Fig. 2

Overview of the quantify-then-validate approach using smoothed perplexity to evaluate the quality of abundance estimates directly on fragment sets in the absence of ground truth. (1) An input fragment set is first partitioned into a quantified and a validation set. (2) Abundance estimates for different candidate models (e.g. for explored hyperparameters as part of model selection) are inferred from the quantified fragment set only. (3) To account for “impossible” fragments and avoid shrinkage to unbounded perplexities, given abundance estimates are smoothed (see Sect. 3.2). (4) Mapping probabilities to the reference transcriptome are computed for fragments in the validation set. (5) Smoothed perplexity is computed given each input abundance estimate and the held-out validation fragment set to evaluate and perform model selection—the lower the perplexity, the better an abundance estimate describes the held-out set of validation fragments