26Genomic enhancers form the central nodes of gene regulatory networks by harbouring combinations of 27 transcription factor binding sites. Deciphering the combinatorial code by which these binding sites are 28 assembled within enhancers is indispensable to understand their regulatory involvement in establishing 29 a cell's phenotype, especially within biological systems with dysregulated gene regulatory networks, 30 such as melanoma. In order to unravel the enhancer logic of the two most common melanoma cell states, 31namely the melanocytic and mesenchymal-like state, we combined comparative epigenomics with 32 machine learning. By profiling chromatin accessibility using ATAC-seq on a cohort of 27 melanoma 33 cell lines across six different species, we demonstrate the conservation of the two main melanoma states 34and their underlying master regulators. To perform an in-depth analysis of the enhancer architecture, 35we trained a deep neural network, called DeepMEL, to classify melanoma enhancers not only in the 36 human genome, but also in other species. DeepMEL revealed the presence, organisation and positional 37 specificity of important transcription factor binding sites. Together, this extensive analysis of the 38 melanoma enhancer code allowed us to propose the concept of a core regulatory complex binding to 39 melanocytic enhancers, consisting of SOX10, TFAP2A, MITF and RUNX, and to disentangle their 40individual roles in regulating enhancer accessibility and activity. 41 63 or co-expressed genes 8,9 ; (2) comparative genomics, by exploiting cross-species data to identify 64 conserved and therefore possible important (parts of) enhancers 10-12 ; (3) genetic screens to measure the 65 effect of mutations on enhancer activity 13,14 ; and (4) machine learning techniques, where mathematical 66 models are trained to recognise specific patterns in enhancers and help to classify them 15 . Particularly 67 the latter has seen a strong boost the last years, with the advent of large training sets derived from 68 genome-wide profiling. Three pivotal methods based on deep learning include DeepBind 16 , DeepSEA 17 69and Basset 18 , the first convolutional neural networks (CNNs) applied to genomics data 19 . Since their 70 emergence in the genomics field, machine learning techniques, and especially CNNs, have been applied 71to model a range of regulatory aspects, including TF binding sites 20 , DNA methylation 21 and 3D 72 chromatin architecture 22 , by exploiting large epigenomics datasets. 73 74Deciphering gene regulation and the underlying enhancer code is not only important during dynamic 75 processes such as development, but also in disease contexts such as cancer, where gene regulatory 76 networks are typically dysregulated due to mutations. Melanoma is a type of skin cancer which mostly 77 develops from a buildup of UV-induced mutations in melanocytes, the pigment-producing cells in the 78 skin 23 . Particularly in this cancer type, gene expression is dysregulated and highly plastic, giving rise to 79 two main me...

Recent genomic and scRNA-seq analyses of melanoma identified common transcriptional states correlating with invasion or drug resistance, but failed to find recurrent drivers of metastasis. To test whether transcriptional adaptation can drive melanoma progression, we made use of a zebrafish mitfa:BRAFV600E;tp53-/- model, in which malignant progression is characterized by minimal genetic evolution. We undertook an overexpression-screen of 80 epigenetic/transcriptional regulators and found neural crest-mesenchyme developmental regulator SATB2 to accelerate aggressive melanoma development. Its overexpression induces invadopodia formation and invasion in zebrafish tumors and human melanoma cell lines. SATB2 binds and activates neural crest-regulators, including pdgfab and snai2. The transcriptional program induced by SATB2 overlaps with known MITFlowAXLhigh and AQP1+NGFR1high drug resistant states and functionally drives enhanced tumor propagation and resistance to Vemurafenib in vivo. Here we show that melanoma transcriptional rewiring by SATB2 to a neural crest mesenchyme-like program can drive invasion and drug resistance in endogenous tumors.