{
  "id": 18093004,
  "title": "Synthetic MPRA perturbation supervision improves variant-effect prediction in sequence-to-function models",
  "url": "https://urgent.news/2026/11/03/synthetic-mpra-perturbation-supervision-improves-variant-effect",
  "topic": "science",
  "section": "Science",
  "published": "2026-11-03T00:00:00.000Z",
  "source": {
    "name": "bioRxiv",
    "slug": "biorxiv",
    "url": "https://www.biorxiv.org/content/10.64898/2026.11.02.763699v1?rss=1"
  },
  "original_language": "en",
  "account": "Deep learning models that translate genetic sequences into functional outcomes, such as gene expression, often learn from data derived from a standard reference genome. Nevertheless, these models encounter challenges when predicting the influence of genetic alterations on gene expression. Massively parallel reporter assays (MPRAs) provide direct insights into the effect of genetic variations on gene expression, yet the sheer volume of data required for fine-tuning sequence-to-function (S2F) models using such assays is constrained. This research explores the potential of fine-tuning Borzoi, a leading S2F model, using synthetic data generated through massively parallel reporter assays (MPRAs) as a feasible substitute for restricted training data.\n\nThe team initially trained a collection of 16 models on 110,359 K562 MPRA measurements. Subsequently, the predictions generated by this ensemble of models were amalgamated into a single model, known as the student model. This student model was trained using synthetic variant-effect labels derived from MPRA elements, focusing on 1,000 genes with 600 variants per gene. The objective was to fine-tune Borzoi by progressively incorporating more genes into the training process, ranging from 50 to 1,000 genes, each with 600 variants.\n\nAs more genes were included in the training phase, the model's performance progressively improved. In scenarios where data was scarce, training the model on a larger number of genes, each with fewer variants, proved more effective than training on fewer genes with a higher number of variants, assuming the same total number of variants. When compared to the pre-trained Borzoi, the model fine-tuned on 600,000 synthetic samples across 1,000 genes demonstrated a significant increase in the Pearson correlation coefficient. Specifically, the correlation improved from 0.28 to 0.54 for CAGI5 promoter variants and from 0.39 to 0.46 for 903 GTEx fine-mapped whole-blood expression quantitative trait loci (eQTLs).\n\nA crucial aspect of this study is that these enhancements occurred without compromising the pre-trained performance across all 7,611 output tracks of the Borzoi model. Additionally, the fine-tuned Borzoi was able to rediscover regulatory motifs identified by the MPRA model in genes that were not part of the initial training data. This suggests that the model has successfully learned features related to regulatory mechanisms, which can be widely applicable across various genetic loci. The findings of this study demonstrate that synthetic supervision derived from MPRAs can effectively enhance variant effect prediction within S2F models, while retaining the original capabilities of the pre-trained model. This approach presents a scalable solution for better understanding the impact of genetic variations on gene regulation.",
  "summary": "Deep learning sequence-to-function (S2F) models learn to predict functional readouts, such as gene expression, from genetic sequence. However, these models are trained predominantly on the reference genome and struggle to accurately predict the effect of genetic variation on gene expression. Massively parallel reporter assays (MPRAs) explore the direct impact of genetic variation on gene…",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 1,
    "also_reported_by": []
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}