Building isoform coexpression networks

We created a coexpression network at the isoform level through the following steps introduced in our previous publications^24,36. First, expression data for isoforms from overlapping cell lines of the same cancer type in the CCLE and gCSI datasets were downloaded from the PharmacoGx platform (version 1.12.0)³⁷, which comprises pharmacological profiles for several hundred cell lines. The updated CCLE and gCSI PharmacoSets contain isoform-level expression data processed from raw RNA-seq profiles extracted from CGHub³⁸ and NCBI GEO³⁹. Zhaleh et al.⁴⁰ aligned the RNA-seq reads to the Ensembl Genome Reference Consortium release GRCh38⁴¹ using HISAT2⁴², annotated the isoforms and calculated their expression with StringTie⁴³. A total of 58,037 genes, including 19,950 protein-coding genes, 15,767 long noncoding RNAs (lncRNAs) and 14,650 pseudogenes, was annotated by Gencode (version 25)⁴⁴. Then, the FPKM values (the number of fragments per kilobase per million mapped reads units) were converted to log2 (FPKM + 1) to obtain the expression values of the isoforms. Noncoding isoforms were removed based on Ensembl identifiers using the R package BiomaRt (version 2.34.3)⁴⁵. We calculated the Pearson correlation coefficients of two isoform expression values for each dataset as follows:

where E is the expression value of protein isoforms i and j. The value log2 (FPKM + 1) was ≥1 in at least 30 cancer cell line types in each dataset. Protein isoforms i and j are also common isoforms in both the gCSI and CCLE datasets.

Interactions between isoforms in the same genes were removed. To find a balance between removing weak interactions and keeping more isoforms in the network, the isoform network was filtered by the threshold s = 0.5, which was calculated as follows: