GEMME

What it is

Proteins are chains of small units called amino acids. Change one unit, a mutation, and the protein may work as before, work less well, or fail. GEMME estimates which it will be without any laboratory test. It starts from an alignment: the same protein collected from many species and lined up position by position, so that matching parts sit in the same column. Where a column is the same in every species, evolution has not tolerated change there, and a mutation is likely to be damaging. Where a column varies freely, change is probably harmless. GEMME turns this pattern into a score for every possible substitution at every position.

What sets it apart from simple column counting is that it treats the alignment as a record of descent. Sequences are related through a branching tree, so a column that looks varied may only reflect a few distant relatives, and a column that looks fixed may just reflect many near-identical ones. GEMME weighs how far back in that history a given change appears. A substitution seen only in remote relatives, whose proteins differ in many other ways too, counts as less safe than one seen in close relatives. The output is a ranking of mutations from tolerated to disruptive, produced from sequences alone.

Why scientists use it

Measuring mutations in the laboratory is slow, and a single modest protein has thousands of possible single changes. Scientists therefore want a cheap score that sorts the likely harmless from the likely harmful, so that experiments can be aimed at the interesting cases. GEMME is attractive because it is light: it needs an alignment and little else, it runs quickly, and it has few settings to tune. That makes it useful as a first pass over a protein nobody has studied, and as a benchmark against which heavier methods are judged.

It also serves as a stand-in for data that does not exist. Modern work often trains models on measured mutation effects, but those measurements are scarce. GEMME scores can fill the gap as substitute labels, giving a learning method something to imitate while real measurements are still being gathered. The same logic appears wherever evolution has left a usable record, in human genetics, in enzyme engineering and in studies of protein stability, because the underlying claim is the same: what life has avoided changing was probably worth keeping.

In the record

Limits

The method knows only what evolution happened to record. A protein with few sequenced relatives gives a thin alignment and unreliable scores, and a position may be conserved for reasons unrelated to the function being studied. It judges whether a change is tolerated in nature, not what the change does: it cannot say that a mutation makes a protein less stable rather than less active, nor how much, and it has nothing to say about engineered proteins that evolution never tested. Scores are a ranking, not a measurement, and still need experiments to anchor them.

Written by AIxSci: the background from general knowledge, the examples from checked records.

Every article