Question 1

Why does split strategy determine reported accuracy?

Accepted Answer

Molecules are tested in families around a scaffold, so random splitting places close relatives on both sides and the model succeeds by recognising them. Splitting by scaffold, target family or deposition date approximates real use, where candidates differ from anything seen before. Report which split was used prominently.

Question 2

What distortions do public activity datasets carry?

Accepted Answer

Inactive results are recorded inconsistently, assay conditions differ between contributing studies, and popular targets are heavily over-represented. Frequent interferers appear active across many unrelated assays. Document every curation decision, including how activity thresholds and duplicate measurements were handled, since these choices move results substantially.

Question 3

Which baselines must be beaten?

Accepted Answer

Similarity search against known actives and simple models built on molecular fingerprints remain strong on many tasks and are often omitted from comparisons. Tune them with the same effort given to the proposed method. A deep model matching a fingerprint baseline has demonstrated nothing beyond additional computational cost.

Question 4

How should uncertainty be handled?

Accepted Answer

Report calibrated confidence alongside every prediction and define the applicability region, since models extrapolate poorly to chemistry unlike their training data. Ensembles and conformal approaches both give usable estimates. A ranked list without uncertainty invites confident wrong answers on exactly the novel candidates a project cares about.

Question 5

What complicates learning from cellular images?

Accepted Answer

Plate position, imaging session and cell batch produce signals stronger than many treatment effects, and a model will learn those instead if permitted. Randomise treatments across plates, include controls on every plate, and hold out entire batches during evaluation rather than individual wells.

Question 6

How are generated molecules assessed?

Accepted Answer

Computational filters for synthesisability, predicted properties and distance from training data give a first pass, but generation claims need synthesis and testing. Report how many suggestions were made, how many could be synthesised and how many showed activity, since selective reporting of successes is common in this area.

Question 7

Why does prospective testing carry so much weight?

Accepted Answer

Retrospective accuracy inherits every property of the dataset, whereas selecting candidates and testing them commits the model in advance. Even a small prospective set with honest reporting of hit rate against a sensible comparison is stronger evidence than another benchmark table.

Biotechnology

Biotechnology

Research Frontiers in Artificial Intelligence Drug Screening