Reaping the Fruits of LLM Pruning: Towards Small Language Models for Efficient Non-Coding Variant Effect Prediction

What this source contributed

Records in this release that cite this source. A record can cite several sources.

• 3 configurations
• 2 protocols
• 1 dataset
• 3 evaluations
• 8 results
• 9 evidence claims
• 1 use case

Results

Values as the source printed them. Compare results only under compatible protocols, datasets and metrics.

Claims

Statements this source supports, each with where in the source it was checked.

• Evidence summary: diagnostic genomics model execution: Summary Eleven published execution benchmarks are linked, alongside the DRAGEN 4.2 developer total runtime for HG002 (1,838.5 seconds); four are withheld pending source concerns (see the end of this summary), and pretraining compute for NTv3 and HyenaDNA is linked only as outside scope. Of the seven active benchmarks, five measure conventional variant-calling pipelines and two measure genomic foundation models outside any diagnostic workflow: inference on microbial genomes, and fine-tuning and test-set scoring on human eQTL variants. Only the foundation-model benchmarks print repeated runs (five trials for GenomeOcean, three seeds for the Nucleotide Transformer); none reruns a variant-calling pipeline to measure concordance of the calls, none prints variant-calling accuracy in the same table as runtime, and hardware differs within most comparisons, so differences mix software and hardware. On ten WGS samples run by a University of Oslo and Oslo University Hospital group with no declared competing interests (Samarakoon et al. 2025), read mapping plus HaplotypeCaller for NA12878 (47.55x) took 2,317.7 minutes on a CPU-only GATK pipeline, 123.45 minutes with Parabricks on four L4 GPUs, 78.12 on four A100 GPUs, 54.22 on eight H100 GPUs and 60.29 with DRAGEN v4.2. On the same H100 machine, Parabricks DeepVariant took 14.58 minutes for NA12878 with the version the pipeline used (printed 4.1.0-1; Methods say 4.3.0-1) and 16.05 minutes with the updated version (printed 4.3.0-1; Discussion says 4.3.1-1). On one HG002 30x genome on Google Cloud (O'Connell et al. 2023; the authors' employer funded the study and is an NVIDIA, AWS and Google alliance partner), HaplotypeCaller took 38.8 hours and cost 67.9 USD on a 32-vCPU machine and 0.59 hours and 17.5 USD on eight A100 GPUs, while the somatic caller LoFreq cost 8.1 USD on the CPU machine and 19 to 30.1 USD on 2 to 8 A100 GPUs. On Google Cloud machines of similar hourly cost (Franzoso et al. 2025, no declared conflicts), Sentieon DNASeq took 0:14:00 to 0:16:00 per exome at 0.82 to 1.03 USD and 3:00:00 to 3:50:00 per genome at 8.02 to 10.7 USD, and Parabricks Germline on one T4 GPU took 0:10:00 to 0:14:00 per exome at 0.71 to 0.93 USD and 4:06:00 to 4:40:00 per genome at 8.13 to 10.6 USD; costs exclude the Sentieon licence, and two genome runtime cells that cannot be genome runtimes are disputed and left out of these ranges. On human WGS at five read depths on the HPC nodes of a regional genome project (Marangoni et al. 2026; the GeNePi developers, no declared competing interests; one unnamed sample per depth), a 30x genome took 17h with the CPU GATK equivalent and 1h 36m, 42 m and 30m with Parabricks v4.3.0 on two V100, two A100 and four A100 GPUs, and a 100x genome 60h, 6h 30m, 2h 30m and 2h 06m. On one 120-vCPU Azure instance (Hu et al. 2026; all authors employed by Sentieon), DNAscope took 18.9 minutes for 30x Illumina, DNAscope LongRead 41.2 minutes for 30x PacBio HiFi and DNAscope Hybrid 85.8 minutes for 10x PacBio HiFi plus 35x Illumina, at 1.1, 2.5 and 5.2 USD on demand excluding the licence; each pipeline had different input data, so the three are not a ranking. For GenomeOcean on one H100 (Yu et al. 2026; two authors are at the DOE Joint Genome Institute, which develops GenomeOcean; GTDB bacterial genomes; mean and standard deviation over five trials), embedding extraction with the 4B model ran at 21,386 ± 126 input tokens per second with 72.14 GB peak GPU memory in BF16 and 22,401 ± 204 with 67.91 GB with FP8 weights, and generation ran at 1818.79 ± 7.95 generated tokens per second in BF16 and 2169.04 ± 24.30 with FP8 weights, activations and KV cache. On a human eQTL variant classification task (Hegde et al. 2025; Kingston University authors, independent of the model developers; 8,862 test variants), LoRA fine-tuning of Nucleotide Transformer 500M for 5 epochs took 7.53, 7.50 and 7.48 hours over three seeds and scoring the test split 8.49, 8.51 and 8.49 minutes, on GPUs the Methods give as two TITAN Xp; fine-tuning to convergence took approximately 19 hours for the Nucleotide Transformer and around 1 hour per pruned DNABERT-2 model on two Quadro RTX 8000. No active linked result measures variant-calling accuracy against a truth set; the GenomeOcean ARI values score metagenomic binning against GTDB species labels. The four withheld studies are Zheng et al. 2026 (Clair3 developers; Clair3 and DeepVariant on HG003 long reads), whose supplementary runtimes, DeepVariant version and DeepVariant accuracy differ from its Results text and whose accuracy uses one GIAB sample's v4.2.1 confident regions, which leave out the hardest regions; Aspidova et al. 2026 (ModernGENA and GENA-LM developers; four DNA language model families on one A100), whose printed tokens per second are far too low for its batch sizes; Pearce et al. 2026 (EVEE developers, three of whom co-authored Evo 2; Evo 2 embedding extraction for all ClinVar variants), whose source carries a concern about another figure; and Dalla-Torre et al. 2025 (Nucleotide Transformer developers; IA3 fine-tuning time), whose Methods fine-tuning times conflict with its Discussion. Apart from the five GenomeOcean trials and the three Nucleotide Transformer seeds, each comparison prints one value per sample and set-up on at most ten samples, and costs are list prices at the time of each study. (amp-source-dragen-supplementary-tables: result amp-result-dragen-hg002-phase4-runtime; model-execution-20261009-source-samarakoon2025-supplement: Supplementary Tables S1 and S3; model-execution-20261009-source-oconnell2023: Table 1; model-execution-20261009-source-franzoso2025-table-s2: Table S2; articles: Methods, Discussion, Funding and Conflicts of interest; model-execution-20261011-source-marangoni2026-genepi-supplement: Table S1; model-execution-20261011-source-marangoni2026-genepi: Competing interest; model-execution-20261011-source-hu2026-dnascope-hybrid: Table 2, Conflict of interest; model-execution-20261011-source-yu2026-fp8-genomeocean: Tables 3 and 4, author affiliations; gfm-guide-20261014-source-hegde2025: Table A1, Methods paragraph 3, Appendices A.2 and A.3, author affiliations)
• gfm-guide-20261014-eval-hegde-dnabert-2-pruned-convergence: attributes.comparison.adaptation: Comparison: adaptation LoRA parameter-efficient fine-tuning (peft; r=1, lora_alpha=32, lora_dropout=0.1) (gfm-guide-20261014-source-hegde2025: Appendix A.2 ('Parameter-efficient fine-tuning using the peft package: r = 1; lora_alpha = 32; lora_dropout = 0.1'))
• gfm-guide-20261014-eval-hegde-nt-500m-human-ref-5-epochs: attributes.comparison.adaptation: Comparison: adaptation LoRA parameter-efficient fine-tuning (peft; r=1, lora_alpha=32, lora_dropout=0.1) (gfm-guide-20261014-source-hegde2025: Appendix A.3 ('Parameter-efficient fine-tuning using the peft package: r = 1; lora_alpha = 32; lora_dropout = 0.1'))
• gfm-guide-20261014-eval-hegde-nt-500m-human-ref-convergence: attributes.comparison.adaptation: Comparison: adaptation LoRA parameter-efficient fine-tuning (peft; r=1, lora_alpha=32, lora_dropout=0.1) (gfm-guide-20261014-source-hegde2025: Appendix A.3 ('Parameter-efficient fine-tuning using the peft package: r = 1; lora_alpha = 32; lora_dropout = 0.1'))
• Relevance of Fine-tuning and evaluation wall time, 5 epochs on the Enformer eQTL dataset (Hegde et al. 2025 Table A1) to "Select an execution workflow for large-scale diagnostic genomics": Assessed by Fine-tuning and evaluation wall time, 5 epochs on the Enformer eQTL dataset (Hegde et al. 2025 Table A1) (gfm-guide-20261014-source-hegde2025: Table A1; Methods paragraph 3; Table 1 (train and test split sizes); Appendix A.3)
• Relevance of Fine-tuning wall time to convergence on the Enformer eQTL dataset (Hegde et al. 2025 Methods) to "Select an execution workflow for large-scale diagnostic genomics": Assessed by Fine-tuning wall time to convergence on the Enformer eQTL dataset (Hegde et al. 2025 Methods) (gfm-guide-20261014-source-hegde2025: Methods paragraph 3; Table 2; Appendices A.2 and A.3)
• use-case-diagnostic-genomics-model-execution: attributes.clinical_scope: Clinical scope Clinical applicability is not established: the linked evidence is published runtime, cost and resource use of conventional variant-calling pipelines on a few samples and two foundation-model benchmarks outside any diagnostic workflow (inference on microbial genomes; fine-tuning and test-set scoring time on human eQTL variants), not a foundation-model execution benchmark inside a diagnostic workflow, a reproducibility study or a clinical validation. (Store: reviewed judgements use-case-mapping-gfm-guide-20261014-hegde2025-eqtl-5-epochs and -eqtl-convergence become active; the EVEE and Nucleotide Transformer IA3 judgements are reviewed and withheld; the NTv3 and HyenaDNA judgements are outside scope)
• use-case-diagnostic-genomics-model-execution: attributes.decision: Decision Compare published per-sample runtime and cost of CPU, GPU and DRAGEN server execution set-ups that resemble the team's hardware and sequencing type when budgeting a workflow; most are conventional variant-calling pipelines, the two active foundation-model benchmarks (inference on microbial genomes; fine-tuning and test-set scoring on human eQTL variants) are outside any diagnostic workflow, and none measures run-to-run concordance of variant calls. (Store: reviewed judgements use-case-mapping-gfm-guide-20261014-hegde2025-eqtl-5-epochs and -eqtl-convergence become active; the EVEE and Nucleotide Transformer IA3 judgements are reviewed and withheld; the NTv3 and HyenaDNA judgements are outside scope)
• use-case-diagnostic-genomics-model-execution: attributes.setting: Setting Published execution benchmarks, each on stated inputs: CPU-only GATK, Parabricks on L4, A100 and H100 GPUs and DRAGEN v4.2 on ten germline WGS samples (Samarakoon et al. 2025); two germline and four somatic callers on CPU and 2, 4 and 8 GPU cloud and DGX machines on one HG002 30x genome (O'Connell et al. 2023); Sentieon DNASeq and Parabricks Germline on Google Cloud machines of similar hourly cost, five exomes and five genomes (Franzoso et al. 2025); the DRAGEN 4.2 developer total runtime for HG002, 1,838.5 seconds on one hardware configuration and 5,521.21 seconds on an AWS f1.4xlarge configuration, whose stage times cannot be summed because of concurrency; Parabricks v4.3.0 on V100 and A100 nodes against the CPU GATK equivalent at five read depths of human WGS (Marangoni et al. 2026); three Sentieon DNAscope pipelines on one Azure instance (Hu et al. 2026); GenomeOcean inference in BF16 and FP8 on one H100 with microbial genomes (Yu et al. 2026); and LoRA fine-tuning and test-set scoring time of Nucleotide Transformer 500M on 2 × TITAN Xp, and fine-tuning time of pruned DNABERT-2 on 2 × Quadro RTX 8000, on a human eQTL variant classification task (Hegde et al. 2025). Four further linked studies are withheld pending source concerns: Clair3 and DeepVariant on HG003 long reads (Zheng et al. 2026), DNA language model inference on one A100 (Aspidova et al. 2026), Evo 2 embedding extraction for all ClinVar variants (Pearce et al. 2026) and Nucleotide Transformer IA3 fine-tuning time (Dalla-Torre et al. 2025). Pretraining compute for NTv3 and HyenaDNA is linked as outside scope. Hardware differs within most comparisons, so differences mix software and hardware. (Store: reviewed judgements use-case-mapping-gfm-guide-20261014-hegde2025-eqtl-5-epochs and -eqtl-convergence become active; the EVEE and Nucleotide Transformer IA3 judgements are reviewed and withheld; the NTv3 and HyenaDNA judgements are outside scope)

Models, benchmarks and other records

Evidence

Source checking verifies the cited claim or transcription. It does not establish independent reproduction.