Question 5 of 6 Advanced Scenario
Why does deterministic training not guarantee biological reproducibility?
Short answer
Deterministic computation repeats a numerical pipeline under controlled conditions. Biological reproducibility asks whether the finding holds across independent samples, experiments and relevant settings.
Why
Fixing random seeds and deterministic operations can reproduce a model trained on one RNA-seq dataset, yet the biological signal may vanish when the experiment is repeated with independent donors, another sequencing batch or a different laboratory. Document software, hardware, data preprocessing, labels and experimental protocol to make the computational result auditable. Then test biological replication using independent units and suitable controls rather than rerunning the same cells or technical replicates.
PyTorch notes that full numerical reproducibility across releases and platforms is not guaranteed even with identical seeds. Separately assess model robustness across cohorts, sites and measurement technologies. A deterministic result can faithfully reproduce batch effects or leakage, so validation must include experimental design and biological relevance.
Common mistakes
- Calling two identical seeded runs independent validation of a biological hypothesis
The interviewer might ask next
What replication study would expose a model that learned sequencing-batch signatures?