Question 1

How should the test split be constructed?

Accepted Answer

Partition by sequence identity at a stated threshold, and where possible hold out entire folds or use a structure release cutoff so nothing homologous to a test target appears in training. Report the exact rule and the resulting set sizes, because a permissive split is the most common reason a reported improvement fails to reproduce.

Question 2

Which accuracy metrics should be reported?

Accepted Answer

Use more than one. Global measures summarise overall topology, distance-based scores capture local geometry, and per-residue confidence indicates where the model is unsure. Report the distribution across targets rather than a mean alone, since a handful of easy homologous cases can carry an otherwise mediocre average.

Question 3

Do evolutionary couplings still matter given language models?

Accepted Answer

Both routes remain in use. Multiple sequence alignments supply strong coevolutionary signal but cost time to build and fail for shallow families. Sequence language models give a fast single-sequence path with weaker accuracy on hard targets. Many projects compare the two directly, which makes a clean baseline and a fair timing comparison essential.

Question 4

How much compute does a project like this need?

Accepted Answer

It varies enormously. Fine-tuning an existing network or building a lightweight module can run on a single accelerator, while training a large architecture from scratch usually needs a cluster allocation. Scope the project to the resources actually available and confirm the allocation before committing, since a reduced training budget changes which questions can be asked.

Question 5

What makes an ablation study convincing?

Accepted Answer

Change one component at a time, hold the split and the seeds fixed, and run each variant several times so the reported gap can be compared against seed-to-seed variation. State the training budget for every variant, because a component that simply lets a network train longer within the same wall clock is not the same as an architectural gain.

Question 6

Should predictions be validated experimentally?

Accepted Answer

Not usually within the thesis itself, since structure determination is a separate programme of work. Where a collaborator can test a specific prediction, that evidence is valuable, but most projects rest on held-out structures and careful benchmarking. Choose a small number of biologically interesting cases for detailed discussion instead.

Bioinformatics

Bioinformatics

Research Frontiers in Protein Structure Prediction Deep Learning