DNA augmentations for semi-supervised learning in genomic sequence classification Authors: Hyunwoo Yoo, Gail L Rosen Published: 2026 ABSTRACT AbstractSemi-supervised learning (SSL) is a natural fit for genomics, where unlabeled sequences are abundant but high-quality labels are scarce due to costly curation. However, most SSL pipelines rely on generic perturbations that are not tailored to biological constraints, and it remains unclear which augmentation designs yield reliable gains for DNA sequence classification. In this application note, we present an empirical study of genome-aware weak–strong augmentation recipes for pseudo-label-based SSL, focusing on FixMatch and FlexMatch with a GROVER DNA language model backbone. Across two classification benchmarks—antimicrobial resistance (AMR) drug-class prediction from AMR gene sequences and multi-class oncovirus family classification from simulated 150 bp tumor-like reads—we systematically evaluate 25 weak/strong augmentation pairs constructed from biologically motivated operators, including reverse-complement, codon-level back-translation, mild masking/substitution (NN), high-rate substitution, and length-preserving insertion–deletion (indel). We find consistent trends: (i) moderate, sequence-preserving perturbations yield the best accuracy–stability trade-off, (ii) excessive masking or aggressive mutation increases variance and degrades performance, and (iii) short-segment masking (NN) for weak augmentation followed by an indel rate (around 5%) for strong augmentation emerge as a robust strategy to learn from unlabeled genomic data. We release implementations and recommended defaults to facilitate reproducible SSL for genomic sequence modeling under limited-label regimes. Introduction High-throughput sequencing technologies have made large-scale genomic data readily available, yet high-quality labels remain scarce, expensive, and often highly imbalanced [1]. This mismatch naturally motivates the use of semi-supervised learning (SSL), where a small labeled set is augmented with abundant unlabeled sequences to improve predictive performance [2, 3]. In recent years, DNA language models have shown strong potential for sequence-level genomic tasks, including antimicrobial resistance (AMR) classification and viral detection [4–9]. However, effective SSL for genomic sequences remains underexplored, particularly with respect to the design of biologically meaningful data augmentations [10, 11]. A central component of modern SSL frameworks such as FixMatch and FlexMatch is the construction of paired weak and strong views of the same input sequence [2, 3]. In vision and natural language processing, these views are typically generated using generic perturbations (e.g. cropping, color jittering, or word dropout) [12, 13]. In genomics, however, naive adaptations of such strategies risk violating biological constraints, distorting functional motifs, or introducing unrealistic sequence artifacts [14, 15]. As a result, it remains unclear which augmentation operators—and at what strength—are appropriate for DNA sequences in SSL settings. Existing genomic SSL studies often adopt a small set of heuristic perturbations, such as reverse-complement transformation or random masking [16, 17], without systematically evaluating their interactions or statistical reliability. Moreover, biologically motivated transformations such as codon-level back-translation (BT), insertion–deletion (indel) operations, or controlled mutation rates have rarely been analyzed in a unified framework [14, 15, 18]. This lack of empirical guidance makes it difficult for practitioners to design stable and effective SSL pipelines for genomic classification tasks, especially under limited-label and long-tailed regimes. In this application note, we present a focused empirical study of genome-aware augmentation strategies for pseudo-label-based SSL. Using a GROVER DNA language model backbone, we evaluate FixMatch and FlexMatch across two representative genomic classification tasks: AMR drug-class prediction from gene sequences and multi-class oncovirus family classification from simulated tumor-like short reads. We systematically construct and evaluate 25 weak–strong augmentation combinations derived from biologically plausible operators, including reverse-complement, codon-level BT, mild masking and substitution, high-rate mutation, and length-preserving indel perturbations. Our results reveal consistent and statistically supported trends. Moderate, sequence-preserving perturbations achieve the best accuracy–stability trade-off, while overly aggressive masking or mutation degrades performance and increases variance across random seeds. In particular, short-segment severe mutations through masking (NN) and low-rate indel operations emerge as robust defaults for strong augmentation, outperforming synonymous mutations from BT-only in both tasks. We further quantify these differences using Friedman–Nemenyi post-hoc tests to distinguish statistically meaningful improvements from stochastic variation (in Supplement S6). Rather than proposing a new SSL objective, our goal is to provide practical, empirically grounded guidance for designing augmentation pipelines in genomic SSL. To support reproducibility and reuse, we release our augmentation implementations, configurations, and recommended defaults. We expect these findings to serve as a useful reference for applying SSL to DNA sequence modeling under realistic, label-limited conditions. Materials and methods Datasets We evaluate genome-aware SSL on two representative DNA sequence classification benchmarks that differ in task structure and difficulty [19]. The first benchmark focuses on AMR drug-class prediction from gene sequences [20, 21], where each sequence is labeled with its associated resistance drug class. This task reflects a biologically meaningful, sequence-level classification setting with limited labeled data and substantial class imbalance. The second benchmark targets multi-class oncovirus family classification from simulated tumor-like short reads [7]. Following prior work, we generate fixed-length reads of 150 bp to emulate realistic sequencing fragments encountered in cancer genomics pipelines. Each read is labeled with its corresponding viral family, resulting in a more challenging short-context classification problem with higher ambiguity and noise. For both benchmarks, we construct semi-supervised splits consisting of a small labeled subset and a substantially larger unlabeled pool [10, 11]. All sequences are treated independently at the sequence level; no sample-level aggregation or abundance weighting is used. This design isolates the behavior of DNA language models under sequence-level pseudo-label-based SSL, without introducing confounding effects from host-level or sample-level labels. Detailed dataset statistics and class distributions are provided in Supplementary material Tables S1 and S2. DNA language model backbone All experiments are conducted using a GROVER DNA language model backbone, which encodes input DNA sequences into contextualized representations suitable for downstream classification [1, 4]. The backbone is initialized from pretrained weights and fine-tuned end-to-end within the SSL framework. A linear classification head is attached to the pooled sequence representation to predict the target class for each task. We adopt GROVER as the backbone due to its ability to capture contextual dependencies in genomic sequences and its availability as a pretrained DNA language model. Its architecture is well suited for fine-tuning under SSL settings and enables systematic evaluation of augmentation strategies without introducing additional architectural complexity. At a first glance, the self-supervised backbone LLM combined with fine-tuning may seem like “semi-supervised” learning, but the definition of this type of setup is transfer learning. In contrast, explicitly pseudo-labeling unlabeled data is considered SSL. Although GROVER is pretrained using self-supervised objectives, its pretrained weights are used in our framework to initialize transfer learning for the downstream classification tasks. Training this backbone with labeled samples alone corresponds to supervised fine-tuning. In contrast, FixMatch and FlexMatch introduce a separate task-level semi-supervised fine-tuning stage that explicitly uses both labeled and unlabeled samples. LoRA provides a parameter-efficient means of adapting the pretrained backbone, whereas the semi-supervised nature of the proposed framework arises from the use of unlabeled samples in the downstream objectives. Semi-supervised learning frameworks We adopt two widely used pseudo-label-based SSL algorithms: FixMatch and FlexMatch. Both methods rely on paired weak and strong augmentations of the same input sequence. Predictions generated from weakly augmented views are used to produce pseudo-labels, which are then enforced as consistency targets on strongly augmented views. FixMatch applies a fixed confidence threshold to select reliable pseudo-labels, whereas FlexMatch dynamically adjusts class-wise thresholds based on pseudo-label statistics. We further analyze pseudo-label dynamics across training in Supplementary Material S5, with detailed statistics provided in Supplementary Tables S9 and S10. Aside from the thresholding mechanism, both frameworks share the same optimization objective and differ only in their pseudo-label selection strategy. Genome-aware augmentation operators We define a set of biologically inspired DNA augmentation operators designed to respect core sequence constraints while introducing controlled perturbations [14, 15, 18]. These include reverse-complement transformation, codon-level BT, short-segment masking and substitution (NN), high-rate random substitution (mutation), length-preserving insertion–deletion (indel), and a hybrid indel+NN variant that combines structural perturbation with masking. Each operator is parameterized to control perturbation strength, enabling systematic exploration of weak–strong augmentation pairs. Detailed parameterizations, such as masking rate, mutation rate, and indel ratio are provided in Supplementary Table S3. To study interaction effects, we evaluate all combinations of weak and strong augmentations drawn from this operator set. This results in a total of 25 weak–strong augmentation pairs evaluated across both datasets and SSL frameworks. Augmentation operator details We provide a concise description of the augmentation operators used in this study. Back-translation BT performs codon-level synonymous substitution by replacing each codon with an alternative encoding the same amino acid. For coding sequences with a known, in-frame reading frame, this transformation preserves the underlying protein sequence while introducing variability at the nucleotide level. Nucleotide masking (NN) Nucleotide masking introduces short contiguous regions of ambiguity by replacing nucleotides with the “N” token. This simulates local uncertainty or missing information and is typically applied at a low masking rate. Mutation Mutation applies random nucleotide substitutions at a controlled rate. Unlike BT, this operator does not preserve the encoded amino acid sequence but introduces biologically plausible point mutations without altering sequence length. Insertion–deletion (indel) Insertion–deletion introduces structural perturbations by randomly deleting and inserting nucleotides in equal proportion, thereby preserving the overall sequence length while modifying local sequence structure. Indel+NN The indel+NN operator combines insertion–deletion with nucleotide masking, increasing augmentation diversity by jointly introducing structural and local perturbations. The biological interpretation of BT and indel depends on the sequence context. BT is most directly applicable to coding sequences such as AMR genes, where synonymous codon substitutions can change the nucleotide sequence while preserving the encoded amino acid. For non-coding regions and short reads with an unknown reading frame, including viral and human background reads in the Oncovirus dataset, BT is better interpreted as a codon-structured nucleotide perturbation rather than a strictly protein-preserving transformation. Indel represents local insertion and deletion variation while preserving the total sequence length. It can model local sequence variation or sequencing noise in non-coding and short-read sequences, although random indels in coding regions may disrupt the reading frame. We therefore use indel as a controlled robustness perturbation rather than an exact model of coding-sequence evolution. Training protocol and evaluation For each experimental setting, models are trained for a fixed number of epochs using identical optimization hyperparameters. All results are reported as mean accuracy and standard deviation over five random seeds to assess robustness to initialization and data order. For each dataset and SSL framework, we compared the 25 augmentation settings using a Friedman test across the five random seeds [22]. When the Friedman test was significant, we applied Nemenyi post-hoc comparisons to examine pairwise differences among augmentation settings. Statistical significance was assessed at α =0.05, and the detailed results are reported in Supplementary Method S6. Unless otherwise specified, evaluation is conducted on held-out test sets using standard classification accuracy as the primary metric. Full training configurations and hyperparameters are provided in Supplementary Material S2. Results Augmentation landscape across weak/strong DNA views Table 1 and Fig. 1 summarize the augmentation design space for semi-supervised genomic learning across four settings (FixMatch/FlexMatch × AMR/Oncovirus). Figure 1 visualizes the accuracy difference of each weak→strong augmentation pair relative to a lower-bound baseline trained only on the initial labeled subset without pseudo-labeling or unlabeled data utilization, where red indicates improvement and blue indicates degradation. Notably, a substantial portion of augmentation combinations result in performance degradation, highlighting that augmentation selection is non-trivial in genomic SSL. Across panels, the highest-performing regions are consistently defined by NN and indel as weak augmentations, particularly when paired with strong sequence-level perturbations. Under FixMatch on AMR, the best-performing combination is NN→mutation (0.8650±0.0044), which exceeds the BT→BT (0.7480±0.0574). Table 1.Performance of evaluated weak–strong augmentation pairs across semi-supervised genomic learning settingsWeak → StrongFixMatch-AMRFixMatch-OncovirusFlexMatch-AMRFlexMatch-Oncovirusbt → bt0.7480 ± 0.05740.5238 ± 0.02740.6417 ± 0.13870.5027 ± 0.0475bt → nn0.7942 ± 0.05480.5346 ± 0.01880.6106 ± 0.12700.4907 ± 0.0444bt → mutation0.8121 ± 0.01570.5288 ± 0.02320.6801 ± 0.17460.4468 ± 0.1011bt → indel0.7831 ± 0.03860.5247 ± 0.02780.7522 ± 0.12780.4720 ± 0.0726bt → indelnn0.7929 ± 0.02330.5274 ± 0.02320.6328 ± 0.08810.4868 ± 0.0657nn → bt0.8285 ± 0.02630.6318 ± 0.01590.8148 ± 0.02560.6524 ± 0.0464nn → nn0.8328 ± 0.01990.6543 ± 0.05740.8246 ± 0.02050.5371 ± 0.0284nn → mutation 0.8650 ± 0.00440.6599 ± 0.04150.8003 ± 0.07820.6183 ± 0.0851nn → indel0.8367 ± 0.0186 0.7080 ± 0.02290.8206 ± 0.01850.6351 ± 0.0570nn → indelnn0.8454 ± 0.01610.7046 ± 0.02270.8288 ± 0.02070.6485 ± 0.0854mutation → mutation0.8026 ± 0.03460.5849 ± 0.03170.6666 ± 0.12910.5146 ± 0.0428indel → nn0.8190 ± 0.01180.6880 ± 0.0283 0.8346 ± 0.0128 0.6610 ± 0.0200Mean test accuracy ( ± std) over random seeds is reported for FixMatch and FlexMatch on the AMR and Oncovirus datasets. Best results per column are shown in bold. Performance of evaluated weak–strong augmentation pairs across semi-supervised genomic learning settings Mean test accuracy ( ± std) over random seeds is reported for FixMatch and FlexMatch on the AMR and Oncovirus datasets. Best results per column are shown in bold. Difference of accuracy of augmentation approach over the baseline (just the labeled data) for semi-supervised genomic learning over all the combinations. Each heatmap reports the mean test-accuracy gain over the supervised lower bound for one setting (FixMatch/FlexMatch × AMR/Oncovirus), with weak augmentations on rows and strong augmentations on columns. Values are averaged over five random seeds. The figure illustrates how different weak–strong augmentation combinations influence semi-supervised performance across datasets and SSL frameworks. BT: back-translation; NN: nucleotide masking; Mutation: random nucleotide substitution; indel: insertion/deletion perturbation; LB: Lower Bound. On Oncovirus, NN →indel achieves the highest FixMatch accuracy (0.7080 ± 0.0229), again outperforming BT-based pairs (e.g., BT →BT at 0.5238 ± 0.0274). For FlexMatch, the same trend persists: indel →NN achieves the strongest results on both AMR (0.8346 ± 0.0128) and Oncovirus (0.6610 ± 0.0200), and NN →indelNN remains competitive (AMR: 0.8288 ± 0.0207; Oncovirus: 0.6485 ± 0.0854). These patterns indicate that NN- and indel-based augmentations consistently yield strong-performing combinations across datasets and SSL algorithms, further validated in Supplementary Material S6. A detailed ablation study on masking ratios, mutation rates, and indel strength is provided in Supplementary Material S3. Detailed ablation results are provided in Supplementary Tables S4–S8. A complementary evaluation using macro-F1 and balanced accuracy across all 25 augmentation pairs generally supports these trends, while indicating that the exact top-performing configuration can vary by dataset, SSL framework, and evaluation metric (Supplementary Table S11). Robustness to random seeds Figure 2 evaluates seed-wise stability by plotting per-seed test accuracy (five seeds) for each augmentation pair. High-performing combinations identified in Fig. 1 also exhibit tight seed distributions, particularly for FixMatch on AMR, where NN/mutation/indel-based strong views show low variance and high median accuracy. On the more challenging Oncovirus setting, dispersion increases, yet NN- and indel-based regimes remain both stronger and more stable than BT-based alternatives. FlexMatch displays greater seed sensitivity overall, but its top-performing pairs (e.g. indel →NN) retain robust performance with consistently high outcomes across seeds. These results suggest that the observed improvements are not driven by favorable random initialization; instead, they reflect reliable benefits of DNA augmentations in semi-supervised genomic learning. Detailed confusion matrix analyses for both tasks are provided in Supplementary Figs S1–S4. Seed-wise robustness of weak to strong augmentation pairs. For each setting (FixMatch/FlexMatch × AMR/Oncovirus), boxplots summarize the distribution of test accuracy across five random seeds for every weak-to-strong augmentation pair, ordered by mean accuracy within each panel. Boxes show the interquartile range with the median indicated by the center line, and points correspond to individual seed results. This visualization highlights how different augmentation pairs vary not only in mean performance but also in seed-wise stability. Further analyses on pseudo-label dynamics and class-wise learning behavior are provided in Supplementary Tables S9 and S10 and Supplementary Figs S5 and S6. As shown in Table 2, FlexMatch was generally faster than FixMatch across the augmentation combinations, with average runtimes of 1.38 and 2.43 min, respectively. FlexMatch was generally faster, potentially because its dynamically adjusted class-wise thresholds enable more adaptive and efficient pseudo-label selection than the fixed global threshold used in FixMatch. Table 2.Runtime(minutes) comparison of AMR classification experiments under different augmentation combinationsStrong augmentationSSL methodWeak augmentationBTNNMutationIndelIndelNNFixMatchBT4.333.203.013.013.25NN3.042.192.032.032.21Mutation2.801.861.721.741.81Indel2.952.121.941.932.12IndelNN3.032.112.001.972.45FlexMatchBT1.330.750.730.740.76NN2.281.321.881.001.75Mutation1.081.560.681.711.69Indel3.111.211.541.801.60IndelNN0.931.480.631.281.75Each entry reports the wall-clock runtime in minutes. All experiments converged. Rows and columns denote the augmentation methods applied to the two augmented views, respectively. Runtime(minutes) comparison of AMR classification experiments under different augmentation combinations Each entry reports the wall-clock runtime in minutes. All experiments converged. Rows and columns denote the augmentation methods applied to the two augmented views, respectively. Conclusion Across FixMatch and FlexMatch, and across AMR and Oncovirus datasets, we find that weak augmentation (through masking, NN) →strong augmentations (indelNN/mutation/indel) consistently outperform weak synonymous mutations (with any strong augmentation method) via back-translation and remain robust under seed variation (Table 1; Figs 1 and 2). Supplementary Material Acknowledgements The authors thank the anonymous reviewers for their valuable suggestions. Author contributions: Hyunwoo Yoo (Conceptualization [equal], Data curation [equal], Formal analysis [equal], Investigation [equal], Methodology [equal], Software [equal], Validation [equal], Visualization [equal], Writing—original draft [equal], Writing—review & editing [equal]) and Gail L. Rosen (Conceptualization [equal], Data curation [equal], Formal analysis [equal], Funding acquisition [lead], Investigation [equal], Methodology [equal], Project administration [lead], Resources [lead], Software [equal], Supervision [lead], Validation [equal], Visualization [equal], Writing—original draft [equal], Writing—review & editing [equal]). Contributor Information Hyunwoo Yoo, Department of Electrical and Computer Engineering, Drexel University, 3141 Chestnut St, Philadelphia, PA 19104, United States. Gail L Rosen, Department of Electrical and Computer Engineering, Drexel University, 3141 Chestnut St, Philadelphia, PA 19104, United States. Supplementary data Supplementary data is available at NAR Genomics & Bioinformatics online. Conflict of interest None declared. Funding This work is supported in part by funds from the National Science Foundation (NSF: # 1936791 and # 2107108).Also, this work was also supported by a minigrant from the National Aeronautics and Space Administration under Grant No. 80NSSC20M0097 issued through the Pennsylvania Space Grant Consortium. Availability and implementation The source code and experimental scripts used in this study are publicly available at https://github.com/EESI/dna-ssl-augmentations and https://doi.org/10.5281/zenodo.23004635. The repository has been organized as a user-oriented software package that enables researchers to apply the proposed augmentation operators and SSL frameworks (FixMatch and FlexMatch) to their own genomic sequence datasets. Users can provide labeled and unlabeled sequence data in a simple CSV format and train models via a unified command-line interface, with configurable augmentation strategies and backbone models. In addition, the package supports inference on new sequences using trained models. Detailed usage instructions and examples are provided in the repository and Software Usage in Supplementary Material. References Associated Data Supplementary Materials