KMA: A Polysemous Research Acronym

• KMA is a polysemous acronym that refers to different technical systems depending on the context, ranging from a bioinformatics toolkit to meteorological data provision and specialized algorithms.
• In bioinformatics, KMA (Keep Me Around) enables intron retention analysis in RNA-seq by quantifying synthetic intronic sequences and employing replicate-aware significance tests.
• In environmental forecasting, KMA denotes the Korea Meteorological Administration, which supplies real-time data essential for water-temperature prediction and balanced radar nowcasting.

KMA is a strongly polysemous research acronym whose meaning depends entirely on disciplinary context. In recent arXiv literature, it denotes at least six distinct entities: a bioinformatics toolkit for intron retention analysis, the Korea Meteorological Administration as an operational data provider, a device-clustering subroutine for decentralized edge learning, a kernelized multilogit classifier for microarray data, the Kirchhoff migration algorithm in synthetic aperture radar imaging, and a chemically amplified nanobot coordination algorithm for diffuse cancer treatment (Pimentel et al., 2015, Lee et al., 2015, Yang et al., 2020, Wagala et al., 2019, Wang et al., 30 Jun 2025, Harasha et al., 8 Sep 2025). The acronym therefore has no stable cross-domain definition; its interpretation is fixed by local methodological vocabulary such as RNA-seq quantification, meteorological nowcasting, Non-IID edge learning, near-range SAR, or chemotactic swarm control.

1. Acronymic scope in the literature

The primary encyclopedic fact about KMA is that it is not a single concept but an overloaded label reused across unrelated technical traditions. In computational biology it appears as Keep Me Around, a toolkit for retained-intron detection from RNA-seq (Pimentel et al., 2015). In environmental informatics and radar nowcasting, it denotes the Korea Meteorological Administration, an institutional data source used for real-time weather-driven forecasting and for construction of a balanced radar benchmark (Lee et al., 2015, Song et al., 5 Feb 2026). In machine learning, it appears both as a Node Clustering Algorithm based on K-Means and Average Accuracy inside E-Tree decentralized learning and as the Kernel Multilogit Algorithm for microarray classification (Yang et al., 2020, Wagala et al., 2019). In microwave imaging, it denotes the Kirchhoff migration algorithm; in nanorobotics, it denotes an algorithm in which bots carry therapeutic chemical $K$ together with an attractive amplification signal $A$ (Wang et al., 30 Jun 2025, Harasha et al., 8 Sep 2025).

This multiplicity is not superficial. Some senses of KMA name full software systems, some name a government agency, and some name narrowly defined algorithmic subroutines. A plausible implication is that acronym-only citation is unusually error-prone here. Precise expansion is often indispensable, especially because nearby acronyms such as KEMA, K-MACE, and KDMA refer to different methods rather than variant spellings of KMA.

2. KMA as “Keep Me Around” in intron retention analysis

In bioinformatics, KMA refers to “Keep Me Around”, a toolkit for detecting and analyzing intron retention (IR) in RNA-seq experiments (Pimentel et al., 2015). Its design goal is to bridge transcript quantification pipelines and downstream IR analysis without introducing a wholly separate alignment-and-counting framework. The preprocessing stage identifies “measurable” intronic regions, called inclusion regions, defined as intronic segments for which none of the overlapping isoforms contains an exon. It also records the corresponding overlap isoforms, i.e. transcript isoforms that could retain the intron. Preprocessing outputs an intron–transcript relationship table containing intron coordinates, intron quantification coordinates, transcripts that could potentially retain the intron, and the gene name; it also emits a BED track for quantification intervals and a FASTA file of intronic pseudo-transcript sequences.

A central implementation detail is the distinction between exact intron coordinates and intron quantification coordinates. KMA extends each intron into neighboring exons by a small amount, chosen to be several bases shorter than the read length, so that reads spanning intron–exon junctions can support the retained-intron isoform. The synthetic intronic sequences are then added to the ordinary transcriptome, and the combined reference is quantified by an existing transcript quantifier. The paper states that KMA “currently uses the transcript quantification method eXpress,” but is compatible with any RNA-seq quantification pipeline that outputs expression in an additive unit such as TPM (Pimentel et al., 2015). This lets KMA exploit established probabilistic assignment of ambiguous and multi-mapping reads rather than relying only on uniquely assigned reads.

In the R package, KMA computes a PSI-like retention quantity from intron abundance and the abundance of overlapping spliced isoforms. For intron $i$,

$$\mathrm{retention}_i = \frac{E(I_i)}{E(I_i)+\sum_{t\in O_i}E(t)},$$

where $E(I_i)$ is the expression of intron $i$ and $O_i$ is the set of overlap isoforms. The data are stored in an IntronRetention object, but operationally the representation is a database-style table in which each row is one intron observation in one sample, with fields such as intron, sample, condition, retention, numerator, denominator, unique_reads, and filter columns. This tabular design supports aggregation across introns, genes, samples, and conditions.

KMA incorporates several reliability filters. These include coverage filters based on relative expression or rank and a zero coverage filter that finds the longest contiguous intronic region with no read starts, denotes its length by $Z$, and removes the intron if the probability of observing such a no-read-start region at the estimated expression level is low. Its most distinctive methodological contribution is a replicate-aware significance procedure: within a condition, filtered retention values are resampled $B$ times by randomly selecting one retention value from each sample, a null distribution of replicate means is formed, and the observed mean is tested against that null. The method is explicitly one-sided for elevated retention and is intended to reduce false positives caused by noise, low coverage, mapping ambiguity, or contamination by incompletely processed RNA (Pimentel et al., 2015). The paper does not provide extensive head-to-head benchmarking or fixed default thresholds, so practical cutoff choice remains analyst-dependent.

3. KMA as the Korea Meteorological Administration

In environmental prediction and radar nowcasting, KMA denotes the Korea Meteorological Administration, an operational provider of meteorological and radar data rather than an algorithmic method. In the WT-Agabus water-temperature study, KMA is the decisive real-time data source because the authors restrict the predictive model to inputs that are available online from KMA: air temperature ($TM$) and rainfall ($A$0) (Lee et al., 2015). Historical data from the Inje Meteorological Station over 1991–2009 were used for model development, and lag-correlation analysis selected the ANN input triplet $A$1, $A$2, and $A$3. KMA data are obtained through a RESTful web service, transformed into logical sensor streams within CSN, processed by SPID using Esper and EPL, and passed to a MATLAB feed-forward ANN with 3 input variables, 1 hidden layer with 10 neurons, sigmoid hidden activation, linear output activation, and learning rate $A$4. The reported test performance was $A$5, $A$6, $A$7, and $A$8, with the authors emphasizing that the model is constrained by online availability of KMA variables rather than by maximal predictor richness (Lee et al., 2015).

A second, distinct KMA usage appears in precipitation nowcasting, where the acronym again refers to the agency and its products. The exPreCast study constructs a balanced radar dataset from KMA Hybrid Surface Rainfall (HSR) composite radar images obtained through the KMA API hub (Song et al., 5 Feb 2026). The raw HSR product has 5-minute temporal resolution, $A$9 coverage, and 500 m spatial resolution. For the benchmark, images from 2014 to 2023 were collected at 10-minute intervals, negative values were clipped to 0 and divided by 10,000, the data were downsampled from 500 m to 4 km by uniform subsampling, and a central $i$0 crop was extracted, yielding $i$1 model inputs. The split is strictly temporal: training on 2014–2021, validation on 2022, and testing on 2023. Radar values are converted to rainfall rate using the Marshall–Palmer $i$2–$i$3 relation, and evaluation uses thresholds $i$4 mm/h. The paper’s central claim is that this KMA dataset is more balanced across ordinary and extreme rainfall regimes than SEVIR or MeteoNet, making it better suited for realistic operational evaluation (Song et al., 5 Feb 2026).

Across both papers, KMA is thus an institutional backbone for real-time forecasting pipelines. One study uses KMA’s meteorological feed to ensure online executability of a water-temperature predictor; the other uses KMA radar products to construct a decade-scale benchmark that spans both normal precipitation and monsoon- or typhoon-associated extremes.

4. Algorithmic KMA in machine learning and statistical classification

One machine-learning sense of KMA appears in decentralized edge AI. In the E-Tree framework, KMA is defined as “Node Clustering Algorithm based on K-Means and Average Accuracy” (Yang et al., 2020). Its purpose is to cluster edge devices before hierarchical aggregation so that the resulting tree is suitable for Non-IID data and heterogeneous communication delays. Inputs include the number of clusters $i$5, topology graph $i$6, nodes $i$7, a common test set $i$8, threshold $i$9, and pre-training rounds $$\mathrm{retention}_i = \frac{E(I_i)}{E(I_i)+\sum_{t\in O_i}E(t)},$$0. The algorithm computes shortest-path transmission delays $$\mathrm{retention}_i = \frac{E(I_i)}{E(I_i)+\sum_{t\in O_i}E(t)},$$1, pre-trains local models, evaluates each node’s pre-trained accuracy $$\mathrm{retention}_i = \frac{E(I_i)}{E(I_i)+\sum_{t\in O_i}E(t)},$$2 on $$\mathrm{retention}_i = \frac{E(I_i)}{E(I_i)+\sum_{t\in O_i}E(t)},$$3, computes the graph-wide average

$$\mathrm{retention}_i = \frac{E(I_i)}{E(I_i)+\sum_{t\in O_i}E(t)},$$4

and attempts to form clusters whose average pre-trained accuracy $$\mathrm{retention}_i = \frac{E(I_i)}{E(I_i)+\sum_{t\in O_i}E(t)},$$5 satisfies

$$\mathrm{retention}_i = \frac{E(I_i)}{E(I_i)+\sum_{t\in O_i}E(t)},$$6

Assignment is restricted to the nearest $$\mathrm{retention}_i = \frac{E(I_i)}{E(I_i)+\sum_{t\in O_i}E(t)},$$7 candidate centers by delay whenever possible, and cluster centers are updated by a medoid-like rule

$$\mathrm{retention}_i = \frac{E(I_i)}{E(I_i)+\sum_{t\in O_i}E(t)},$$8

The paper presents KMA as a heuristic constrained clustering procedure, gives its complexity as $$\mathrm{retention}_i = \frac{E(I_i)}{E(I_i)+\sum_{t\in O_i}E(t)},$$9, and uses it to shape lower layers of the E-Tree hierarchy (Yang et al., 2020). Here, “KMA” is neither a generic k-means variant nor the whole E-Tree method; it is a specific data-aware, network-aware clustering engine.

A different statistical-learning sense appears in microarray classification, where KMA denotes the Kernel Multilogit Algorithm (Wagala et al., 2019). This method maps categorical class probabilities into a multilogit coordinate system, fits a regularized regression model in that transformed space, and then maps predictions back to class probabilities via the inverse multilogit. With $E(I_i)$0 classes and class $E(I_i)$1 as reference, the transformed response is

$E(I_i)$2

and the ridge-type objective is

$E(I_i)$3

Kernelization yields the dual solution

$E(I_i)$4

The paper adopts KMA from Dalmau et al. (2015) as a benchmark classifier and reports that, on the Colon microarray dataset of size $E(I_i)$5, KMA achieved the lowest classification error rate in both reported settings: 1.7 on un-preprocessed data and 11.2 on preprocessed data (Wagala et al., 2019). In this literature, KMA is therefore a kernelized probabilistic classifier, not a clustering routine.

5. KMA in imaging and nanobot control

In handheld microwave imaging, KMA denotes the Kirchhoff migration algorithm (Wang et al., 30 Jun 2025). The paper on near-range 3-D reconstruction treats KMA and the backprojection algorithm (BPA) as standard time-domain imaging algorithms for arbitrary handheld synthetic-aperture trajectories. Their defining advantage is geometric fidelity: because they evaluate contributions using actual sensor positions $E(I_i)$6 and exact propagation distances $E(I_i)$7, they can compensate trajectory deviations at each scan position and remain applicable to non-uniform synthetic arrays. Their defining limitation is computational cost. The paper states that time-domain imaging algorithms including BPA and KMA are widely adopted because of their direct applicability to arbitrary scan trajectories, but “suffer from time complexity issues that hinder their practical application” (Wang et al., 30 Jun 2025). It therefore proposes HHFFBPA as a fast factorized substitute. Notably, the paper does not derive KMA explicitly or benchmark it numerically; KMA is a reference algorithmic category within the imaging landscape rather than the primary mathematical object of that study.

In nanorobotic cancer treatment, KMA has yet another meaning. The paper on diffuse cancer defines three swarm-control algorithms: KM, KMA, and KMAR (Harasha et al., 8 Sep 2025). Here KMA is the variant in which agents carry therapeutic chemical $E(I_i)$8 and an additional attractive payload $E(I_i)$9 that amplifies the natural marker signal $i$0 emitted by cancer sites. The total sensed field is

$i$1

for KMA, and bots follow noisy gradient ascent on this combined field. Upon reaching a site, a bot always drops both $i$2 and $i$3, then terminates. The purpose of $i$4 is to create a positive-feedback cascade: once a site is found, future bots are attracted more strongly to it. The paper’s main conclusion is a speed–allocation tradeoff. KMA significantly improves treatment speed but often reduces eventual treatment success because excessive convergence to one site wastes payloads when cancer is diffuse; the exception is concentrated-demand patterns, where the cascade is advantageous (Harasha et al., 8 Sep 2025). The study reports that increasing $i$5 can make treatment up to 30 times faster while making final success up to 40% worse, illustrating that KMA is an amplification mechanism rather than a balanced allocation mechanism.

6. Neighboring acronyms and recurrent misconceptions

A common source of confusion is the proximity of KMA to several distinct acronyms that are methodologically related only in the broadest sense. KEMA is Kernel Manifold Alignment, a semi-supervised multi-domain manifold-alignment and domain-adaptation method that aligns several domains in a common latent space using block-diagonal kernels and graph Laplacians, and it is explicitly presented as a kernelized generalization of SSMA rather than as a KMA variant (Tuia et al., 2015). K-MACE or KMACE is k-minimizing Average Central Error, a clustering-validity and correct-number-of-clusters estimator for k-means and kernel k-means, again distinct from all KMA senses (Beheshti et al., 2019). KDMA is the Knowledge-driven Memetic Algorithm for the energy-efficient distributed homogeneous flow shop scheduling problem, with collaborative initialization, PMX crossover, swap mutation, key-factory local search, and a carbon-reduction rule based on machine shutdown during sufficiently long idle periods (Xu et al., 2024).

These nearby names matter because they demonstrate that acronym similarity is not evidence of conceptual lineage. KEMA concerns manifold alignment, K-MACE concerns cluster-number estimation, and KDMA concerns multi-objective scheduling. Their resemblance to KMA is bibliographic rather than substantive. The most reliable scholarly practice is therefore to expand the acronym on first use and to specify the field immediately: Keep Me Around in RNA-seq, Korea Meteorological Administration in forecasting, Node Clustering Algorithm based on K-Means and Average Accuracy in E-Tree, Kernel Multilogit Algorithm in classification, Kirchhoff migration algorithm in SAR, or the $i$6 nanobot strategy in diffuse cancer treatment.