Loading
Life sciences · Preprint
arXiv · September 4, 2026
Early or partial results. Treat as a signal, not a conclusion.
GenDA, a 202M-parameter bidirectional diffusion model, achieves modest improvement in supervised ClinVar variant-effect prediction (AUROC 0.774, +0.103 vs. baseline autoregressive model), but this improvement does not result from the hypothesized entropy-guided reconstruction mechanism. More critically, the model fails to outperform a simple compositional-control shuffle in zero-shot functional sequence generation across promoters, enhancers, and splice boundaries, suggesting that strong fine-tuned performance on variant calls does not translate to functional capability.
Machine learning model development and benchmarking study with supervised fine-tuning and zero-shot functional validation. Genomic sequences; ClinVar SNVs for fine-tuning and evaluation; promoter, enhancer, and splice-boundary sequences for functional validation. No clinical population or human subjects.. Intervention: GenDA bidirectional diffusion model with entropy-guided span placement for genomic sequence reconstruction and variant-effect prediction. Compared with: Autoregressive baseline model (similarly scaled, 202M parameters) for variant prediction; random-span variant (entropy guidance removed); control shuffle preserving 3-mer composition for functional inpainting.
GenDA 202M-parameter model reaches pooled ClinVar SNV AUROC of 0.774, exceeding autoregressive baseline by 0.103 Matched random-span variant reaches AUROC of 0.777, providing no evidence entropy guidance caused ClinVar improvement GenDA fails zero-shot functional inpainting: does not consistently outperform control shuffling native gap while preserving 3-mer composition across promoters, enhancers, exon boundaries, and intron boundaries
Safety was not reported in the material analysed. Check the source before drawing any conclusion about harm.
This work does not provide sufficient evidence to support clinical or functional deployment. The discordance between variant-prediction performance and functional generation capacity highlights that in silico genomic models may learn spurious patterns that improve on specific benchmarks without capturing true biological function, necessitating orthogonal validation before use in variant interpretation or synthetic sequence design.
A single-center machine learning model development study with mixed results on surrogate endpoints (variant prediction AUROC, functional inpainting), no clinical validation, and failure of the core hypothesis that entropy guidance improves function, warranting independent replication before clinical application.
As stated by the source record.
Quoted from the source exactly as published.
This work does not provide sufficient evidence to support clinical or functional deployment. The discordance between variant-prediction performance and functional generation capacity highlights that in silico genomic models may learn spurious patterns that improve on specific benchmarks without capturing true biological function, necessitating orthogonal validation before use in variant interpretation or synthetic sequence design.
Graded across the dimensions that decide whether you should act, each from what the source actually supports. There is no single score, and where a dimension was not assessed it says so.
Bidirectional discrete diffusion model appears naturally suited to genomic modeling because it can reconstruct missing sequence from both flanks. We developed GenDA (Genomic Density-optimized Absorbing Diffusion) under the additional hypothesis that entropy-guided span placement would concentrate reconstruction pressure on compositionally complex regions, improving both downstream variant-effect prediction and functional sequence generation. Our results only partially support this premise. After supervised fine-tuning, the 202M-parameter GenDA model reaches a pooled ClinVar SNV AUROC of 0.774, exceeding a similarly scaled autoregressive model by 0.103. However, a matched random-span variant reaches 0.777, providing no evidence that entropy guidance causes the ClinVar improvement. More unexpectedly, GenDA fails a zero-shot functional inpainting stress test: across promoters, enhancers, exon boundaries, and intron boundaries, it does not consistently outperform a control that shuffles the native gap while exactly preserving 3-mer composition. Failure is already present for 50--500-bp gaps, although enhancer degradation worsens at longer gaps. Diagnostics identify several boundary conditions: entropy measures local sequence complexity rather than functional importance; 1-mer tokenization limits physical context; training spans are capped at 300 bp; and high absolute AlphaGenome fidelity can coexist with negative control-normalized restoration. These results show that strong fine-tuned variant prediction, a plausible corruption prior, and functional generation are distinct claims that require separate validation.
Taken from the source record, never inferred. Follow any of these and new work involving them reaches your briefing.