Question 1

How should a validation split be constructed for escape prediction?

Accepted Answer

Random splitting leaks information, because closely related sequences appear on both sides. Splits should be stratified by collection time and by lineage, so the model is asked to score variants it has no phylogenetic neighbours for. Reporting performance on a genuinely prospective window, against a frequency-ranked baseline, is the strongest evidence available.

Question 2

Why must fitness be modelled alongside escape?

Accepted Answer

A substitution that abolishes antibody recognition but also impairs receptor binding or folding will not spread. Escape scores taken alone therefore over-nominate positions that never matter epidemiologically. Joint scoring, or filtering escape candidates through a measured or predicted fitness term, produces shortlists that correspond much more closely to variants observed later.

Question 3

What role does mutational scanning play in this work?

Accepted Answer

Scanning experiments measure antibody binding and often replication for large substitution libraries at defined epitopes, giving quantitative ground truth. Model calibration against these measurements is more informative than benchmark ranking alone. Coverage is limited to the epitopes assayed, so claims about unmeasured regions remain predictions rather than validated results.

Question 4

How can sampling bias in sequence records be handled?

Accepted Answer

Public sequence collections over-represent regions with strong surveillance capacity and periods of intense concern. Reweighting by collection region and time, subsampling per lineage, and reporting results separately for under-sampled clades all help. The bias should be characterised explicitly in the thesis rather than assumed to average out during training.

Question 5

Which uncertainty measures are appropriate for these models?

Accepted Answer

Predictions are made outside the training distribution by design, so confidence estimates must reflect distance from observed sequence space rather than in-sample variance alone. Ensembles, conformal approaches and explicit distance-to-training measures are all in use. Whatever is chosen, calibration should be checked on the time-held-out window, not on random data.

Question 6

How do prediction results feed into antibody design?

Accepted Answer

Predicted escape maps identify epitopes where resistance is likely to emerge quickly, which argues against advancing candidates that bind there alone. Combinations targeting sites with divergent escape profiles and high fitness cost are favoured. Design teams generally need ranked, interpretable site-level output rather than a single aggregate score per variant.

Biotechnology

Biotechnology

Research Frontiers in Viral Escape Mutation Prediction Models