Research map / Architectures and efficiency / Pruning and efficient tuning
Genomics Data Lossless Compression with (S, K)-Mer Encoding and Deep Neural Networks
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DeepGeCo, a lossless genomics compressor with three modes (MINI static, PLUS adaptive, ULTRA semi-adaptive) that combines (s,k)-mer encoding, a frozen BiGRU Warm-Start model, and a Transformer Supporter model to address cold-start and throughput issues.
Problem
Learning-based genomics compressors either give inferior compression ratios or low throughput, and adaptive compressors suffer from model cold-start in early batches.
Approach
Combine a frozen BiGRU Warm-Start model pretrained on multi-species genomes with an adaptive Transformer Supporter model, mixed via a time-decaying probability mixer gated by a threshold controller, and pre-encode the sequence with a (s,k)-mer encoder to raise throughput; select (s,k) with a ranking metric RTCR.
Key results
- DeepGeCo-PLUS achieves average 1.677 bits/base vs 1.731 for TRACE and 2.429 for DZIP, a 3.103% and 30.955% improvement respectively.
- DeepGeCo MINI reaches average throughput of about 96.568 KB/s vs 4.208 KB/s for DNA-BiLSTM, with throughput gains up to 22.949x over baselines.
- The (3,3)-mer encoder gives the lowest RTCR and is used as default; ablation shows removing modules degrades compression ratio or throughput.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ULTRA instead of TRACE | DeepGeCo · PlFa | compression ratio (bits/base): 1.834 -> 1.816 | Helped |
| Use ULTRA instead of DNA-BiLSTM | DeepGeCo · WaMe | compression ratio (bits/base): 1.945 -> 1.948 | No clear effect |
| Use ULTRA instead of TRACE | DeepGeCo · DrMe | compression ratio (bits/base): 1.912 -> 1.906 | Helped |
| Use ULTRA instead of DZIP | DeepGeCo · OrSa | compression ratio (bits/base): 1.881 -> 1.899 | Hurt |
| Use ULTRA instead of DZIP | DeepGeCo · GaGa | compression ratio (bits/base): 1.840 -> 1.858 | Hurt |
| Use ULTRA instead of TRACE | DeepGeCo · SnSt | compression ratio (bits/base): 1.873 -> 1.866 | Helped |
| Use ULTRA instead of DZIP | DeepGeCo · MoGu | compression ratio (bits/base): 1.651 -> 1.650 | No clear effect |
| Use ULTRA instead of DNA-BiLSTM | DeepGeCo · AtAl | compression ratio (bits/base): 1.845 -> 1.850 | Hurt |
| Use ULTRA instead of TRACE | DeepGeCo · ArTh | compression ratio (bits/base): 1.905 -> 1.898 | Helped |
| Use ULTRA instead of TRACE | DeepGeCo · HuMa | compression ratio (bits/base): 0.070 -> 0.044 | Helped |
| Remove the (s, k)-mer encoder (ablation) | DeepGeCo PLUS · SnSt | compression ratio (bits/base): 1.860 -> 1.879 | Hurt |
| Remove the (s, k)-mer encoder (ablation) | DeepGeCo PLUS · SnSt | throughput (KB/s): 23.566 -> 9.064 | Hurt |
Surprising findings
- TRACE achieves lower average CUDA and total memory than DeepGeCo modes despite lower compression ratio performance, and DeepGeCo uses slightly more memory than TRACE.
- MINI with larger batch size shows increased compression ratio and GPU memory but reduced throughput.
Limitations
- MINI mode has a low compression ratio
- Warm-Start model depends on pretraining on similar multi-source genomics data
- Model structure hyperparameters were not optimized in detail
Experimental setup
Datasets: 10 real-world genomics datasets (DNA corpus, Pratas and Pinho 2019) multi-species genomics data for Warm-Start pretraining
Benchmarks: 10 real-world genomics datasets (PlFa, WaMe, DrMe, OrSa, GaGa, SnSt, MoGu, AtAl, ArTh, HuMa)
Metrics: compression ratio (bits/base) throughput (KB/s) CPU peak memory (GB) GPU memory (GB) RTCR
Baselines: DNA-BiLSTM DZIP TRACE DeepDNA GeneFormer MFCompress GenoZip
Builds on: TRACE DZIP DNA-BiLSTM k-mer encoding
Closest papers by meaning
- GeneZip: Region-Aware Compression for Long Context DNA Modeling (NeurIPS 2026 · similarity 0.84)
- EvoPress: Accurate Dynamic Model Compression via Evolutionary Search (ICML 2025 · similarity 0.84)
- MeLLoC: Lossless Compression with High-order Mechanism Learning (NeurIPS 2024 · similarity 0.84)
- The KoLMogorov Test: Compression by Code Generation (ICLR 2025 · similarity 0.83)
- Forget the Data and Fine-Tuning! Just Fold the Network to Compress (ICLR 2025 · similarity 0.83)
- Fast and Low-Cost Genomic Foundation Models via Outlier Removal (ICML 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.