A generalizable deep learning framework for inferring fine-scale germline mutation rate maps

Key results

1. 01

   ShownAcross the human autosomes, MuRaL gave the highest genome-wide regional (binned) mutation rate correlations of all compared models, and its advantage grew as bins got smaller. Its k-mer correlations matched the best models even though it used less training data and no functional genomic inputs.

   Versus Aggarwala 7-mer, Carlson 7-mer, Carlson 7-mer+features and Karczewski 3-mer on 2,390,435,721 shared autosomal bases (1Mb, 100Kb, 10Kb bins); MuRaL used 1.5 millions training mutations and 260,712 parameters vs 392,128 for Carlson 7-mer+features; Carlson models slightly better for A>C and CpG>GpG 5-/7-mers, MuRaL better in almost all 9-mer correlations. · MuRaL outperforms existing models; Comparison of MuRaL models with previously…

2. 02

   ShownEffective models can be trained on singleton rare variants from just 100 diploid genomes. Scored against denser rare-variant sets, their predictions correlated well and were comparable to models trained on singletons from a larger downsampled population.

   '1in200' AT and non-CpG models performed like '1in2000' models trained on equal singleton numbers for Chr20 k-mer (3- to 9-mer) and regional (1Mb, 100Kb, 10Kb) correlations scored on '10in20000' variants; the '1in200' CpG model was close to the '5in1000' model. Default: 500,000 mutated plus 10,000,000 non-mutated training sites. · MuRaL can build effective models with relatively…; Rare variants from gnomAD; Figure legends

3. 03

   ShownTransfer learning from rare-variant models improves de novo mutation (DNM) models trained on limited data. Rare-variant models alone approximate DNM rates about as well as transfer-learned models, so they can stand in when no DNMs are available.

   With 150,000 DNMs (AT, non-CpG) or 50,000 (CpG) from gene4denovo, retraining all pre-trained weights gave significantly lower validation losses than ab initio DNM models over ten trials; rare-variant models had lower or similar losses than ab initio and similar Chr1 k-mer and 1Mb regional correlations to transfer models; retraining only last FC layers performed poorly. · Training with DNMs and transfer learning; De novo mutations; Transfer learning; Figure legends

4. 04

   ShownThe framework carries over to other species with modest data. Transfer-learned macaque models match ab initio ones while using less data, and ab initio models for Arabidopsis and Drosophila reach good k-mer and regional correlations from small training sets.

   M. mulatta (853 genomes): transfer models trained on 30% of the ab initio data performed similarly on rheMac10 Chr20 with less computation time; A. thaliana (1135 inbred lines; 100,000 training variants per model, 50,000 for CpG): 10Kb regional correlations >0.5 for all mutation types; D. melanogaster: singletons from 205 DGRP inbred lines gave promising correlations. · Generating mutation rate maps for other species; Applying MuRaL to other species

5. 05

   ShownMuRaL rate maps support downstream analyses. Clustering human coding genes by fine-scale rate profiles gives three groups with distinct GO enrichments, including a high-rate developmental cluster, and depletion rank (DR) scores built on MuRaL rates rank pathogenic variants better.

   K-means (K=3) on GENCODE v37 genes matched '10in20000' observed rare-variant profiles, unlike clusters from Carlson 7-mer+features or Karczewski 3-mer rates; lowest 10% DR windows (500bp, 50-bp step) had higher ClinVar pathogenic enrichment with MuRaL than Carlson rates (Fisher's exact test, p < 2.2e-16). · Using MuRaL-predicted mutation rates for other…; Mutation rates around human genes; Depletion rank (DR); Figure legends