The Average Repetitive Substring Index (ARSI) is an unsupervised genome-only measure that scores a genetic region by the mean length, over nucleotide positions, of the longest substring starting at each position that also occurs in a reference set of the organism's genetic elements; high ARSI in 5' UTRs, coding regions, introns, and 3' UTRs of E. coli, S. cerevisiae, and S. pombe correlates with elevated gene expression. Zafrir and Tuller (2017) explicitly left open four linked questions: (1) whether high-ARSI recurrent substrings causally drive expression, including position- and strand-specific enhancer/repressor roles, testable e.g. by CRISPR/Cas9 targeted genome editing or reporter assays; (2) which condition-specific or low-abundance regulatory motifs ARSI misses (its false-negative regime), requiring benchmarking against experimentally derived motif sets; (3) principled optimization of the ARSI scoring scheme, e.g. reweighting substring lengths and counts, with model selection against expression and splicing outcomes; and (4) molecular-evolution models that incorporate locally varying concentrations of regulatory codes across transcript regions.

Background

Gene expression in all organisms is regulated by sequence signals interleaved through every part of a transcript: promoters, 5' and 3' untranslated regions (UTRs, the non-coding flanks of mature messenger RNA), introns (non-coding segments removed by splicing), coding sequences (CDS, the protein-coding part, also called ORFs for open reading frames), and splice junctions. Most computational methods for finding these signals are supervised, i.e. trained on measured expression data or known motifs. Zur and Tuller (2015) introduced an unsupervised alternative for coding regions: with no expression data or motif library, exploit the redundancy of the genetic code by measuring how often a gene's substrings are shared with the rest of the genome, on the premise that selection for expression-related signals leaves long recurrent substrings in well-expressed genes. Zafrir and Tuller (2017) generalized this to the Average Repetitive Substring Index (ARSI): for each nucleotide position of a sequence S one finds the longest substring starting there that also appears in at least one element of a reference set G (typically all other elements of that region type in the same organism), and ARSI is the mean of these lengths, ARSI(G,S)=∑i,j∣sij∣/∣S∣\mathrm{ARSI}(G,S)=\sum_{i,j}|s_{ij}|/|S|. Sliding-window ARSI profiles, compared against randomized genomes that preserve GC content and codon usage, showed elevated information concentration near ORF boundaries and intron boundaries in E. coli, S. cerevisiae, and S. pombe, and ARSI scores of most regions correlated significantly with expression level, enough to rank genes by expression from sequence alone. The authors themselves stressed, however, that these are correlations: ARSI says nothing about which substrings are functional, in which direction they act, or what the measure misses. In their discussion they pose four open issues that jointly define what is needed to turn the correlational ARSI signal into validated biological knowledge: (i) causality and functional directionality — ARSI–expression correlation cannot distinguish substrings that drive expression from passengers linked to other features, and the same motif may act as an enhancer in one position or strand and a repressor in another (their example: a splicing motif that activates splicing downstream of an exon but represses it elsewhere), so perturbation experiments are required — CRISPR/Cas9 genome editing (targeted rewriting of chosen DNA segments) or reporter assays (placing the test sequence before an easily measured reporter gene to quantify its regulatory effect) that mutate or relocate high-ARSI motifs; (ii) false negatives — ARSI is expected to detect abundant broadly acting motifs but to miss condition-specific motifs and splicing regulatory elements whose cognate binding factors are lowly expressed, and this regime must be characterized by benchmarking against experimentally derived motif sets and condition-specific regulatory data; (iii) scoring optimization — the plain average over substring lengths is a first design that could be improved, e.g. by reweighting the contributions of different repetitive substring lengths and their counts, with principled model selection against expression and splicing outcomes; and (iv) evolutionary modeling — the demonstrated selection on these substrings calls for molecular-evolution models (probabilistic models of nucleotide substitution along lineages, whose rates can be made to depend on position and context) that incorporate locally varying concentrations of regulatory codes across UTRs, introns, junctional windows, and coding regions, to study transcript evolution under such constraints.

Problem Statement

For the ARSI measure of Zafrir and Tuller (2017) — the mean longest substring, per nucleotide position of a region, that recurs in a reference set of the organism's genetic elements — determine what the detected recurrent substrings actually do and what the measure actually captures, along the four linked axes the source itself poses: (1) Causality and directionality: by targeted perturbation of high-ARSI substrings (e.g. CRISPR/Cas9 sequence edits or reporter assays that mutate or relocate them), establish whether specific ARSI-detected substrings drive expression and splicing outcomes, act directly or via linkage to other features, and how their effect depends on position and strand (e.g. whether the same splicing-associated motif activates downstream of an exon but represses elsewhere). (2) False-negative regime: benchmark ARSI against experimentally derived motif sets and condition-specific regulatory data to characterize which regulatory motifs it recovers and which it misses (e.g. condition-specific motifs and splicing regulatory elements whose cognate factors are lowly expressed). (3) Scoring optimization: improve and systematically optimize the ARSI scoring scheme — for example by reweighting the contributions of different repetitive substring lengths and counts — with principled model selection to determine which variants best predict expression and splicing outcomes. (4) Evolutionary integration: develop and validate molecular-evolution models that incorporate locally varying concentrations of regulatory codes across UTRs, introns, junctional windows, and coding regions, so that transcript evolution under these constraints can be studied quantitatively. The common objective is to convert the correlational, genome-only ARSI signal into causally validated, quantitatively optimized, and evolutionarily grounded understanding of the gene-expression code in transcript regions. The four axes are independent, independently attackable sub-questions rather than a single conjunctive requirement, and resolving any one of them resolves the task (the wet-lab perturbation of axis (1) is desirable but not required when the resolution rests on another axis); an axis counts as resolved only when its operational criterion is met — axis (1): a targeted edit or reporter assay demonstrating that mutating or relocating a specific high-ARSI substring causally changes expression or splicing, with the direction of the effect and its position and strand dependence established; axis (2): a benchmark on experimentally derived motif sets and condition-specific regulatory data reporting which motifs ARSI recovers and which it misses; axis (3): a scoring variant that beats baseline ARSI on held-out expression and splicing outcomes by an explicitly stated metric; axis (4): a substitution model that reproduces a named observed ARSI/expression signature. The common objective stated above describes the union of all four axes, not a condition that each resolved axis must satisfy on its own.

References

  1. Zafrir Z, Tuller T. 'Unsupervised detection of regulatory gene expression information in different genomic regions enables gene expression ranking.' BMC Bioinformatics 18:77 (2017). DOI: 10.1186/s12859-017-1497-z. https://doi.org/10.1186/s12859-017-1497-z
  2. Zur H, Tuller T. 'Exploiting hidden information interleaved in the redundancy of the genetic code without prior knowledge.' Bioinformatics 31(8):1161-1168 (2015). DOI: 10.1093/bioinformatics/btu797. https://doi.org/10.1093/bioinformatics/btu797
  3. Goz E, Zafrir Z, Tuller T. 'Universal evolutionary selection for high dimensional silent patterns of information hidden in the redundancy of viral genetic code.' Bioinformatics 34(19):3241-3248 (2018). DOI: 10.1093/bioinformatics/bty351. https://doi.org/10.1093/bioinformatics/bty351
  4. Diament A, Weiner I, Shahar N, Landman S, Feldman Y, Atar S, Avitan M, Schweitzer S, Yacoby I, Tuller T. 'ChimeraUGEM: unsupervised gene expression modeling in any given organism.' Bioinformatics 35(18):3365-3371 (2019). DOI: 10.1093/bioinformatics/btz080. https://doi.org/10.1093/bioinformatics/btz080
  5. Galili M, Tuller T. 'CSN: unsupervised approach for inferring biological networks based on the genome alone.' BMC Bioinformatics 21:190 (2020). DOI: 10.1186/s12859-020-3479-9. https://doi.org/10.1186/s12859-020-3479-9
  6. Bahiri-Elitzur S, Tuller T. 'Codon-based indices for modeling gene expression and transcript evolution.' Computational and Structural Biotechnology Journal 19:2646-2663 (2021). DOI: 10.1016/j.csbj.2021.04.042. https://doi.org/10.1016/j.csbj.2021.04.042
  7. Burghardt M, Diament A, Tuller T. 'Designing genetically stable multicopy gene constructs with the ChimeraUGEM web server.' NAR Genomics and Bioinformatics 7(4):lqaf191 (2025). DOI: 10.1093/nargab/lqaf191. https://doi.org/10.1093/nargab/lqaf191