guanine
gpra-c¶
tags: gene expression, promoter sequences, GPRA, dual-reporter assay, yeast, synthetic biology
tl;dr¶
gpra-complex (aka gpra-c) is a synthetic biology task measuring gene expression in yeast. gpra-c solves the issue of finite natural gene-promoter diversity by relying on random oligonucleotides.
overview¶
The central dogma suggests that genes exist to produce RNA, which is then used to manufacture proteins. Upstream promoter elements heavily regulate the amount of RNA a gene can produce, known as its gene expression. Because genes (and their adjacent promoters) tend to be constrained evolutionarily, there is limited sequence diversity in and around gene regions – including promoters. This means that only a small, finite set of promoter sequences (~ 20k in hg38) are ‘trainable’ for any given model. To circumvent this bottleneck, the de Boer lab has developed ways of injecting randomized promoter sequences (complete with their own control) into yeast, known as Gigantic Parallel Reporter Assays.
Class labels are ordinal and range from zero (0, minimal expression) to seventeen (17, maximal expression). Intermediate scores of one to sixteen (1-16) represent increasing levels of gene expression.
See also
The sibling task gpra-d is intended to allow transfer learning research – but both are sizeable stand-alone tasks.
example models¶
model |
\(\rho\) |
|---|---|
T5 (baseline) |
84.6738 |
nt-v2-500m |
72.6726 |
Evo2_1B_base |
72.6487 |
nt-v2-250m |
72.436 |
Caduceus-PS |
72.4355 |
5-mer LinSVR (baseline) |
36.3022 |
GC-content (baseline) |
20.5533 |
interpretation¶
gpra-c is a difficult, but insightful task. While its dynamic range (the typical upper and lower bound of scores) is slightly constrained – it nonetheless produces rankings that correlate to model quality and other tasks (i.e. the newest, fanciest models increasingly do well).
example usage¶
first, clone the dataset from huggingface (make sure you have Git LFS installed):
git clone https://huggingface.co/datasets/guanine/gpra_c
then, read the file into main memory with your favorite file parser
import pandas as pd
# 1per is the recommended few-shot training split
# there are no bed files for GPRA, as it is not in a reference genome
train_dat = pd.read_csv('gpra_c/1per/1per.csv', sep=',') # csv
train_dat.head()
finally, splice the sequence out with your preferred genome reader, e.g. twobitreader
CONTEXT_SIZE = 8192 # change this for your model
row = train_dat.iloc[0]
seq = row['seq']
# we recommend pre/appending a yeast scaffold for large context models, e.g.
seq = scaf_a + seq + scaf_b
assert len(seq)==CONTEXT_SIZE # we recommend checking for truncation
build details¶
Compared to the source dataset, gpra-c has undergone slight refinement. Specifically, non-canonical length datapoints (i.e. those differening from the standard 80 bp of randomized sequence) have been pruned. While the variability in length likely represents some biological signal, it trivialized a significant portion of the final scoring (as non-80 lengths clustered heavily across the class labels).
controlled factors¶
sequence length
appears in¶
original citation¶
Eeshit Dhaval Vaishnav, Carl de Boer, & Aviv Regev. (2022). The evolution, evolvability and engineering of gene regulatory DNA. https://doi.org/10.1038/s41586-022-04506-6