The core problem COSIGT addresses is one of depth. Recent methods such as Locityper have already shown that pangenome references, which store many haplotypes as paths through a graph rather than a single consensus sequence, can dramatically improve targeted genotyping at difficult loci when read data are abundant. Locityper aligns reads to locus-specific haplotypes and selects the best-fitting diploid pair by optimizing alignment accuracy, insert-size concordance, and coverage balance. But that alignment-based likelihood machinery has an Achilles heel: as sequencing coverage falls, the signal-to-noise ratio degrades and genotyping accuracy drops with it. That matters enormously for population-scale studies and biobank cohorts sequenced at variable and often shallow depth, and it matters even more for ancient DNA, where postmortem degradation and microbial contamination routinely push effective coverage below 2X.

COSIGT takes a fundamentally different mathematical route. The pipeline first constructs a local variation graph for the target locus from haplotype-resolved assemblies, rather than querying an entire genome-wide pangenome. This locus-specific design allows parameter tuning, rapid incorporation of new assemblies, and faster genotyping. Reads pre-aligned to the region are extracted, supplemented with unmapped reads rescued by a k-mer filtering tool called kfilt, and then mapped to the local graph. From the graph nodes that reads traverse, the pipeline builds a coverage vector for the sample, in which each element reflects length-normalized, multi-mapping-aware coverage at a node. Every haplotype in the graph is likewise represented as a coverage vector of node traversal counts. COSIGT then enumerates all possible diploid haplotype pairs, sums their vectors to create synthetic genotype profiles, and computes the cosine similarity between each synthetic profile and the observed sample vector. The pair with the highest similarity wins.

The elegance of the approach lies in what cosine similarity measures. Because the metric captures the orientation of a vector rather than its magnitude, it is inherently invariant to overall sequencing depth. A sample sequenced at 1X coverage produces the same relative coverage profile shape as one sequenced at 30X, merely scaled down, and cosine similarity is blind to that scaling. Likelihood-based methods, by contrast, depend on absolute read counts and lose power as counts dwindle. In benchmarks across 326 challenging medically relevant genes and 265 structurally variable regions using short-read data from the 1000 Genomes Project, with pangenome graphs built from assemblies of the Human Pangenome Reference Consortium and the Human Genome Structural Variation Consortium, both COSIGT and Locityper performed well at 5X and 30X coverage, with more than 93 percent high-quality calls. Locityper retained an edge at 30X, reaching 98.2 percent versus 93.9 percent for COSIGT at the medically relevant genes.

At low coverage, the picture reversed dramatically. At 1X, COSIGT delivered 93.4 percent of calls at mid-or-higher quality compared with 84.5 percent for Locityper, and at 2X the gap persisted at 95.8 percent versus 93.4 percent. The advantage became even starker on simulated ancient DNA. Across 48 medically relevant genes enriched for pharmacogenetic and immunogenetic content, including HLAs, CYPs, and mucins, COSIGT maintained roughly 94 percent mid-or-higher quality calls at 1X while Locityper fell to about 46 percent; at 2X the figures were roughly 95 percent versus 60 percent. COSIGT outperformed Locityper at all 48 genes and showed only marginal degradation relative to matched modern DNA at the same coverage. The simulations, generated with a purpose-built simulator called ancestralsim, incorporated realistic ancient DNA damage patterns, short fragment lengths, and contamination levels of 0 or 10 percent, making the result a demanding test.

Robustness to missing references was assessed through leave-all-out benchmarks, in which the true haplotypes of each sample were excluded from the pangenome graph. Even then, COSIGT remained near-optimal, achieving at least 87 percent of the best possible quality at most loci, with 87.6 percent of medically relevant genes and 90.8 percent of structural variant regions falling in the top quintile of achievable quality. An ancestry-mismatched test, in which Peruvian and Colombian samples were genotyped against a graph containing no admixed American assemblies, yielded 87.2 percent high-quality calls at the medically relevant genes, suggesting the method degrades gracefully when the reference panel does not perfectly match the population under study.

The team also demonstrated the tool at genuine population scale. Applying COSIGT to 1,085 whole-genome samples from the Italian Moli-sani cohort, sequenced at roughly 20X coverage, the researchers performed HLA typing across seven classical HLA genes and achieved a mean haplotype-level accuracy of 89.5 percent against types imputed with HLA*IMP:02 from microarray data, comparable to the 90.8 percent achieved by the dedicated HLA genotyping tool T1K. More broadly, the authors report having applied COSIGT to more than 6,000 modern and ancient human genomes, demonstrating population-scalable analysis of complex repeats and multi-copy genes. The pipeline is implemented in Snakemake with containerized deployment, parallelizes across regions and samples, and its per-sample genotyping steps are computationally light, with a median runtime of about 0.02 minutes and median memory use of 52 megabytes for the core genotyping step.

The authors are candid about limitations. Genotyping accuracy ultimately depends on the quality and completeness of the input pangenome: haplotypes absent from the reference panel cannot be called exactly, and novel structural variants will be assigned to the most similar available haplotype. The leave-all-out benchmarks quantify how much accuracy survives this constraint, but they also make clear that expanding pangenome references remains essential. COSIGT currently requires pre-aligned BAM or CRAM files, creating a dependency on reference-based alignment, though future work aims to accept raw FASTQ input directly and to implement haplotype subsampling to keep the method scalable as pangenomes grow. The framework also currently supports diploid genotyping, although the underlying mathematics generalizes to arbitrary ploidy by enumerating k-tuples of haplotypes.

The implications reach well beyond a single software release. Pangenome references are expanding rapidly in size, quality, and taxonomic breadth, and as they do, the bottleneck shifts from reference completeness to scalable genotyping. By making complex-locus genotyping reliable at 1 to 2X coverage, COSIGT opens the door to systematically mining the vast archives of low-coverage sequencing data already sitting in biobanks and legacy datasets, and to bringing archaeological and ancient genomes into population-level biomedical analyses from which their shallow depth previously excluded them. Larger cohorts become affordable at reduced sequencing cost, samples of heterogeneous quality can be combined in a single analysis, and the allelic diversity captured in pangenome references becomes practically accessible across the full spectrum of sequencing depths. The pipeline is compatible with long-read input as well, which could extend its utility to low-coverage long-read datasets too sparse for assembly. For a field that has long treated structurally complex loci as terra incognita in shallow data, the message is that the map, and the compass to read it, may finally have arrived together.

Subject of Research: Pangenome graph-based genotyping of structurally complex genomic loci from low-coverage sequencing data

Article Title: COSIGT: population-scalable genotyping of complex loci from low-coverage sequencing data using pangenome graphs

Article References: Bolognini, D., Guarracino, A., Paleni, C., Dudley, T. S., Iacoviello, L., Raveane, A., Sudmant, P. H., Garrison, E., & Soranzo, N. (2026). COSIGT: population-scalable genotyping of complex loci from low-coverage sequencing data using pangenome graphs. Genome Biology, 27(1), Article 286. https://doi.org/10.1186/s13059-026-04242-4

Image Credits: AI Generated

DOI: 10.1186/s13059-026-04242-4

Keywords: pangenome graphs, genotyping, low-coverage sequencing, ancient DNA, structural variation, copy-number variation, HLA typing, CYP2D6, cosine similarity, population genomics, Genome Biology, bioinformatics

The research, conducted by Saloua Lagnaoui and Khalid Haddouch of the Laboratory of Applied Sciences and Emerging Technologies at ENSA, Sidi Mohamed Ben Abdellah University in Fez, Morocco, together with Zakariae En-naimani of the Laboratory Computer Science, Artificial Intelligence and Cyber Security at ENSET, Hassan II University of Casablanca, introduces a fundamentally different way of thinking about redundancy. Rather than waiting until training is complete and then surgically removing filters, or randomly resetting them in the hope that diversity will emerge, the proposed method models the relationships between filters dynamically during training itself. It continuously measures how similar the responses of different filters are and applies similarity-aware weighting that dampens redundant feature responses before they can dominate the network’s internal representations.

To understand why this matters, it helps to consider what filters actually do inside a convolutional neural network. Each filter is a small pattern detector that slides across an image, responding strongly to particular visual features such as edges, textures, color gradients, or the characteristic lesions and discolorations that signal disease on a leaf. In an ideal network, every filter would specialize in a distinct feature, collectively forming a rich and diverse vocabulary of visual descriptors. In practice, however, gradient-based training frequently drives multiple filters toward nearly identical behavior. When that happens, the network wastes computational capacity re-learning the same feature, its effective representational power shrinks, and its convergence during training can become unstable as redundant filters compete for the same gradient signals.

Previous approaches to this problem have generally fallen into two camps. Pruning techniques, such as those that rank filters by importance and remove the least useful ones after or during training, treat redundancy as something to be excised. Stochastic resetting methods, which randomly reinitialize filters mid-training, gamble that randomness will restore diversity. Both strategies have drawbacks: pruning can be brittle and often requires careful tuning and retraining, while stochastic approaches introduce nondeterminism that makes results harder to reproduce and optimize. The CNN-SA-RFR framework, by contrast, is deterministic and adaptive. By embedding a cosine similarity measure into a self-attention mechanism, it quantifies the angular alignment between filter responses and uses that information to reweight the network’s feature maps continuously, suppressing responses that duplicate what other filters already capture.

The choice of cosine similarity is technically significant. Unlike Euclidean distance, which is sensitive to the magnitude of activation vectors, cosine similarity measures the orientation between vectors, making it a natural gauge of whether two filters are responding to the same underlying pattern regardless of response strength. Combined with self-attention, a mechanism popularized by the transformer architecture that allows a model to weigh the relationships among different elements of its input, this creates a layer that is acutely aware of its own internal redundancy. The attention weights effectively act as a soft, learned form of redundancy control: filters whose responses are highly similar to others receive diminished influence, while distinctive, informative features are amplified and allowed to propagate through the network.

The authors evaluated their framework on two publicly available benchmark datasets: Plant Pathology 2020, released for the FGVC7 Kaggle competition, and the Plant Disease Recognition dataset, also hosted on Kaggle. These datasets present genuinely difficult classification challenges, requiring models to distinguish healthy leaves from those affected by multiple diseases under varying lighting conditions, viewing angles, and stages of disease progression. The results were striking. The proposed method achieved up to 78 percent accuracy on the Plant Pathology 2020 dataset and 94 percent on the Plant Disease Recognition dataset, outperforming baseline convolutional neural network models on both benchmarks.

Beyond raw accuracy, the study reports improvements in three additional dimensions that matter greatly for practical deployment. Feature diversity increased, indicating that the network’s filters genuinely specialized rather than duplicating one another, which suggests the similarity-aware weighting is doing exactly what it was designed to do. Convergence stability also improved, meaning training runs were less prone to the oscillations and plateaus that plague redundant networks, a benefit that translates directly into reduced engineering time and compute cost. Together, these gains position the method as a practical optimization strategy rather than a laboratory curiosity, particularly for agricultural applications where models may need to run on edge devices with limited processing power.

The work builds on a growing body of research at the intersection of attention mechanisms and efficient convolutional networks. Self-attention has already proven its value in plant disease recognition, with prior studies showing that attention-enhanced convolutional models can better localize disease symptoms and ignore background clutter. At the same time, research into kernel and filter redundancy reduction has explored low-rank expansions, knowledge distillation, and differentiable pruning masks. What distinguishes CNN-SA-RFR is the fusion of these two threads: it uses attention not to improve the network’s view of the input image, but to improve the network’s view of itself, continuously auditing its own filters for redundancy and correcting it in real time.

The implications extend well beyond plant pathology. Filter redundancy is a general problem in convolutional architecture design, affecting applications from medical imaging to autonomous driving, and any deterministic, training-time method that mitigates it without sacrificing accuracy is of broad interest to the machine learning community. For agriculture specifically, the stakes are high. Plant diseases cause substantial global crop losses each year, and smartphone-based diagnostic tools powered by efficient CNNs are increasingly seen as a frontline defense for farmers who lack access to expert agronomists. Leaner, more diverse, and more stable networks make such tools cheaper to train, faster to run, and more reliable in the field.

The Moroccan team’s findings also carry a methodological lesson for the wider deep learning community: architecture and training procedure need not be designed independently. By making redundancy awareness an intrinsic property of the network layer rather than an external post-processing step, CNN-SA-RFR suggests a future in which models are not merely trained and then optimized, but are self-optimizing from the first forward pass. The datasets used in the study are publicly available, and the authors note that the code can be obtained from the corresponding author, opening the door for other researchers to replicate, refine, and extend the approach across new domains and larger architectures.

Subject of Research: A cosine similarity-based self-attention framework for reducing filter redundancy in convolutional neural networks applied to plant disease classification

Article Title: CNN-SA-RFR: A cosine similarity-based self-attention approach for reducing filter redundancy in CNNs for plant disease classification

Article References: Lagnaoui, S., En-naimani, Z., & Haddouch, K. (2026). CNN-SA-RFR: A cosine similarity-based self-attention approach for reducing filter redundancy in CNNs for plant disease classification. Multimedia Tools and Applications, 85(10), Article 764. https://doi.org/10.1007/s11042-026-21921-3

Image Credits: AI Generated

DOI: 10.1007/s11042-026-21921-3

Keywords: convolutional neural networks, self-attention, filter redundancy, cosine similarity, plant disease classification, deep learning, computer vision, agricultural AI, model efficiency, feature diversity, Plant Pathology 2020, neural network pruning