Splicing: Multifaceted Recombination

• Splicing is a recombination operation applied across various fields like RNA processing, formal language theory, graph theory, digital forensics, topology, and software management.
• It involves cutting, re-linking, and controlled recombination to create diverse outputs such as alternative RNA isoforms, new graph structures, forged images, and modified software dependency trees.
• Recent advances include deep learning for splice junction prediction and graph-based transcript comparisons that enhance both biological and computational splicing analysis.

Splicing denotes several distinct operations in which elements are cut, recombined, glued, or re-linked. In molecular biology it is the processing of precursor RNAs into mature transcripts; in formal language theory it is a binary word operation inspired by DNA recombination; in graph theory it is an operation that generates new graphs from given graphs; in digital forensics it is a local image forgery formed by pasting a region from one image into another; in low-dimensional topology it is a gluing operation on knot exteriors and, operadically, a way to compose embeddings; and in software package management it is a representation of binary re-linking in Spack (Lee et al., 2015, Kari et al., 2011) 0702034.

1. Molecular splicing in RNA processing

In eukaryotes, RNA splicing is the process by which introns are removed from pre‑mRNA and exons are joined to form a mature transcript. The spliceosome recognizes donor and acceptor splice sites and catalyzes two transesterification reactions; alternative splicing then allows one gene to produce multiple isoforms through exon skipping, alternative $5'$ or $3'$ splice sites, intron retention, and mutually exclusive exons (Chan et al., 2021). The basic junction terminology is stable across the literature considered here: the donor site is the exon–intron boundary and the acceptor site is the intron–exon boundary (Lee et al., 2015).

The sources also stress that splice-site recognition is not exhausted by consensus motifs. Pre‑mRNA secondary structure may facilitate or hinder the interaction with factors and small nuclear ribonucleoproteins that regulate splicing, and in yeast the analysis of $3'$ splice-site recognition is used as a working example of structure-dependent regulation (Plass et al., 2015). A complementary physical account attributes part of splice-site recognition modality to entropy effects and depletion attraction in a crowded nucleus, arguing that these effects may explain the relevance of the aspecific intron length variable in the choice between intron definition and exon definition, and may also motivate qualitative features of higher-eukaryotic genome architecture (Osella et al., 2009).

Epigenetic regulation appears in the sources as an additional layer. In colon carcinoma cells, breast-cancer progression models, and B‑ALL, intragenic DNA methylation is systematically correlated with alternative splicing of the CD44 transmembrane receptor; loss of methylation is associated with skipping of CD44 variant exons, altered recruitment of MBD1 and HP1$\gamma$, and a shift toward the short standard isoform CD44s (Batsché et al., 2019).

The term also includes trans-splicing systems. In Trypanosoma brucei, trans-splicing of leader sequences onto the $5'$ ends of mRNAs is treated as a genome-wide regulatory layer rather than a marginal curiosity. Spliced leader trapping detected the $5'$ splice sites of 85% of the annotated protein-coding genes, found up to 40% of transcripts to be differentially expressed, and discovered more than 2500 alternative splicing events, many of which appear to be stage-regulated (Nilsson et al., 2021).

2. Computational analysis and prediction of biological splicing

Short-read RNA-seq turned splicing into a computational pipeline with several distinct tasks: splice-aware mapping and read assignment, transcript reconstruction or de novo assembly, quantification of isoforms and events, differential splicing or differential isoform expression, and visualization. Junction reads, exon coverage, splice graphs, PSI-like event measures, and probabilistic isoform quantification are the standard abstractions in this workflow (Alamancos et al., 2013).

At the sequence-classification level, splice-junction prediction has been formulated as a supervised problem on fixed windows centered at exon boundaries. One deep recurrent approach uses 60‑mer DNA sequences, trainable dense 4-dimensional nucleotide embeddings, and 3‑class classification into acceptor, donor, and non-site; on UCSC hg19/hg38, the proposed LSTM-based RNN achieved test accuracy $0.943$ and $F1 = 0.9430$, compared with $0.888$ for a DBN and $0.868$ or $3'$0 for SVM baselines (Lee et al., 2015). The same source emphasizes a common misconception: canonical GT/AG patterns are well known, but they are too short and ambiguous to define real splice junctions on their own.

For quantitative alternative splicing, the Discrete Compositional Energy Network models transcript probabilities through constituent splice junction energies, then maps transcript probabilities to exon-boundary inclusion values $3'$1. On CAPD, DCEN reached Spearman $3'$2 and Pearson $3'$3 on all withheld test genes, outperforming direct SpliceAI-style regression and ablation variants (Chan et al., 2021). This suggests that explicit composition from sites to junctions to transcripts can be advantageous when the target is isoform-aware regression rather than local classification.

Self-supervised learning has also been reformulated around splicing. IsoCLR treats alternative splice isoforms and homologous transcripts as contrastive “views” of the same functional object, and pre-trains on mature RNA isoforms with explicit splice-site and coding-sequence tracks. On downstream prediction of RNA half-life and mean ribosome load, the pre-training strategy yields competitive linear-probe performance and up to a two-fold increase in Pearson correlation in low-data conditions (Fradkin et al., 2023).

For graph-based transcript comparisons, recent work moves beyond binary events. A GrASE splicing graph is a DAG per gene; bubbles are source–sink subgraphs with at least two valid paths; and shared versus distinct exonic parts are computed directly from read-supported exonic part edges. The framework supports all-pairwise contrasts, a multinomial $3'$4-way comparison, and lower-set bipartitioning in a containment graph, specifically to handle complex bubbles where more than two transcript paths compete (Witoslawski et al., 8 Nov 2025).

3. Formal-language and graph-theoretic splicing

In formal language theory, splicing is a binary word operation inspired by DNA recombination under the action of restriction enzymes and ligases. The classical theory distinguishes Head/Păun splicing rules and Pixton splicing rules, and finite splicing systems generate regular languages; however, these languages form a proper subclass of the class of regular languages (Kari et al., 2011). A central structural result is decidability: if a regular language is generated by a splicing system, then it is also generated by a splicing system whose size is a function of the size of the syntactic monoid of the input language, and this bounded system can be effectively constructed (Kari et al., 2011).

Graph-theoretic generalization is motivated by the mismatch between linear strings and three-dimensional molecules. String splicing was introduced as an abstract model for DNA recombination, but earlier graph splicing systems were limited to particular graphs interpretable as linear or circular graphs. The “Graph Splicing System” introduces a splicing system for graphs that can be applied to all types of graphs, treating splicing as a new operation among graphs that generates many new graphs from the given two graphs, and studying graph-theoretical results of that operation [0702034].

These two traditions meet at the level of abstraction. In the formal-language setting, splicing is a rule-governed closure operation on sets of words. In the graph setting, it is a rule-governed closure operation on sets of graphs. A plausible implication is that the shift from strings to graphs preserves the recombinational intuition while widening the admissible ambient structure from one-dimensional sequences to arbitrary graph topologies.

4. Image splicing in digital forensics

In digital image forensics, image splicing is a local forgery in which a donor region from one image is pasted into a host image. One paper adopts the definition

$3'$5

where $3'$6 is the foreground authentic image, $3'$7 the background authentic image, and $3'$8 a binary mask designating the pasted region; the additional constraint

$3'$9

encodes that the spliced region covers a smaller area than the background region (Bi et al., 2023). This distinguishes splicing from copy-move, inpainting, and global enhancement.

A recurring difficulty is that detectors trained on semantically poor datasets may learn semantics rather than splicing traces. GreatSplicing was introduced specifically to address insufficient semantic varieties and inconsistent experimental settings: it comprises 5,000 spliced images, covers spliced regions with 335 distinct semantic categories, includes 2,887 object-aware spliced images and 2,113 shape-random spliced images, and yields lower misidentification on authentic images together with stronger cross-dataset detection than existing datasets (Bi et al., 2023). The same source proposes five benchmark modes, including Self-Sufficiency Mode, Cross-Dataset Validation Mode, Synthetic-Finetuning Mode, Large-Sample Mode, and Authentic-Image Introduction Mode.

Physical-consistency approaches define image splicing differently. One line of work detects splicing by verifying the consistency of lens radial distortion across an image. Using line-based calibration, the method estimates the first-order radial distortion parameter $3'$0 from straight edges in different regions, relying on the expectation that a genuine image taken by a single camera at a fixed zoom has globally consistent radial distortion behavior, whereas a splice disturbs that consistency (Chennamma et al., 2011). On the authors’ Spliced Image Data Set, the reported splicing detection rate is 86% (Chennamma et al., 2011).

These two forensic perspectives address a common misconception from opposite sides. Dataset-centered work argues that semantics can dominate learning unless the data are semantically rich. Optics-centered work argues that camera-intrinsic geometry can reveal splicing even when scene content appears visually coherent.

5. Splicing in topology and operad theory

In low-dimensional topology, splicing refers to gluing operations on knot exteriors and to operadic composition of embeddings. One paper introduces a topological operad $3'$1, the splicing operad, acting on

$3'$2

and shows that this action extends the action of the operad of $3'$3-cubes (Budney, 2010). In the case $3'$4, the action encodes Larry Siebenmann’s splicing construction for knots in $3'$5; moreover, the space of long knots in $3'$6, $3'$7, is free with respect to the splicing operad action, and the free generating space is the subspace $3'$8 of torus and hyperbolic knots (Budney, 2010).

A related but distinct use appears in quantum topology. If a rational homology 3‑sphere $3'$9 is obtained by gluing the exteriors of two framed knots $\gamma$0 and $\gamma$1, a splicing formula expresses the LMO invariant of $\gamma$2 in terms of the Kontsevich–LMO invariants of $\gamma$3 and $\gamma$4 (Massuyeau et al., 2020). In low degrees this recovers Fujita’s formula for the Casson-Walker invariant, and the same source observes that the second term of the Ohtsuki series is not additive under standard splicing (Massuyeau et al., 2020). The formula extends to the case where each $\gamma$5 comes with an additional link $\gamma$6, yielding a satellite formula for the Kontsevich–LMO invariant (Massuyeau et al., 2020).

The topological usage differs sharply from the biological and forensic ones, but a shared structural pattern remains visible: splicing is again a controlled cut-and-glue operation whose iteration is naturally organized by algebraic composition.

6. Binary re-linking and provenance in software package management

In Spack, splicing is a mechanism for re-linking already-built binaries so that some of their dependencies are replaced by other, ABI-compatible binaries, while retaining full provenance of how everything was actually built (Gouwar et al., 9 Sep 2025). A spliced spec is obtained by taking an existing, already-built spec and replacing one of its dependency nodes in the DAG—and transitively any common dependencies—with a different, ABI-compatible spec. The result has a runtime view, which says that the binary uses dependency $\gamma$7 instead of $\gamma$8, and a build spec reference, which says that the binary was actually compiled against $\gamma$9 (Gouwar et al., 9 Sep 2025).

The compatibility model is declarative rather than inferred. Packagers add can_splice(target_spec, when=source_spec) rules to the packaging DSL, and the ASP-based concretizer chooses whether to keep the original dependency hash or to execute a splice (Gouwar et al., 9 Sep 2025). This directly addresses a likely misconception: splicing in Spack is not automatic ABI detection. ABI compatibility still comes from developers’ knowledge, but the solver can exploit it during dependency resolution and installation (Gouwar et al., 9 Sep 2025).

The main motivation is the gap between binary and source package managers in HPC. Spack can build from source and use buildcaches, but without a binary compatibility model it cannot mix binaries that were not built together. Splicing augments Spack’s packaging language and dependency resolution engine so that compatible binaries can be reused while maintaining the flexibility of source builds, including for ABI-sensitive dependencies such as MPI (Gouwar et al., 9 Sep 2025). The reported overheads are modest relative to build-time savings: with the new encoding but splicing disabled, average concretization time increases by about 4.7% with a local buildcache and about 7.1% with a public buildcache; with splicing enabled for MPI-dependent RADIUSS specs, the increases are 17.1% with the local buildcache and about 153% with the public buildcache (Gouwar et al., 9 Sep 2025).

Across these domains, splicing consistently names a structured recombination operation rather than a single discipline-specific mechanism. In RNA processing it joins or redefines transcript segments; in formal systems it closes languages or graph families under recombination; in image forensics it denotes the compositing act to be detected; in topology it glues exteriors and composes embeddings; and in software infrastructure it re-links dependency graphs while preserving provenance. The sources together suggest that the persistence of the term reflects a common operational schema—local cut, controlled reconnection, and global consequences—even when the underlying objects are nucleic acids, words, graphs, images, manifolds, or binaries.