the saved example shows −1.19 and +0.42 kcal/mol. The first supports stabilization, the second suggests the opposite direction. This is a useful uncertainty signal. Compare against available measurements; without them, make an informed experimental choice.

What to check before an experiment

• Whether the original amino acid and numbering match your construct.
• Whether a functional site is affected and how confidently its environment is modeled.
• Whether a reported measurement refers to the same protein, construct and conditions.
• Whether a single-substitution estimate is being applied to a combination without separate testing.

A large negative number does not remove these checks. Ranking helps allocate attention, but does not establish commercial success probability. A missing second channel does not confirm the first prediction.

Assessing a laboratory series

Illustrative example: WT measurements of 54, 55 and 56 °C and variant measurements of 59, 60 and 61 °C give means of 55 and 60 °C and a +5 °C difference. These fictional values demonstrate table arithmetic, not a Foldiff validation experiment.

Next, check separate activity measurements and reproducibility. A candidate must meet effect-direction and activity-threshold criteria relative to WT. The “Candidate for repeat testing” status describes this filter, not established statistical significance.

What accuracy can you expect?

The service claims no single accuracy percentage across enzymes. Saved dataset metrics vary substantially, and database lookup cannot replace an independent pilot. Your task needs new measurements, a simple selection baseline and evaluation of transfer to the protein family.

Keep the original report and conditions with the series CSV and JSON. This lets you repeat the analysis and identify confirmed hypotheses. This cycle is more useful than treating every prediction as a precise measurement.