← Blog

From region to expression in one run: the annotation-to-expression workflow

The composite annotation-to-expression workflow finds genes in a raw region and predicts expression for each one — walk through a worked example over the HBB locus.

Genomic Intelligence
  • release
  • annotation
  • expression
  • workflow

The expression model on the DNA analysis platform reads a DNA sequence centered on a gene’s transcription start site (TSS) and returns a predicted expression level. That works when you already know where the TSS is. Often you start from a raw stretch of genome and don’t. The annotation-to-expression workflow covers that case: it runs gene finding to locate the TSSs in a region first, then predicts expression for each one, in a single run.

The workflow pairs two tasks, annotation (gene finding) and expression, so you can go from an arbitrary genomic window to per-gene expression without hand-centering anything.

What the workflow takes as input

The workflow reads the same two inputs as the expression task, but relaxes the requirement on the sequence:

  1. A DNA sequence — this no longer has to be TSS-centered. You can hand it a raw region (a chr:start-end range pulled from Ensembl, pasted nucleotides, or an uploaded FASTA). On the platform, enable the Find genes first checkbox in the Expression task: it runs DNA annotation to locate each gene’s TSS in your sequence, centers a 9,198 bp window on it, and then predicts expression for that window.
  2. An experimental context description — the same natural-language description of cell type, assay, and related metadata that conditions the expression estimate. Every gene found in the region is scored under this one context.

The result is one expression prediction per transcript the annotation step locates, all reported as log(TPM+1).

A concrete example: a 16 kb window over HBB

To see the workflow end to end, I loaded a raw 16,000 bp region on chromosome 11 spanning the HBB (hemoglobin β) locus (chr11:5,219,000–5,235,000, GRCh38) via the coordinate search, deliberately a plain region rather than a centered gene window. I kept the platform’s built-in K562 experimental context (an ENCODE-style polyA RNA-seq description of the K562 chronic myelogenous leukemia line), enabled Find genes first, and clicked Run Analysis.

Gene Expression Prediction panel for a 16,001 bp region over the HBB locus run through the annotation-to-expression workflow: a '3 genes found' badge; summary cards reading Genes Found 3 (by annotation), Predicted 3 expression values, Max Expression 0.75 log(TPM+1), and Sequence 16,001 base pairs; an Expression per Gene table listing transcript_1, transcript_2, and transcript_3, all on the minus strand, with expression levels 0.75, 0.70, and 0.70 log(TPM+1), each labelled Not Expressed; the K562 experimental-context description; and a footer noting the annotation g0-annotation and expression g0-expression models with a total runtime of 8,458 ms (annotation 7,641 ms + expression 817 ms).
The annotation-to-expression workflow over the 16 kb HBB region: annotation found three transcripts, and the expression model scored each one under the K562 context — summarized in the headline cards and the per-gene table.

The Gene Expression Prediction panel reports the two steps together. The headline cards read Genes Found 3 (by annotation), Predicted 3 expression values, Max Expression 0.75 log(TPM+1), and Sequence 16,001 base pairs. The annotation step located three transcripts in the region, and the expression step returned a value for each. The Expression per Gene table lists them: transcript_1, transcript_2, and transcript_3, all on the minus strand, with predicted expression of 0.75, 0.70, and 0.70 log(TPM+1) respectively, each labelled Not Expressed under the K562 context. The panel echoes the K562 experimental-context description used for scoring, and the footer records the two models, g0-annotation and g0-expression, plus the runtime: 8,458 ms total (annotation 7,641 ms + expression 817 ms).

Everything here started from a plain 16 kb region: the workflow found the genes, centered a window on each TSS, and scored them, with no manual step in between.

These are computational predictions intended for research and development, not clinical or diagnostic decisions. Try the annotation-to-expression workflow — on your own regions — at app.genomicintelligence.ai.