Evaluating retrieval-augmented generation (RAG) systems requires benchmarks that capture diverse question characteristics, yet practitioners lack empirical guidance on which dimensions to vary and at what granularity. We present HieraRAG, a hierarchical framework for studying granularity in RAG benchmark construction, defining optimal granularity as the level that maximizes discriminative power (the standard deviation of generation quality across categories) within a given RAG configuration. As a case study, we generate 5,872 synthetic question-answer (QA) pairs from FineWeb-10BT across 3 dimensions (Question Complexity, Answer Type, Linguistic Variation) at 3 granularity levels (2, 4, and 8 categories). With a BM25+Falcon-3-10B pipeline, optimal granularity varies by dimension: complexity benefits from fine-grained distinctions (discriminative power: 0.053) while answer type and linguistic variation peak at medium granularity. We introduce a Coherence Ratio metric to quantify whether fine-grained splits cleanly subdivide parent categories, revealing structural differences across dimensions (Question Complexity: 0.40 vs. Answer Type: 1.44). Human evaluation of 110 stratified QA pairs confirms synthetic quality. While these specific findings reflect a single configuration, HieraRAG provides a portable procedure and validation metric for practitioners to determine evaluation granularity within their own RAG settings.

Abstract The reference indexing problem for $$k$$ -mers is to pre-process a collection of reference genomic sequences $$\mathcal {R}$$ so that the position of all occurrences of any queried $$k$$ -mer can be rapidly identified. An efficient and scalable solution to this problem is fundamental for many tasks in bioinformatics. In this work, we introduce the spectrum preserving tiling (SPT), a general representation of $$\mathcal {R}$$ that specifies how a set of tiles repeatedly occur to spell out the constituent reference sequences in $$\mathcal {R}$$ . By encoding the order and positions where tiles occur, SPTs enable the implementation and analysis of a general class of modular indexes. An index over an SPT decomposes the reference indexing problem for $$k$$ -mers into: (1) a $$k$$ -mer-to-tile mapping; and (2) a tile-to-occurrence mapping. Recently introduced work to construct and compactly index $$k$$ -mer sets can be used to efficiently implement the $$k$$ -mer-to-tile mapping. However, implementing the tile-to-occurrence mapping remains prohibitively costly in terms of space. As reference collections become large, the space requirements of the tile-to-occurrence mapping dominates that of the $$k$$ -mer-to-tile mapping since the former depends on the amount of total sequence while the latter depends on the number of unique $$k$$ -mers in $$\mathcal {R}$$ . To address this, we introduce a class of sampling schemes for SPTs that trade off speed to reduce the size of the tile-to-reference mapping. We implement a practical index with these sampling schemes in the tool . When indexing over 30,000 bacterial genomes, reduces the size of the tile-to-occurrence mapping from 86.3 GB to 34.6 GB while incurring only a 3.6 $$\times $$ slowdown when querying $$k$$ -mers from a sequenced readset. Availability: is implemented in Rust and available at https://github.com/COMBINE-lab/pufferfish2 .

There has been rapid development of probabilistic models and inference methods for transcript abundance estimation from RNA-seq data. These models aim to accurately estimate transcript-level abundances, to account for different biases in the measurement process, and even to assess uncertainty in resulting estimates that can be propagated to subsequent analyses. The assumed accuracy of the estimates inferred by such methods underpin gene expression based analysis routinely carried out in the lab. Although hyperparameter selection is known to affect the distributions of inferred abundances (e.g. producing smooth versus sparse estimates), strategies for performing model selection in experimental data have been addressed informally at best. Thus, we derive perplexity for evaluating abundance estimates on fragment sets directly. We adapt perplexity from the analogous metric used to evaluate language and topic models and extend the metric to carefully account for corner cases unique to RNA-seq. In experimental data, estimates with the best perplexity also best correlate with qPCR measurements. In simulated data, perplexity is well behaved and concordant with genome-wide measurements against ground truth and differential expression analysis. To our knowledge, our study is the first to make possible model selection for transcript abundance estimation on experimental data in the absence of ground truth.

MOTIVATION: Mapping genetic interactions (GIs) can reveal important insights into cellular function and has potential translational applications. There has been great progress in developing high-throughput experimental systems for measuring GIs (e.g. with double knockouts) as well as in defining computational methods for inferring (imputing) unknown interactions. However, existing computational methods for imputation have largely been developed for and applied in baker's yeast, even as experimental systems have begun to allow measurements in other contexts. Importantly, existing methods face a number of limitations in requiring specific side information and with respect to computational cost. Further, few have addressed how GIs can be imputed when data are scarce. RESULTS: In this article, we address these limitations by presenting a new imputation framework, called Extensible Matrix Factorization (EMF). EMF is a framework of composable models that flexibly exploit cross-species information in the form of GI data across multiple species, and arbitrary side information in the form of kernels (e.g. from protein-protein interaction networks). We perform a rigorous set of experiments on these models in matched GI datasets from baker's and fission yeast. These include the first such experiments on genome-scale GI datasets in multiple species in the same study. We find that EMF models that exploit side and cross-species information improve imputation, especially in data-scarce settings. Further, we show that EMF outperforms the state-of-the-art deep learning method, even when using strictly less data, and incurs orders of magnitude less computational cost. AVAILABILITY: Implementations of models and experiments are available at: https://github.com/lrgr/EMF. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.