Lecture 3.5

Structure prediction

Learning objectives

After this lecture, you should be able to do the following.

Judge what a published analysis establishes, given the evidence standing behind it, what it leaves out, and where its claims are disputed.

  1. Explain what the sentence “no evaluation of this method on this system has been published as of a stated date” licenses you to conclude, and what it does not.
  2. Explain why a method’s reported performance on a benchmark set need not hold for one particular target, and name what about the benchmark decides whether it does.
  3. Given excerpts from several sources on one open method question, identify the position each source takes and the evidence each position rests on.
  4. Given several sources disagreeing about one method, differentiate a disagreement about what was observed from a disagreement about what the observations mean.
  5. Given a disagreement between published claims about one method, choose the observation that would best settle it, and identify which position each possible outcome would favor.

Account for what a structure predictor is asked to do, and compare the sources of information prediction has drawn on.

  1. State what a structure predictor takes as input and what it returns.
  2. Compare the sources of information structure prediction has relied on in turn, and name what each change added that the previous source could not supply.
  3. Explain why a predictor is given a family of related sequences rather than the target sequence on its own.
  4. Explain why a predicted model is a prediction rather than a measurement, and name what a deposited structure carries that a predicted one does not.
  5. Explain how a solved structure of a relative lets you model a target, and state what limits that route.

Relate correlated change across a sequence family to spatial contact in the folded structure, and account for what the signal needs in order to be visible.

  1. Define a pair of coevolving positions in an alignment, and state what their correlation suggests about the folded structure.
  2. Explain why a correlation between two columns need not mean the positions touch, and name what else produces one.
  3. Explain why two positions whose side chains touch in the fold are under pressure to change together.
  4. Predict what happens to the coevolution signal when the alignment available for a target holds few sequences, or many that are nearly identical.
  5. Distinguish what a single conserved column reports from what a pair of correlated columns reports, and state which of the two places a contact in the structure.

Read the confidence a predictor reports alongside a model, and determine which parts of that model the numbers support.

  1. State what a per-residue confidence value reports about a predicted model.
  2. Explain what a confidence value is built to predict, and why a high value for a region is not the same thing as that region being correct.
  3. Compare a per-residue confidence value with a B-factor, naming what each one is a statement about.
  4. Given a reported relation between confidence values and the accuracy later observed at those values, determine whether the values are calibrated, meaning whether a stated confidence matches the accuracy seen at it.
  5. Given per-residue confidence values across a predicted chain, read off which regions of the model are supported and which are not.
  6. Distinguish the reasons a region of a predicted model can carry low confidence, and state which of those reasons a further experiment could settle.

Determine what a predicted model can be used for, given that it returns one conformation of one chain.

  1. Explain what a predicted model of a single chain leaves unreported about a target whose functional unit is an oligomer.
  2. Explain why a predictor that returns a single structure answers only part of the question for a molecule whose function requires more than one state.
  3. Given a downstream analysis that depends on where a side chain sits or on the shape a target takes with a partner bound, determine whether a predicted model supplies what the analysis needs.
  4. Given both a predicted model and a low-resolution deposited structure of one target, differentiate what each one supports, and state which you would use for a stated analysis.
  5. Given a predicted model together with its confidence values and a stated purpose, differentiate the purposes the model serves from the ones it does not.
Back to top