The Pencil Room
Questions about predicting protein structures that nobody can answer yet. Each is written in pencil, with a source that says it is still open. When one is settled, it is inked with the year.
Fold Commons shows AlphaFold DB models. Some of these questions are open for those models; some are open for every method. Each entry says which.
In pencil: still open
-
In pencil · still open · for the models shown here
Can a model show how a protein moves between its shapes, not just one frozen pose?
Many proteins switch between two or more shapes to do their job. A model like the ones on this site usually shows one of them, and you cannot choose which.
“Where a protein is known to have multiple conformations, AlphaFold usually only produces one of them. The output conformation cannot be reliably controlled.”
-
In pencil · still open · for the models shown here
How much does one changed letter change a protein's stability or its job?
A model of the changed sequence usually looks almost the same as the original, even when the change causes disease. A similar shape is not evidence of a small effect.
“AlphaFold has not been validated for predicting the effect of mutations.”
-
In pencil · still open · for the models shown here
What do the floppy, disordered parts of a protein actually look like?
Low confidence can mark a region as disordered, but the loose ribbon drawn there is not a picture of what that region does. It is one line through many possible shapes.
“the prediction makes no statement about the relative likelihood of different conformations”
-
In pencil · still open · for every method
Can a model reliably predict how several proteins fit together, including an antibody on its target?
In the CASP16 blind test (2024), even the best groups predicted only a little more than half of the complexes to high accuracy. The assessors called antibody–antigen pairs difficult targets.
“complex structure prediction remains an unsolved challenge”
-
In pencil · still open · for every method
Can RNA's 3-D shape be predicted as well as a protein's now can?
Not yet. In CASP16, no prediction of a new natural RNA structure came close to the experimental one, and accuracy had not clearly improved since earlier blind tests.
“Accounting for template availability, there has not been a notable increase in nucleic acid modeling accuracy between previous blind challenges and CASP16.”
Kretsch RC, et al. Assessment of Nucleic Acid Structure Prediction in CASP16. Proteins (2026).
-
In pencil · still open · for the models shown here
Where do the metals, cofactors and small molecules sit in the models shown here?
The models on this site contain protein only. Some newer methods place small molecules too, so this one is open for these models rather than for the whole field.
“AlphaFold does not predict the positions of any non-protein components found in experimental structures”
In ink: settled
-
Inked 2020 · for every method
Can a computer predict the shape of one protein chain from its sequence alone, close to what an experiment shows?
Open for about fifty years. In the CASP14 blind test (2020), AlphaFold's models came close to experimental structures for most single chains. Most, not all: the other entries in this room are what is left.
“demonstrating accuracy competitive with experimental structures in a majority of cases”
Reviewed 2026-10-13; next review by 2027-10-13. Each year every source is read again and anything answered is inked. This room claims only that these questions are open and names who says so. A sibling room stands on Spark & Whetstone.