See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/318402958

Classification and function of small open reading frames

Article  in  Nature Reviews Molecular Cell Biology · July 2017

DOI: 10.1038/nrm.2017.58

CITATIONS READS

106 915

2 authors, including:

Pedro Patraquim
Brighton and Sussex Medical School
15 PUBLICATIONS   452 CITATIONS   

SEE PROFILE

All content following this page was uploaded by Pedro Patraquim on 12 January 2018.

The user has requested enhancement of the downloaded file.

 ANALYSIS
Classification and function of small
open reading frames
Juan-Pablo Couso1,2 and Pedro Patraquim2
Abstract | Small open reading frames (smORFs) of 100 codons or fewer are usually — if arbitrarily
— excluded from proteome annotations. Despite this, the genomes of many metazoans,
including humans, contain millions of smORFs, some of which fulfil key physiological functions.
Recently, the transcriptome of Drosophila melanogaster was shown to contain thousands of
smORFs of different classes that actively undergo translation, which produces peptides of mostly
unknown function. Here, we present a comprehensive analysis of smORFs in flies, mice and
humans. We propose the existence of several functional classes of smORFs, ranging from inert
DNA sequences to transcribed and translated cis-regulators of translation and peptides with a
propensity to function as regulators of membrane-associated proteins, or as components of
ancient protein complexes in the cytoplasm. We suggest that the different smORF classes could
represent steps in gene, peptide and protein evolution. Our analysis introduces a distinction
between different peptide-coding classes of smORFs in animal genomes, and highlights the role
of model organisms for the study of small peptide biology in the context of development,
physiology and human disease.

1
Centro Andaluz de Biologia
del Desarrollo, CSIC-UPO,
Sevilla 41013, Spain.
2
Brighton and Sussex Medical
School, University of Sussex,
Brighton BN1 9PX, UK.
Correspondence to J.-P.C.
jpcou@upo.es
doi:10.1038/nrm.2017.58
Published online 12 Jul 2017;
corrected online 1 Aug 2017
Gene sequences were initially identified as containing
open reading frames (ORFs) that are potentially trans-
latable into proteins. More recently, the full molecular
complexity of genes has been exposed, culminating in the
updated concept of a gene that includes both regulatory
regions and transcripts1. Intriguingly, many genes have
been discovered that produce transcripts with mRNA-like
features, such as capping and polyadenylation, but that are
not apparently translated into proteins, and these tran-
scripts are known as long non-coding RNAs (lncRNAs).
These genes and their products have revolutionized our
understanding of gene regulation and RNA metabolism2,3.
There is another class of genetic elements that also
challenges the understanding of the coding potential of
the genome: putatively functional small ORFs (smORFs;
also known as sORFs) of 10 to 100 codons4. Millions of
smORF sequences are found in eukaryotic genomes5–7,
and thousands can be mapped to transcripts, in many
cases, to putative lncRNAs8–10. It is as if we have a genome
within our genome: a hidden genome about which we
know very little. smORFs have been deemed non-coding
on the basis of their short length, which defeats stand-
ard computational detection of protein-coding capacity;
the fact that there has been little experimental corrobor­
ation of their function; and for convenience, as their very
high numbers present a challenge for annotation and
­curation. Consequently, functional smORFs are often
not annotated because they have not been experimen-
tally corroborated, and they have not been corroborated
because they are not annotated, a difficulty which is
rarely (and only serendipitously) overcome.
The problem with computational annotation is
that, as is the case for canonical protein-coding ORFs,
it relies fundamentally on sequence similarities, which
reveal the conservation of the putative coding sequence,
thereby indicating a selective value and hence func-
tion; and the similarity to proteins and protein domains
with an experi­mentally corroborated function, thereby
suggest­ing a similar function for the smORF. However,
true conservation and homology of smORFs is difficult
to establish owing to two fundamental problems: short
sequences accrue lower quantitative conservation scores
(that is, the sum of the scores from each amino acid)
than longer canonical proteins, whereas, reciprocally,
the probability of short sequences obtaining such a ‘low
conservation score’ by chance is higher. For example, the
BLAST search tool penalizes the identification of pro-
tein sequences of fewer than 80 amino acids, and fails
with those that have fewer than 20 (REF. 6). Given the high
­number of smORFs in the genome, and the expectation
that most of them are not functional (see below), arbitrary
cut-offs for a minimal ORF length of 50 or 100 codons are
used in genome annotation, discarding shorter ORFs that
lack experimental evidence of function.

NATURE REVIEWS | MOLECULAR CELL BIOLOGY VOLUME 18 | SEPTEMBER 2017 | 575

©
2
0
1
7
M
a
c
m
i
l
l
a
n
P
u
b
l
i
s
h
e
r
s
L
i
m
i
t
e
d
,
p
a
r
t
o
f
S
p
r
i
n
g
e
r
N
a
t
u
r
e
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.
A N A LY S I S

Obtaining experimental evidence for smORF func-
tion is also difficult. Standard biochemical methods for
protein isolation fail to detect peptides below 10 kD,
which escape a typical gel or filter, and which can be
masked by degradation peptides from larger proteins.
Genetics also encounters problems, as short sequences
such as smORFs offer a small target for random
mutagenesis screens and other gene-discovery proto-
cols, and the large number of smORFs in the genome
(see below) makes it impractical to carry out systematic
directed mutagenesis. In the unlikely event of a smORF
mutation being isolated, it is often assigned to adjacent
canonical genes, as most smORFs are not annotated.
Nevertheless, smORFs are finally receiving atten-
tion and breaking out of this impasse. There is a grow-
ing realiza­tion that hundreds, if not thousands, of
smORFs are translated8,9,11 and that smORF-encoded
peptides (SEPs) can have important functions and can
be widely conserved across metazoans12,13 (reviewed in
REFS 14–16). However, the full repertoire of SEP func-
tions is still unknown, and the genomic functions and
evolutionary impact of smORF sequences have not been
revealed. Experimental characterization of smORFs at a
genomic level in bacteria, yeast and plants showed that
hundreds of smORFS can produce a phenotype5,17,18.
In metazoans, experimental evidence has also accumu-
lated to support these conclusions, although it is still
anecdotal. Some SEPs16,19 (sometimes known as micro-
peptides (REFS 20,21)) are annotated as having biological
activity as antibacterial peptides22, cell signalling mol­
ecules23, cytoskeletal regulators24 and other regulators of
canonical proteins13,25. These functions can be essential
for viability, but only a small minority (a few hundred)
of the putative smORFs in each genome have a suspected
function (inferred by homology)26, and even fewer (tens)
have an experimentally corroborated function14. These
functional smORFs tend to be longer (approximately
80 codons) and thus are more amenable to standard
homology searches, as well as to bio­chemical and genetic
analyses. However, 90% of smORFs are much shorter
(approximately 20 codons), but can display sequence
conservation and evidence of being translated, similarly
to ‘functional’ smORFs6,8,10. Furthermore, even these
shorter smORFs can have crucial developmental and
physiological functions and homologues across great
evolutionary distances12–14. Therefore, there could be
many more uncharacterized smORFs with biomedically
relevant functions, but until now we have not identified
such smORFs, or predicted their functions.
Current experimental evidence shows that, in ­animal
genomes, only approximately 1.2% of smORFs are tran-
scribed, and of these only about one-third seem to be
translated (see below). However, even this small per-
centage of functional smORFs could theoretically prod­
Ribosome profiling uce tens of thousands of uncharacterized peptides in
A technique that globally each animal species. Even if only a small proportion of
probes RNA molecules these peptides have biological activity, we could be miss-
that are being actively ing hundreds of peptides that could shed light on many
translated by ribosomes
by analysing ribosome-
aspects of biology and medicine. Thus, the challenge is
protected RNA fragments how to identify the biologically active smORFs and their
(ribosomal footprints). peptides, or the “beautiful needles in the haystack” (REF. 4).
In this article, we present an analysis of both previ-
ously available and new smORF data. Although there is
evidence for the pervasive use of non-AUG start codons
and the translation of overlapping ORFs14,27,28, we focus
on the population of non-overlapping ORFs with canon-
ical AUG start codons in the reference genomes of three
metazoans (fruit flies, mice and humans). We propose
that sufficient information has accrued to approach
smORFs not as an undersized ‘discard bin’, but as a group
of novel molecular actors with specific characteristics,
evolutionary origin and biological functions at both
the RNA (non-coding) and the peptide (coding) levels.
We present a classification of animal smORFs based
on the characteristics of their sequence and the struc-
ture of their RNAs and encoded peptides, a classifica-
tion that provides predictions on the function of not yet
fully characterized smORFs. Finally, we show evidence
indicating that novel smORFs can randomly and con-
tinuously appear in a­ nimal genomes, and that different
smORF classes r­ epresent steps in the evolution of new
canonical proteins.
Classification of smORFs
Translated Drosophila melanogaster smORFs have been
identified at the genomic level using a combination of
ribosome profiling, peptide tagging and bioinformatics
analyses. Two main groups of translated smORFs have
been identified, according to their profiling metrics and
bioinformatics characteristics, with one such group
enriched in peptides allocated to cell membranes and
organelles8. This classification is important, because it
seemed to identify smORFs with high chances of being
functional. Such a classification could direct research
to more promising smORFs by linking their sequence to
biochemical properties and molecular functions, thereby
greatly reducing the need for exploratory experiments.
Accordingly, we subsequently characterized hemotin,
which is a fly smORF from the membrane-associated
group, and revealed its activity in phagosome matur­
ation and its conservation in vertebrates29. We have
also obtained further experimental and bioinformatics
data that support a functionally relevant classification
of D. melanogaster smORFs. Thus, although this pre-
liminary classification is far from being able to predict
smORF function with the kind of precision that we enjoy
with most canonical proteins, it offers heuristic value,
and therefore we refine and expand it here to include
vertebrate smORFs. We propose the existence of at least
five types of smORFs with distinct transcript organiza­
tion, size, conservation, mode of translation, amino acid
usage (frequency of each amino acid) and peptide struc-
ture properties. We suggest that these five classes are
likely to have different cellular and molecular functions,
from inert DNA sequences to transcribed and translated
cis-regulators of translation, to the expression of func-
tional peptides with the propensity to function as regu-
lators of canonical proteins. Before we elaborate on the
basis of this classification and its functional i­ mplications,
we introduce the five classes below (FIG. 1).
First, intergenic ORFs are the most numerous (96% of
smORFs (FIG. 2a)). These intergenic ORFs are stretches

576 | SEPTEMBER 2017 | VOLUME 18 www.nature.com/nrm

©
2
0
1
7
M
a
c
m
i
l
l
a
n
P
u
b
l
i
s
h
e
r
s
L
i
m
i
t
e
d
,
p
a
r
t
o
f
S
p
r
i
n
g
e
r
N
a
t
u
r
e
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.
 A N A LY S I S

Median size
ORF class RNA type (codons) Translation15 Conservation Coding features Function
Intergenic None 22 None None6,8 Non-canonical AA None
ORFs
uORFs 5´ UTRs of mRNAs 22 Low None8,30 • Nonrandom AA • Non-coding
• No domains • Translation
regulation

Transcribed smORFs
lncORFs lncRNAs 24 Low None8,10 • Nonrandom AA Non-coding or
• No domains coding

Short Short mRNAs 79 High Class • Positively charged AA • Coding

CDSs • Transmembrane • Regulators of
α-helices canonical proteins
Short Spliced mRNAs 79 High Kingdom • Canonical AA • Coding
isoforms • Protein domain • Small interfering
loss peptides
Canonical mRNAs 491 High Kingdom • Canonical AA • Coding42
ORFs • Multiple protein • Structural,
domains enzymatic,
regulatory
Untranslated region DNA RNA splicing
ORFs Other coding sequences Ribosome profiling signal
Figure 1 | Properties of smORFs in fruitflies and mammals. Most open reading frames (ORFs) in animal genomes are
Nature Reviews | Molecular Cell Biology
small ORFs (smORFs) in untranscribed regions (intergenic ORFs; light blue) and are considered non-functional. ORFs of
101 codons or more (canonical ORFs; purple) are generally translated and produce annotated proteins with predictable
functions. In between these two extremes, our genome also encodes transcribed and putatively functional smORFs of
10–100 codons, which can be divided into different classes according to the type of their encoding transcript: upstream
ORFs (uORFs; teal) in the 5ʹ untranslated regions (5′ UTRs) of canonical mRNAs; long non-coding ORFs (lncORFs), which are
present in long non-coding RNAs (green); short coding sequences (short CDSs), which are annotated ORFs of 100 codons
or fewer that are present in short mRNAs (yellow); and short isoform ORFs of 100 codons or fewer, which are generated by
alternative splicing of canonical mRNAs (pink). We extracted all ORFs from both the annotated transcriptomes and the
non-transcribed regions of Drosophila melanogaster, Mus musculus and Homo sapiens, and divided the ORFs into these
classes, which differ in several characteristics: size (indicated as the median number of codons per ORF), average rate of
translation, average taxonomic level of ORF conservation, and features of the encoded amino acids, such as the prevalence
of transmembrane α-helices. We propose that these characteristics correlate with a favoured function for each smORF
class and that they are conserved in flies, mice and humans. The data presented is from this work unless indicated.
The translation and conservation characteristics of the short isoforms is extrapolated from their long isoforms.

Translation efficiency
A measure of the rate of
translation for a given mRNA
feature, obtained in ribosome
profiling experiments. It usually
consists of the ratio between
ribosomal footprints and RNA
sequencing reads generated
by the mRNA region.
of DNA between an ATG and a stop codon, and have
a median size of 22 codons (FIG.  2b). Judging from
high-throughput data, they do not seem to undergo
transcription or translation, and thus it is likely that most
are non-functional, and are instead simple, random con-
sequences of nucleotide permutations in regulatory or
‘junk’ DNA. We suggest the name intergenic ORFs in
order to distinguish them from the transcribed, putatively
functional smORFs that constitute the other classes.
Second, upstream ORFs (uORFs) constitute the
second most abundant class, potentially doubling
the number of currently annotated coding sequences
(FIG. 2a). uORFs are smORFs found in the 5ʹ untranslated
regions (UTRs) of mRNAs that encode canonical pro-
teins, and they also have a median length of 22 codons
(FIG. 2b). Almost 50% of annotated animal mRNAs con-
tain uORFs (FIG. 3a; see also REFS 14,30), and the trans-
lation of a proportion of these uORFs has been reported
in all organisms in which it has been examined, includ-
ing yeast, flies, zebrafish and mice31–34, albeit with low
translation efficiency8,28,30,35. uORFs are thought to regu­
late the translation of the downstream, canonical
ORFs in their transcripts. uORFs have low average
conservation8,30 and their amino acid usage is differ-
ent from random values, but is subtly different from
canonical proteins.
Third, long non-coding ORFs (lncORFs) are smORFs
that are found in putative lncRNAs and are the third
most abundant class (FIG. 2a). They have a median size of
24 codons (FIG. 2b) and their translation character­istics
are similar to those of uORFs: low efficiency and only
detected in one-third of the lncORFs assessed. A typical
lncRNA of approximately 3 Kb could theor­etically con-
tain 100 lncORFs of some 100 codons (30 per frame);
in fact, 98% of annotated lncRNAs contain at least one
smORF in the metazoan species that we have analysed,
with a median of six smORFs per lncRNA (FIG. 3b). Even
the 40 lncRNAs with a non-­coding function that have
been characterized in humans36 contain between 1 and
15 lncORFs. Thus, lncORFs are typically found in poly-
cistronic, sometimes overlapping, arrangements, which
hinders their experimental characterization. Their
amino acid usage is nonrandom, but different from
canonical proteins. Their function is currently gener-
ally unknown, and there is considerable debate about
whether lncORFs are translated, and whether such
translation is productive27,34,37–39. lncRNAs do not seem
to be conserved3,8–10, but several RNAs that were initially
classified as lncRNAs were later shown to encode and
translate peptides with biomedically important func-
tions in development and ­physiology, and to be highly
conserved in evolution12,13,40.

NATURE REVIEWS | MOLECULAR CELL BIOLOGY VOLUME 18 | SEPTEMBER 2017 | 577

©
2
0
1
7
M
a
c
m
i
l
l
a
n
P
u
b
l
i
s
h
e
r
s
L
i
m
i
t
e
d
,
p
a
r
t
o
f
S
p
r
i
n
g
e
r
N
a
t
u
r
e
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.
A N A LY S I S

a Drosophila melanogaster Mus musculus Homo sapiens short isoforms may be higher than currently appreciated
smORFs Annotated because detecting them depends on experimental data,
ORFs 855
18,328 27,988 50,270
and the unambiguous assignment of proteomic and ribo-
828
somal profiling sequences to specific mRNA isoforms
1,207
remains difficult.
16,904 20,430
129 128,307 31,954 43,206 Coding versus non-coding functions
172,115
520 Evidence of transcription and/or translation are two
1,651 objective criteria for assuming that smORFs are func-
tional, whether coding or non-coding. We analysed
Intergenic ORFs: Intergenic ORFs: Intergenic ORFs:
635,402 24,119,916 21,369,201 the genomes and transcriptomes of flies, mice and
humans26,41 and compared the characteristics of differ-
ent RNAs that contain smORFs, and the characteristics
uORFs IncORFs Short CDSs Short isoforms Canonical ORFs of the smORFs themselves.

b Median length (codons) Transcribed smORFs Transcription of smORFs. Extensive RNA-seq studies
in metazoans have been carried out, especially in model
Intergenic Short Short Canonical organisms such as D. melanogaster and Mus musculus.
ORFs uORFs lncORFs CDSs isoforms ORFs
The repertoire of transcribed sequences that these
D. melanogaster 23 20 25 79 82 490
studies provide may not be complete, as not all organs
M. musculus 22 22 23 81 79 424
and cell types have been sampled, but in general most
H. sapiens 22 23 24 78 75 421 transcription should have been detected. The small
Predicted Annotated populations of short CDSs of approximately 79 codons
Figure 2 | Conservation of smORF numbers and lengths in animals. a | Small open (FIG. 2a) in these data sets are found in polyadenylated
reading frame (smORF) classes are colour-codedNature Reviews | Molecular Cell Biology
as in FIG. 1. Circle sizes are proportional monocistronic transcripts that are often annotated as
to the number of smORFs. The number of annotated, short coding sequences (short putatively coding 26,41, even though direct experimental
CDSs) is similar in all the animals analysed, and low when compared with both annotated corroboration of their translation is lacking in most
canonical ORFs and non-annotated transcribed smORFs (upstream ORFs (uORFs) and cases8. Short CDS transcripts are shorter and have fewer
long non-coding ORFs (lncORFs)). The higher number of uORFs and lncORFs in mammals exons than canonical mRNAs (FIG. 3c,d). This could
is related to their higher number of mRNAs and long non-coding RNAs (lncRNAs), ­follow a trend that was observed in eukary­otes — fewer
whereas the number of intergenic, non-transcribed smORFs is related to the genome
exons are found in shorter coding transcripts42. More
size in each species. b | The lengths of different smORF classes are conserved across
metazoans. Intergenic ORFs, uORFs and lncORFs have a median length of 20–25 codons; surprisingly, a large proportion of the transcripts that
short CDSs and short isoforms have a median length of 75–82 codons; and canonical were detected by RNA-seq lacked a canonical ‘long’
ORFs have a median length of 421–490 codons. ORF, and thus have been considered lncRNAs2,3, even
though they contain lncORFs and display coding-like
mRNA features, such as similar length and structure to
Fourth, short coding sequences (short CDSs; previ- short CDSs (FIG. 3c,d), transcription by RNA polymer-
ously known as longer smORFs8,15), have a median size ase II, capping, polyadenylation and accumulation in
of 79 codons (FIG. 2b) and are preferentially found in the cytoplasm43,44,45.
functionally monocistronic transcripts that have mRNA As large numbers of intergenic ORFs are found in
characteristics, albeit they are shorter and s­ impler in non-transcribed genomic regions (FIG. 2a), do they also
structure than canonical mRNAs. They seem to be trans- represent smORFs in uncharacterized transcripts?
lated as efficiently, and are detected as often, as canonical What is their origin and function? Their median size
ORFs8. There are hundreds of short CDSs in flies and of 23 codons across species is expected by chance:
humans (FIG. 2a), but only a small proportion has been among 60 possible codons, 3 are stop codons, which is
functionally characterized. The characterized examples a ratio of 0.05. Counting from an ATG start codon, the
and the average amino acid features of the class suggest length of the resulting ORF depends on the probability
that they have membrane-related functions as regulators of encountering a stop codon. This probability is inde-
of canonical proteins. pendent at each codon, but the accumulated probability
Fifth, short isoforms, which are generated as alter- of encountering any stop codon obviously increases with
native transcripts or as splice forms of longer, canoni- length. Thus, the following exponential decay function
cal protein-coding genes, are the least abundant class of f(x) = λe–λx
smORFs, according to annotated data (FIG. 2a). Annotated where λ (the decay rate parameter) is 0.05 and x is
short isoforms are translated and have a median size of the length of the ORF in codons, indicates the f­ requency
79 codons (FIG. 2b), thereby resembling short CDSs in at which ORFs of each size are expected to occur at
size and in transcript structure, although their amino random, and generates a size distribution that closely
Protein isoforms acid sequences are closer to those of canonical proteins, fits that observed for intergenic ORFs (FIG. 3e). Thus,
Variants of a given protein as expected. Short isoforms merit separate classification intergenic ORFs, unlike short CDSs (FIG.  3f), seem
generated by the translation of
alternative mRNA sequences,
and study on two bases: the certainty of their genetic to be randomly generated by our genomes, so it would
in distinct mRNAs produced by ­origins, and their potential for functions that are directly be expected that most are not functional (just as most
the same gene. related to their canonical protein isoforms. The number of mutations are not advantageous).

578 | SEPTEMBER 2017 | VOLUME 18 www.nature.com/nrm

©
2
0
1
7
M
a
c
m
i
l
l
a
n
P
u
b
l
i
s
h
e
r
s
L
i
m
i
t
e
d
,
p
a
r
t
o
f
S
p
r
i
n
g
e
r
N
a
t
u
r
e
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.
 A N A LY S I S

Other data indicate that we do not have millions of A fundamental filter for considering a smORF as ‘genic’,
genes in our genomes. Although computational gene-­ or as putatively functional, must be the existence of solid
annotation protocols are biased against smORFs6,46, transcriptional data. For example, computational and
classi­cal estimates of gene numbers that were obtained RNA-seq evidence that was used to re‑examine previ-
from biochemical and genetic studies 47 are com­ ous computational estimates of putatively functional
patible with the annotated numbers of genes that are non-annotated smORFs in various species could only
obtained with computational methods, but differ by corroborate a small percentage of these smORFs21. This
several orders of magnitude from the high numbers transcriptional filter does not contradict the possibility
of intergenic ORFs. Therefore, it is likely that most of that some intergenic ORFs could be transcribed and even
the intergenic ORF sequences are not active genes and translated in tissues that have not yet been subjected to
thus, in order to avoid inflating the estimates of func- RNA-seq, but focuses experimental and ­computational
tional smORFs, they should not be considered functional. efforts towards more likely functional targets.

a b
60 15
Relative frequency (%)

Relative frequency (%)

50 uORFs Median IncORFs
40 10
30
20 5
10
0 0
0 2 4 6 8 10 12 14 16 18 20 0 2 4 6 8 10 12 14 16 18 20 ≥21
uORFs per mRNA IncORFs per IncRNA
c 30 d 50
Canonical Canonical
25 40
Short isoforms Short isoforms
Relative frequency (%)

Relative frequency (%)

20 Short CDSs Short CDSs
IncRNAs 30 IncRNAs
15
20
10
10
5

0 0
0 500 1,000 1,500 2,000 2,500 3,000 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
Length (bp) Exons (absolute number per mRNA)
e 6 f 6
Intergenic ORFs Short isoforms
5 5
uORFs Short CDSs
Relative frequency (%)

Relative frequency (%)

4 IncORFs 4 Exponential distribution

Median Exponential distribution λe–λx (λ = –0.05) Median
3 λe–λx (λ = –0.05) 3

2 2

1 1

0 0
20 40 60 80 100 20 40 60 80 100
Length (codons) Length (codons)
Figure 3 | RNA characteristics of transcribed smORFs. The mean and standard error of the mean (SEM) of fruit fly,
Nature Reviews | Molecular Cell Biology
mouse and human values are plotted in dark and light shades, respectively. a | Number of predicted upstream ORFs
(uORFs) per annotated mRNA. Nearly 50% of the animal mRNAs analysed contain one uORF or more. b | Most annotated
long non-coding RNAs (lncRNAs) contain one or more small open reading frames (smORFs) (median = 6). c | Metazoan RNA
transcript length according to ORF classes. mRNAs containing short coding sequences (short CDSs) and lncRNAs are
similar in size (on average, 400 bp in length). Short isoform transcripts are, on average, 600 bp in length, whereas canonical
ORF-containing mRNAs are larger on average (3.1 kb). d | Short CDS-containing mRNAs and lncRNAs have lower
transcript complexity, as measured by the number of different exons in annotated transcripts of each class, than mRNAs
that contain shorter isoforms; canonical mRNAs are, on average, more complex than all of the other classes. e | The size
distributions of uORFs and long non-coding ORFs (lncORFs) in animal genomes are similar to that of intergenic ORFs.
Interestingly, their distributions fit an exponential decay distribution, suggesting that intergenic ORFs, uORFs and
lncORFs are randomly generated in animal genomes. The median (23 codons) of the exponential decay distribution for
smORFs of 10–100 codons is indicated (compare with FIG. 2b). f | The size distributions of short CDSs and short isoforms.
The median length (79 codons) of short CDSs is indicated. Short CDS and short isoform smORFs have similar size
distributions, which are different from the exponential distribution of other smORFs (dotted line, compare with part e).

NATURE REVIEWS | MOLECULAR CELL BIOLOGY VOLUME 18 | SEPTEMBER 2017 | 579

©
2
0
1
7
M
a
c
m
i
l
l
a
n
P
u
b
l
i
s
h
e
r
s
L
i
m
i
t
e
d
,
p
a
r
t
o
f
S
p
r
i
n
g
e
r
N
a
t
u
r
e
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.
A N A LY S I S

ORF tagging
A technique to probe the
translation of a specific open
reading frame (ORF), whereby
a reporter sequence without a
start codon is cloned in‑frame
with the assessed ORF.
Two modes of smORF translation: lncRNAs and short
CDS mRNAs. Ribosome profiling provides quantita-
tive and qualitative measures of translation, both at the
single gene scale and at the genomic scale27,28. It offers
a direct readout of ribosome occupancy on mRNAs at
the single nucleotide level, and can be compared with
RNA-seq data of the same biological sample to provide
quantitative metrics, such as translation efficiency,
which are directly related to the rate of translation at
each ORF (see REF. 48 for a review). Results obtained
from a wide variety of species show that translation
occurs in a more pervasive manner than expected, with
numerous ribosome footprints detected in l­ncRNAs
and in the UTRs of annotated transcripts; many of
these newly identified translated regions coincide with
smORFs8,9,27,49. Although non-coding functions of sev-
eral lncRNAs are well established3,50, the non-coding
functions of the vast majority are currently unknown,
and it is plausible that some of these lncRNAs actually
encode translated smORFs.
Using a variation of ribosomal profiling, smORF
translation in D. melanogaster was shown to occur in
two different modes, which correlated with smORF
class8. There were 220 annotated short CDSs transcribed
in a fly cell line, which were translated at a ­similar effi-
ciency, and are detected as often (about 80% of tran-
scripts were detected as translated in both cases), as
canonical ORFs. Short CDS translation correlated with
mRNA abundance and followed canonical models,
with multiple ribosomes covering the ORF with regu­
lar spacing. Unlike short CDSs, approximately 2,000
uORFs and lncORFs did not follow this canonical mode
of translation, but were detected as translated in one-
third of the cases and showed about half the translation
efficiency of canonical ORFs8. These differences have
also been observed in vertebrates (reviewed in REF. 15).
Ribosomal profiling of zebrafish embryos validated the
translation of 302 previously annotated smORFs, and
identified 190 novel smORFs in previously uncharacter­
ized transcripts and putative lncRNAs, as well as 311
uORFs and 93 ORFs in 3ʹ UTRs9. These smORFs tended
to be more than 50 codons in length and showed con-
servation across vertebrates, which classifies them as
short CDSs. However, the authors noted the existence
of a class of ORFs of fewer than 20 codons with a ribo-
somal profiling signal in lncRNAs, which would belong
to the lncORF class. Similarly, up to 50% of lncRNAs in
mouse embryonic stem cells had a ribosome profiling
signal28, which could potentially give rise to thousands
of peptides49.
There has been a rather technical debate on whether
a low level of ribosome profiling signal represents
prod­uctive translation27,34,37–39. It has been suggested
that some lncRNAs could be associated with the trans­
lation machinery to regulate the translation of canon-
ical mRNAs, without actually being translated; or that
the detected ribosome binding is incidental and non-­
productive; or that the footprints detected are not
gener­ated by ribosomes. In summary, it has been sug-
gested that the ribosome profiling signals in ­lncRNAs
are noise, which yields false positives. However, whereas
there is a linear correlation between the ­levels of canon-
ical mRNAs in polysomes and their ribosome profil-
ing signal8,11,51, such a positive correlation could not be
observed with lncRNAs. Many l­ncRNAs were present
in high quantities in polysomes, but only some ­lncRNAs
produced a ribosomal profiling ­s ignal 8,11, suggest­
ing that the ribosome profiling signal of ­lncRNAs
is not prod­uced by generic background noise, but by
the specific translation of a subset of l­ncRNAs, even
if this occurs at a modest efficiency. Further­more,
lncORF translation has been corroborated by ORF
tagging and proteomics 8,9,11,46 (see below). Finally,
smORFs in transcripts that were initially annotated as­
lncRNAs can produce biologically active peptides with
important functions12–14,20,25,40,52.
An alternative high-throughput method to detect
translation is proteomics, which matches mass spectro­
metry signatures of digested peptides to expected pro-
tein sequences53. Improvements in proteomics methods
have allowed the detection of SEPs, but, in general,
proteomics has lagged behind ribosome profiling in
smORF detection, and the detection of lncORF peptides
has been lacking. Even with the use of specific and new
size fractionation methods and custom libraries that
include non-annotated smORFs, ‘peptidomic’ studies
have only corroborated 8 new SEPs in D. melanogaster
brains54 and 23 SEPs in human cell lines46. Peptides
smaller than 80 amino acids are preferentially degraded
by endogenous proteases55, which hampers their detec-
tion, but they have fewer amino acids to generate two
non-overlapping peptides after in vitro trypsinization,
which is required by standard proteomics validation.
Two ribosomal profiling studies failed to obtain paral-
lel mass-spectrometry evidence for any peptide smaller
than 50 amino acids8,9, which is a common limitation of
proteomic studies of smORF translation. These factors
could explain the low SEP detection, but it could also be
possible that lncORFs only produce unstable peptides
in small quantities, which would be in agreement with
their low ribosome profiling metrics.
Could lncORF-encoded peptides in general have a
biological role or, in other words, are lncORFs functional
and do they have a function that is conveyed by the pep-
tides that they encode? Or are the peptides irrele­vant to
lncORF function? Long ‘non-coding’ RNAs that actu-
ally produce biologically active ­peptides could simply be
misclassified and the annotated ­lncRNAs could actually
comprise two subpopulations: true non-coding lncRNAs
and protein-coding short CDS mRNAs. Alternatively,
some, many or all ­lncRNAs could have dual coding
and non-coding functions, as in the case of the plant
pri‑MIR171b (which produces both a microRNA
and peptides 56), or the mammalian humanin and
MOTS‑c peptides that are produced by mito­chondrial
ribosomal RNA (rRNA) 57,58; or they could simply
produce inactive peptides that are quickly degraded.
lncORF translation may be the functionally relevant pro-
cess, whereas the peptides themselves may convey little
or no function; that is, lncORFs could have a non-coding
function that involves ribosomes. A precedent for such a
scenario is presented by uORFs (see below).

580 | SEPTEMBER 2017 | VOLUME 18 www.nature.com/nrm

©
2
0
1
7
M
a
c
m
i
l
l
a
n
P
u
b
l
i
s
h
e
r
s
L
i
m
i
t
e
d
,
p
a
r
t
o
f
S
p
r
i
n
g
e
r
N
a
t
u
r
e
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.
 A N A LY S I S

uORFs are cis-regulators of translation in canonical SEPs regulate canonical proteins

genes. The function of uORFs is connected to two fail-
safe features of eukaryotic translation: re‑­initiation
and leaky scanning 31,59. In eukaryotic translation, the
small ribosomal subunit (40S) binds to the mRNA at
the 5ʹ cap complex and scans the transcript until it
encounters an AUG codon that is preceded by a 4 nt
CA‑rich Kozak sequence. The 60S ribosomal subunit
then joins to form a complete ribosome (80S) and a
tRNAMet complex joins to initiate translation. Once
translation is terminated at the stop codon, the ribo-
some dissociates from the mRNA, but the 40S sub­
unit can re-initiate scanning for downstream ORFs to
translate. Such re‑initiation of translation can occur if
the initially translated ORF is no longer than 30 amino
acids, and if an additional ORF is found approx­
imately 100–200 bp downstream of the stop codon of
the initially translated ORF. In addition, weak Kozak
sequences are occasionally not recognized, which leads
to leaky scanning in which ORFs may be scanned but
not translated, thereby allowing continued scanning
and translation of d ­ ownstream ORFs.
Both processes are stochastic, but they can facilitate
the translation of polycistronic eukaryotic genes, and can
act as a canonical translation regulatory mechanism31,32.
The classical model of uORF function is the yeast gene
GCN4, in which four uORFs repress the translation of
the Gcn4 protein59,60. uORF translation precludes the
translation of the downstream Gcn4‑encoding ORF,
but in conditions of starvation, the uORFs are scanned
and not translated, allowing the translation of the Gcn4
protein. In this case, the sequences of the peptides that
are produced by the Gcn4 uORFs are irrelevant, and
the physical presence of the uORFs themselves acts
as the regulatory mechanism; in other genes, however,
the uORF-encoded peptides can stall the ribosome
upstream of the main ORFs in a sequence-dependent
manner 61. This inhibitory cis-regulatory uORF function
does not necessarily preclude that some uORF peptides
could have functions in trans that are independent of
their locus62. A pure cis-regulatory role for uORFs fits
with their low translation levels8,30,39, their low sequence
conservation8,30 and their lack of propensity to form
known protein domains (see below). A prediction of
having a repressive cis-regulatory function is that there
should be a negative correlation between the transla-
tion of the uORFs and that of the main ORFs. Such
a genome-wide negative correlation has been found
in mammals32 and zebrafish30, but not in D. melano­
gaster 8 or yeast 35,63, perhaps hinting at undiscovered
uORF functions.
In summary, there is evidence for the translation for
all types of transcribed smORFs, albeit with different
translation efficiencies and chance of detection, which
correlate with their proposed class; in other words,
their size and type of transcript-of‑origin. There are
also well-established smORF functions, which can be
separated into ‘coding’ functions — the production of
a biologically active peptide — and ‘non-coding’ regu­
latory effects of engaging the translation machinery.
We explore the functions of SEPs below.
The small size of SEPs does not support the typical,
multidomain structure of canonical proteins, but instead
accommodates only one or, at most, two simple domains
(considering the simplest domain is a transmembrane
α‑helix (TMH) that is 30 amino acids in length, and
the need for an unstructured spacer region between
domains)64. Interestingly, protein domains in isolation
or incomplete protein domains can have functions that
are unrelated to those observed in their native configur­
ation, as part of large multidomain proteins65. For
example, artificially expressed peptides that contain the
Antennapedia homeodomain or the HIV TAT domain
are cell-penetrating peptides, a function that is unrelated
to that of their endogenous counterparts66. It follows
that even the function of smORFs that contain known
protein domains cannot be easily predicted; in fact, in
most cases, it is still unknown. A bioinformatics examin­
ation of smORF peptide sequences, informed by exam-
ples for which the functions have been experimentally
­characterized, might clarify their roles.
Characterized SEPs. The molecular and organismal
functions of several SEPs have been characterized in
metazoans and in unicellular organisms. We have identi­
fied a group of approximately 80 short CDS-encoded
peptides, which are conserved from flies to humans
and, in some cases, even in yeasts and plants41. The
most ­common function of these ancient SEPs is as pos-
itive regulators of cytoplasmic processes (FIG. 4). These
processes include ubiquitin-like post-translational
modifi­cation67, cytoskeleton dynamics24, translation68
and cyclin function in mitosis69 (FIG. 4a). The second most
common function is related to mitochondria, such as
apoptosis-related functions70 and mitochondrial respir­
ation16,71 (FIG. 4b). These peptides offer concrete examples
of crucial functions that are conserved for hundreds of
­millions of years, but they constitute a minority (8–9%)
of short CDSs in each species. As most smORFs are not
as widely conserved, it is unclear whether these ancient
smORFs offer a functional blueprint for all short CDSs
and, even less so, for lncORFs and uORFs.
Other functionally and molecularly well-character­
ized smORFs are perhaps not as ancient but still show
conservation comparable to canonical proteins, and have
a wider range of functions, as negative regulators. Plants
have short CDSs and short isoforms that encode small
interfering peptides (also known as micro­proteins72)
that function predominantly as transcription repres-
sors73,74. These small interfering peptides are short,
dominant-negative isoforms or duplications of canon-
ical transcription factors, and they usually contain one
known protein domain. They interfere with the function
of canonical transcription factors, either by sequestering
them into unproductive dimers (when the small inter-
fering peptide contains a dimerization domain but not
a DNA-binding domain) or by competing with them
for binding to DNA (when the small interfering pep-
tide contains a DNA-binding domain but lacks other
domains that are required for its activity)72–74. No small
interfering peptides have been characterized in animals,

NATURE REVIEWS | MOLECULAR CELL BIOLOGY VOLUME 18 | SEPTEMBER 2017 | 581

©
2
0
1
7
M
a
c
m
i
l
l
a
n
P
u
b
l
i
s
h
e
r
s
L
i
m
i
t
e
d
,
p
a
r
t
o
f
S
p
r
i
n
g
e
r
N
a
t
u
r
e
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.
A N A LY S I S

Helix–loop–helix
(HLH). A DNA-binding domain
that characterizes members
of a transcription factor family.
It is composed of two α-helices
connected by a short loop.
but there is no reason why they should not exist. There
are examples in humans of dominant-negative peptides
that are close to 100 amino acids in length. The inhibitor
of DNA-binding (Id) family of helix–loop–helix (HLH)-
like peptides, which are 119–168 amino acids in length,
sequesters basic HLH proteins into inactive complexes
(FIG. 4c). They thereby regulate development, the cell cycle
and the circadian rhythms in organisms from flies to
humans, and these peptides have also been implicated in
stem cell renewal and cancer75. Small interfering peptides
could be as prevalent in animals as they are in plants, but
these peptides and the short RNAs that encode them may
have been discarded as artefacts or as non-functional in
studies, even if they are detected experimentally.

a b BCL-2
MOMP Caspase
activation
5′ cap 60S
AAA CKS85A
Humanin Electron transfer
RPS21 40S CDK1
Mitochondrion complexes
CDK2
IV

Z600 I III UQCR10

ATP
Eukaryotic translation Cell cycle progression Oxidative Apoptosis
phosphorylation

c Active Id
Inactive d
homodimer homodimer
PLN or SclA Muscle
HLH Ca2+ contraction
SERCA
On Off
Promoter Promoter ER Cytoplasm Myofibril
ER
membrane
e f Defensin Drosocin
Bacterial cell Transmembrane
Plasma membrane membrane diffusion
CDC42 Phagocytosis Cytoplasm
Hemotin
14-3-3ζ Cytoplasm Pore

SPEC2 Efflux of
cytoplasmic DnaK
PI3K components

Early endosome Bacterial

Mature Degradation metabolism
endosome
Figure 4 | Functions of smORF-encoded peptides. Ancient small open reading frame (smORF)-encoded peptides
Nature
(SEPs; blue ellipses) are conserved across eukaryotes (yeast, insects and vertebrates) Reviews
and act | Molecular
as positive Cell
regulators ofBiology
canonical proteins. Evolutionarily more-recent SEPs (pink ellipses) act as negative regulators. a | Ancient cytoplasmic SEPs
promote the activity of canonical protein complexes. 40S ribosomal protein S21 (RPS21; 83 amino acids in length) is a
component of the small ribosomal subunit (40S) and aids in mRNA translation; cyclin-dependent kinase subunit 85A
(CKS85A; 96 amino acids in length) interacts with cyclin-dependent kinase 1 (CDK1) and CDK2 and promotes cell cycle
progression. Its activity is opposed by the negative regulator Z600, which is an SEP that represses CDK1. b | Ancient SEPs
that are produced in the cytoplasm can promote mitochondrial processes. BCL-2 peptides are localized to the
mitochondrial outer membrane to promote mitochondrial outer membrane permeabilization (MOMP) and the subsequent
release of apoptotic signals through caspase activation. Other SEPs, such as ubiquinol-cytochrome c reductase complex III
subunit 10 (UQCR10), are involved in electron transport at the inner mitochondrial membrane. The evolutionarily recent
humanin peptide is produced by a mitochondrial ribosomal RNA (rRNA) and represses BCL-2 function. c | Small interfering
SEPs act as dominant-negative repressors of transcription factors. The Id (inhibitor of DNA-binding) peptides contain a
helix–loop–helix (HLH) domain, but lack a basic DNA-binding domain. Id peptides bind to HLH transcription factors
and sequester them in inactive complexes that cannot bind to DNA. d | Sarcolamban A (SclA; 28 amino acids in length)
is a Drosophila melanogaster SEP. Similarly to its human orthologue phospholamban (PLN), SclA regulates calcium uptake
in the endoplasmic reticulum (ER) of muscle cells through repression of SERCA, which is the Ca2+ pump that transfers Ca2+
from the cytosol to the ER to terminate muscle contraction. e | SEPs can repress the activity of canonical proteins in
phagocytic membranes and organelles. Hemotin29 represses the phosphoinositide 3‑kinase (PI3K) activator 14‑3‑3ζ, thereby
slowing endosomal maturation to allow phagocytic digestion, whereas small effector of CDC42 protein 2 (SPEC2) represses
the small GTPase CDC42 to regulate the formation of phagocytic pockets. f | Antimicrobial peptides penetrate the cell
membranes of microorganisms. Defensin creates membrane pores that lead to the leakage of cytoplasmic content;
drosocin binds to and interferes with cytoplasmic proteins, such as the chaperone DnaK, to slow bacterial metabolism.

582 | SEPTEMBER 2017 | VOLUME 18 www.nature.com/nrm

©
2
0
1
7
M
a
c
m
i
l
l
a
n
P
u
b
l
i
s
h
e
r
s
L
i
m
i
t
e
d
,
p
a
r
t
o
f
S
p
r
i
n
g
e
r
N
a
t
u
r
e
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.
 A N A LY S I S

Small interfering and regulatory peptides could be a
general feature of SEP function, as dominant-negative
interference does not need to be limited to transcription
factors. This fits well with the small size of SEPs, which
cannot form the large globular proteins with ­buried
active sites that are characteristic of enzymes, or the
large multidomain structural proteins64,76. However,
SEPs could be perfect for interfering with larger pro-
teins. Indeed, some short CDS-encoded peptides have
a demonstrated negative regulatory role in mitosis77
(FIG. 4a), apoptosis78 (FIG. 4b), ubiquitylation52,79, muscle
contraction13,20 (FIG. 4d), phagocytosis29,80 (FIG. 4e) and as
antimicrobial peptides22,81,82 (FIG. 4f). Interestingly, most
of these functions involve cell membranes.
Protein structure and amino acid usage of smORFs
without annotated functions. In baker’s yeast (Saccharo­
myces cerevisiae), 299 non-annotated smORFs with
evidence of transcription, translation or sequence con-
servation were tested in genetic experiments5. Of these,
247 were required for cell growth during starv­ation
and stress conditions. In Escherichia coli, 217 smORFs
were identified by bioinformatics criteria, of which
24 were tested and 18 were found to be required for
bacterial growth17. Of these, 10 SEPs were observed
at membranes and were predicted to encode TMHs,
­similarly to 65% of annotated bacterial proteins of fewer
than 50 amino acids17. These genetic studies indicate
that many uncharacterized SEPs may not be absolutely
required for survival, but may improve overall fitness,
similarly to the genetic requirement for the negative
regulatory SEPs (discussed above)14,22. These studies
also suggest that many uncharacterized SEPs may be
located at cellular membranes (thereby complicating
their biochemical detection).
In D. melanogaster, a bioinformatics analysis revealed
more predicted TMHs in short CDSs than in canoni-
cal proteins, lncORFs and uORFs8. This trend is shared
by vertebrate smORFs (FIG. 5a). TMHs could easily be
related to putative functions in cell membranes and
organelles13,17,20,29. Indeed, the limited gene ontology data
available in flies displays an enrichment for membrane-
r­elated terms8, and tagging a sample of short CDS pep-
tides revealed a tendency for these peptides to localize to
cell membranes, including mitochondrial membranes8.
Finally, the characterization of hemotin29 (FIG. 4e), as well
as that of other short CDSs (E. Magny and J. I. Pueyo,
personal communication) corroborates the predic-
tion of membrane- and organelle-related functions for
TMH‑containing short CDS-encoded peptides.
The amino acid composition of smORFs could also
be a source of information pertaining to their potential
functionality. In flies, we observed putatively different
amino acid usage among canonical proteins, short CDSs,
lncORFs, uORFs and random RNA sequences8. These
differences could underlie different molecular functions
and thus could be an indicator of coding potential83.
We analysed the amino acid usage of mouse and human
smORFs and randomized RNA sequences in compari-
son to flies. We found a remarkable degree of similarity
between the fly and the vertebrate amino acid frequencies
in each smORF class (FIG. 5b). Pooling the data from flies,
mice and humans reveals significant correlations and
differences in the amino acid usage of different smORF
classes, and when compared with canonical proteins
and randomized RNA. As observed in flies, the amino
acid compositions of metazoan canonical proteins, short
isoforms and short CDSs resemble each other and differ
significantly from those of randomized RNA sequences
and intergenic smORFs (FIG. 5c,d).
Despite their overall similarity to canonical proteins,
short CDSs have an increased frequency of some pos­
itively charged amino acids, and that of some negatively
charged amino acids is decreased, thereby producing
an overall positive charge bias (FIG. 5d). Artificial cell-­
penetrating peptides have an overall positive charge and
can cross plasma membranes, which is of great interest
to the pharmaceutical industry 84,85. Given the preva-
lence of TMHs in short CDS peptides (FIG. 5a) and the
membrane localization of short CDS peptides (FIG. 4),
in particular to the negatively charged mitochondria8,16,
it is tempting to speculate that this charge bias similarly
favours their traffic across membranes and organelles.
Given that amphipathic cations (molecules that contain
both positively charged and hydrophobic regions) can
cross the mitochondrial outer membrane86, short CDS
peptides may also be able to cross membranes.
The amino acid usage of short CDSs would also fit
with a role as antimicrobial peptides. Antimicrobial pep-
tides belong to the humoral (macromolecule) branch of
innate immunity, and their production is regulated by
a signalling mechanism that is conserved from flies to
humans81,87. Antimicrobial peptides tend to be amphi­
pathic, approximately 50–150 amino acids in length and
have a propensity to form TMHs88,89. Their amphipathic
nature confers solubility and the ability to bind to and
integrate into microbial membranes. These character­
istics, which are identical to those found for short CDS
peptides (FIG. 5a,d), are sufficient for the design of artifi­
cial antimicrobial peptides that act even more effici­
ently than natural peptides88. Antimicrobial peptides
can form pores in the membranes of microorganisms,
leading to cell leakage and death81, but can also behave
as cell-penetrating peptides that, once inside the cells,
can interfere with vital cellular processes81,82,89. In this
regard, antimicrobial peptides could also be seen as
negative regulators89. As the overall molecular character­
istics of antimicrobial peptides are identical to those
of short CDS peptides, some of the hundreds of short
CDS peptides with currently unknown function might
act as antimicrobial peptides. Indeed, several well-­
characterized antimicrobial peptides are encoded by
short CDSs22,81,82 (FIG. 4f).
Another possible function of positively charged
peptides is to bind to nucleic acids. DNA and RNA
are negatively charged, and transcription factors and
other DNA-binding proteins, such as histones, bind to
DNA and to RNA through positively charged domains.
As discussed above, plant small interfering peptides
are transcription regulators, and animal SEPs with
possible DNA-binding activities have been described,
such as the human MRI‑2, which is a regulator of the

NATURE REVIEWS | MOLECULAR CELL BIOLOGY VOLUME 18 | SEPTEMBER 2017 | 583

©
2
0
1
7
M
a
c
m
i
l
l
a
n
P
u
b
l
i
s
h
e
r
s
L
i
m
i
t
e
d
,
p
a
r
t
o
f
S
p
r
i
n
g
e
r
N
a
t
u
r
e
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.
A N A LY S I S

DNA end-joining protein KU90; the fly SEP Pgc, which of randomized RNA (FIG. 5c). uORFs resemble short
epigenetically represses transcription25; and, interest- CDSs in having a slight positive charge bias (FIG. 5d).
ingly, those produced by the dual-function (coding and If  uORF-derived peptides have no function, their
non-coding) RNA pri‑MIR171b56. ‘coding-­like’ amino acid usage might be an irrelevant
The amino acid usage of lncORFs and uORFs is consequence of their location near canonical ORFs. It is
intriguing. Overall uORF amino acid propensities corre- possible that uORFs are derived ‘nonsense’ fragments
late highly with those of canonical ORFs, short isoforms of nearby canonical ORFs. An interesting, alternative
and short CDSs (FIG. 5c), but they also correlate highly possibility is that uORF peptide sequences could reflect
with those of lncORFs, and somewhat less with those a specific but as-yet undiscovered function62.

a 0.5 P = 0.0065

0.4 ns
TMHs per 100 AAs

b
P = 0.0133
0.3
ORF class Fly–mouse Fly–human Mouse–human
ns uORFs 0.77 0.76 0.99
0.2
lncORFs 0.92 0.92 0.98
Short CDSs 0.96 0.96 0.99
0.1
Short isoforms 0.94 0.95 0.99
Canonical ORFs 0.97 0.97 0.99
0.0
s

s
RF

RF

RF

rm

DS
uO

cO

lO

fo

tC
so
ca
In

or
ti
ni

Sh

d
or
no

Sh
Ca

Short CDSs
Difference from canonical AA frequencies (%)

3
c lncORFs
Intergenic

isoforms
Random

lncORFs

uORFs
uORFs

2
ORFS

Short

Short
CDSs

1
Canonical ns
ORFs
0
Short ns
CDSs
–1
Short ns
isoforms
–2
lncORFs
–3
uORFs
MC K RH EDQTSN W Y F I V P G A L
Intergenic
ORFs MC K ED S Short CDSs
Significant

MC ED A IncORFs
AAs

MC R ED uORFs
0 0.2 0.4 0.6 0.8 1
Pearson’s r S + – P, H
Figure 5 | Coding features of smORFs. a | Number of putative transmembrane α-helix domains
Nature (TMHs;
Reviews predicted
| Molecular Cellby
Biology
TMHMM 2.0) encoded per 100 amino acids (AAs) in each open reading frame (ORF) class. Short coding sequences
(short CDSs) are significantly enriched in TMHs compared with canonical ORFs (p values for t‑tests indicated), whereas
upstream ORFs (uORFs) are significantly depleted. Mean and standard error of the mean (SEM) of fly, mouse and human
values are plotted. b | Correlation of amino acid composition across flies, mice and humans by small ORF (smORF) class,
quantified as Pearson’s r values for each pairwise correlation. All correlations are significant (P < 0.05). c | Correlation of
overall amino acid composition among smORF classes using pooled values from flies, mice and humans, quantified as in
part b. Short CDSs and short isoforms most resemble canonical ORFs and differ from random sequences; long non-coding
ORFs (lncORFs) and uORFs display an amino acid composition that is intermediate between canonical ORFs and
intergenic ORFs and random values. d | Amino acid propensities of smORFs, normalized to canonical propensities
(data as in part a). Some short CDS amino acid frequencies differ significantly from canonical ORFs (multiple t‑tests with
Bonferroni corrections, P < 0.05). Some short CDSs encode more sulfur-containing amino acids, as well as being enriched
in positively charged and depleted in negatively charged amino acids. lncORFs and uORFs resemble each other and
display a pattern related to short CDSs, but with more extreme variations. Short isoforms are not plotted, as their only
significant difference from canonical ORFs is Met usage, the higher frequency of which is attributable to their shorter
length. Amino acids with significantly different frequencies (enrichment or depletion) from canonical ORFs are indicated.
ns denotes correlations that are not statistically significant (P > 0.05). +, positively charged amino acids; –, negatively
charged amino acids; H, hydrophobic; P, polar; S, sulfur-containing amino acids.

584 | SEPTEMBER 2017 | VOLUME 18 www.nature.com/nrm

©
2
0
1
7
M
a
c
m
i
l
l
a
n
P
u
b
l
i
s
h
e
r
s
L
i
m
i
t
e
d
,
p
a
r
t
o
f
S
p
r
i
n
g
e
r
N
a
t
u
r
e
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.
 A N A LY S I S

Transcription
Translation ‘Elongation’

Intergenic ORFs

lncORFs
Short CDSs
Random smORF
generation

Duplication Canonical ORFs

Random
ORF loss
Transcribed
pseudogenes Short isoforms
Pseudogenes

Pseudogenization Truncation
Loss of transcription

Coding potential Translation Coding Protein Canonical

efficiency function structure protein
content
DNA RNA splicing
Protein
Untranslated region Ribosome function
ORF profiling signal

Figure 6 | Stepwise evolution of smORFs towards proteins. Stepwise model for small open reading frame (smORF)
evolution towards new canonical ORFs. It has been suggested that protein-coding Nature
genesReviews
emerge| Molecular Cell Biology
from a continuum of
non-functional sequences (‘protogenes’ (REF. 100)). We propose that the various classes of smORFs (FIG. 1) represent
different steps in this process. Intergenic ORFs appear at random in non-transcribed DNA, but during evolution they can
become part of a transcription unit, giving rise to long non-coding ORFs (lncORFs). lncORFs (and upstream ORFs (uORFs))
can also appear randomly in transcribed sequences, and have been shown to be translated in up to a third of the cases and
with a low translation efficiency. The main functional outcome of this low level of translation may not be to produce
biologically active peptides, but to provide the organism with a reservoir of peptides that are translated at a low level,
which can be subjected to natural selection. The translation efficiency of these coding lncORFs could increase, giving rise
to short coding sequences (short CDSs). In turn, short CDSs could grow or integrate into canonical proteins (‘elongation’).
The creation of new canonical ORFs is counterbalanced by their conversion into shorter isoforms, pseudogenes and
perhaps lncRNAs111, and also by the random disappearance of intergenic ORFs through the loss of ATG codons.

The amino acid usage of lncORFs resembles that of
short CDSs and uORFs in its higher proportion of sulfur-­
containing amino acids (Met and Cys) and lower pro-
portion of the negatively charged residues Asp and Glu
(FIG. 5d). However, lncORFs are overall most similar to
random amino acid usage (FIG. 5d). It is not clear that these
propensities would confer an overall positive or amphi-
pathic nature to lncORF-derived peptides, and thus we
cannot speculate on a putative membrane function for
lncORFs. Nevertheless, the sarcolamban (Scl) family of
SEPs, which function in the endoplasmic reticulum (ER)
of muscle cells (FIG. 4d), are encoded by transcripts that
were previously annotated as lncRNAs, and are approx-
imately 30 amino acids in length13,20, which is typical of
lncORFs. Similarly, the Tal SEPs are 11–32 amino acids
in length and influence adjacent cells23,91, implying that
they diffuse across membranes. Translated lncORF
peptides were observed at mitochondria and other
organ­elles8, and human mitochondrial rRNAs prod­
uce SEPs, such as humanin (24 amino acids in length)
(FIG. 4b), which is a generic inhibitor of cell death of bio-
medical importance57, and MOTS‑c (16 amino acids
in length), which acts outside the mitochondria to regu-
late insulin activity 58, also implying that it crosses mem-
branes. However, the short lncORF peptides might be
avidly degraded55, which would hinder a possible cell-­
penetrating function92. There are too few examples of
characterized lncORF-­encoded peptides, and more func-
tional studies of lncORFs are needed. Similarly to uORFs,
lncORFs might be a ‘mixed bag’, containing c­ oding
smORFs (thus actually being short CDSs) and non-­
coding smORFs, or else they might represent a group of
sequences that are poised for coding function but that
have not yet done so. We explore this possibility below.
The genomic role of smORFs in the birth of genes
There are indications that smORFs have a general geno­
mic function as a source of new protein-coding genes.
Despite the marked differences in conservation, average
lengths, amino acid usage, translation efficiency, protein
structure and functionality of the different classes of
smORFs, there is a degree of overlap among the classes
(FIG. 3), which may suggest the existence of an evolution-
ary continuum among classes (FIG. 6). We examine below

NATURE REVIEWS | MOLECULAR CELL BIOLOGY VOLUME 18 | SEPTEMBER 2017 | 585

©
2
0
1
7
M
a
c
m
i
l
l
a
n
P
u
b
l
i
s
h
e
r
s
L
i
m
i
t
e
d
,
p
a
r
t
o
f
S
p
r
i
n
g
e
r
N
a
t
u
r
e
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.
A N A LY S I S

Pseudogene
A paralogue of a functional
protein-coding gene, which has
lost its gene expression and/or
protein-coding capacities.
Paralogue
Homologous gene within
a given species, usually
generated by gene duplication.
the two possibilities for the generation of smORFs: from
existing protein-coding sequences, or de novo formation
from previously non-coding sequences.
smORFs can emerge as fragments of longer protein-­
coding genes through alternative RNA processing,
intron retention and premature stop codons. Short pro-
tein isoforms can be generated by alternative transcrip-
tion, splicing and polyadenylation (or a combination of
the three) from canonical proteins. It seems that higher
eukaryotes have leaky splicing mechanisms93, and this
can lead to the production of short isoforms with proper
mRNA and translation features. We have seen that short
isoforms can produce dominant-negative peptides,
which could have deleterious consequences, in a simi-
lar manner to amyloid-β peptides of 36–43 amino acids
in length, which form plaques in Alzheimer disease94.
However, if the deleterious consequences are small,
late-onset (past reproductive age), or pleio­tropically
linked to positive traits95, even such ‘deleter­ious iso-
forms’ could be temporarily carried by our genomes.
A short isoform could be selected as a negative regu-
lator and could eventually become an independent
gene following a retrotransposition or a gene dupli-
cation event. After further evolution and divergence,
this duplicated isoform could become a pseudogene
or, alternatively, a new short CDS (FIG. 6). The similar
sizes (FIG. 3f) and amino acid usage (FIG. 5) of short CDSs
and short isoforms could indicate that short CDSs have
been generated in this manner, especially those with
protein domains or with clear homology to canonical
proteins. Alternatively, such paralogue short CDSs could
also arise directly from canonical ORFs. Splice junction
mutations and intron retention can introduce intronic
sequences into a protein, but introns contain stop codons
either by chance, or by selection, presumably to stop the
production of abnormal, long proteins93. Thus, intron
translation would lead to the production of truncated
proteins with new carboxyl termini, and long 3ʹ UTRs.
In this scenario, evolution would be expected to favour
smORF-producing transcripts, if the encoded peptides
are advantageous. Altogether, paralogue short CDSs
that emerge from canonical proteins could be part of
canonical protein evolution, and represent an opposing
mechanism to processes that usually inactivate proteins
and that result in the formation of pseudogenes (FIG. 6).
Other short CDSs, such as hemotin 29,96 or toddler
(also known as elabela)97,98, seem to have no paralogues
in the genome, so we cannot assign their origin to a pre-­
existing canonical protein-coding gene. Where do these
‘singletons’ come from? Short CDSs could evolve from
shorter lncORFs and uORFs, by mechanisms favouring
‘ORF extension’. In principle, the extension of any ORF
only necessitates changing the stop codon to the codon
for an amino acid, which is the most likely outcome
(95%) in the event of stop codon mutation, as observed
in the Tal genes12 (FIG. 7a). Stop codon readthrough,
an event that is readily detected in vivo by ribosome
profiling 99, could be an alternative means, or an inter­
mediate step, to stop codon mutation during ORF exten-
sion. In either case, such elongated ORFs will again be
subjected to the exponential function λe–λx (see above),
moderating elongation to 25‑amino-acid-long steps.
Alternatively, amino‑terminal elongation (as observed
in the PLN peptides of the Scl family (FIG. 7b)) could also
occur, although this necessitates a new in‑frame ATG in
an appropriate Kozak context. Either way, if such elong­
ated peptides preserve their original function, then
they could be positively selected owing to the higher
stability that results from increased length55. In time,
the elong­ated ends of the peptide would be subjected
to selection to improve the peptide function or to add
new functions.
The size distribution of lncORFs and uORFs fits
closely the exponential random distribution of inter-
genic ORFs (FIG. 3e), suggesting that they also appear ran-
domly in the genome. lncORFs and uORFs may evolve
from intergenic ORFs that become transcribed (ORF
first), or appear directly in non-coding RNA sequences
(RNA first). The possible de novo generation of proteins
from previously non-coding sequences is increasingly
debated100,101. Up to 1% of protein-coding genes could
be species-specific (without homologues) and therefore
of recent origin102–105, but this idea is controversial106 and
strongly depends on the computational ability to detect
homologues. The mechanism for the emergence and
the spread of such de novo genes has not been clarified,
although a role for lncORFs has been proposed10,107.
Most lncORFs and de novo genes seem devoid of dis-
tant homologues, appearing and disappearing in the
genome of different species in the same Order 10,104,107,108,
suggesting both recent origins and dynamic evolution-
ary behaviour. However, the Tal and Scl families of short
CDSs have been conserved over hundreds of millions of
years of evolution12,13 (FIG. 7a,b), showing that lncORFs
can become short CDSs and can be fixed in the genome,
perhaps in correlation with increased coding potential.
Further observations are compatible with the evolution
of lncORF and uORFs to short CDSs and canonical
coding-genes (FIG. 6). First, short CDSs have an inter-
mediate evolutionary age, as they are more conserved
than uORFs8,30 and lncORFs8,10, but less conserved than
canonical ORFs (FIG. 7c). Second, smORFs reveal a pro-
gression towards canonical amino acid usage, starting
from intergenic ORFs to lncORF and uORFs to short
CDSs, and then to short isoforms and canonical pro-
teins (FIG. 5c,d). Finally, although the size of lncORFs and
uORFs suggest a random origin from intergenic ORFs
or non-coding RNA sequences (FIG. 3e), their amino acid
sequences do not fully reflect such an origin, but instead
more closely resemble coding ORFs (FIG. 5c). In conclu-
sion, lncORFs and uORFs may appear at random, but
once they appear their nucleotide sequences are sub-
jected to selection. Whether this selection is initially
due to a coding or a non-coding function needs to be
ascertained; however, it can result in the development of
amino acid sequences with full peptide function12,13,20,40,52.
In summary, some smORFs might have emerged
from canonical genes, and others from non-coding
sequences. Either way, the processes that drive the
evolution of canonical proteins (gene duplication and
protein neofunctionalization109) should also act on
smORFs and give rise to smORF ‘families’, offering yet

586 | SEPTEMBER 2017 | VOLUME 18 www.nature.com/nrm

©
2
0
1
7
M
a
c
m
i
l
l
a
n
P
u
b
l
i
s
h
e
r
s
L
i
m
i
t
e
d
,
p
a
r
t
o
f
S
p
r
i
n
g
e
r
N
a
t
u
r
e
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.
 A N A LY S I S

more raw materials for smORF evolution. Indeed, short
CDSs and short isoforms can undergo duplication, both
inside their transcript and as part of gene families12,13,72–75
(FIG. 7a,b), and can be generally classified according to
their sequence similarity (J.-P.C. and P.P., unpublished
observations). Finally, smORFs might not only appear
from or develop into new canonical proteins, but might
also be added to existing canonical proteins by exon
shuffling 110, thereby driving the development of new
protein domains. In our view, smORFs could repre-
sent a genomic protein factory, which utilizes materials
both new (lncORFs and intergenic ORFs) and recycled
(canonical genes), constantly producing putative new
peptides and protein domains.
Conclusions and future perspectives
Bioinformatic and experimental limitations have
­hampered the study of smORFs and the identification
of functional smORFs. Overcoming these limitations
and increasing the pool of experimentally characterized
smORFs is the foremost challenge in the field. CRISPR–
Cas-based gene editing should herald a new phase of
faster progress, by allowing the targeted manipulation
of individual smORFs. This is especially important in the
case of lncORFs, uORFs and small isoforms, in which
specific ORFs within the transcript or gene must be
mutated individually. Nonetheless, our analysis of
the available data suggests some emerging principles
of smORF classification and function.

a c
100 Canonical ORFs
tal-1A Short CDSs
RNA atggcagcctacttggatcccactggccagtactaa
80
Protein M A A Y L D P T G Q Y * Internal
duplications
and 60

% of ORFs
1A 2A 3A AA START and
tal STOP codon
degeneracy 40

tal-AA 20
RNA atgctggatcccactggaacataccggcgaccacgcgacacgcaggac
Protein M L D P T G T Y R R P R D T Q D 0

n
m
r

s
tcccgccaaaagaggcgacaggactgcctggatccaaccgggcagtactag

ai
as
RNA

do
ylu
rd

m
Cl

ng
O

Do
Ph
Protein S R Q K R R Q D C L D P T G Q Y *

Ki
Conservation level

b Disordered region TMH

Structure

Human PLN MEKVQYLTRSAIRRASTIEMPQQARQKLQNLFIN--FCLILICLL-LICIIVMLL--

Fruit fly SclA MSEARNLFTT--FCILAICLF-FLYLIYAVL-- Multiple sequence
Human SLN MGINTRELFIN--FTIVLITVI-LMWLLVRSYQY alignment
Human MRLN -----------MTGKNWILISTTTPKSLEDEIVGRLLKILFVIFVDLISIIYVVITS
Human DWORF---------------------MAKGSTFSHLLVP--ILLLIGWIVGCIIMIYVVFS-

Conservation level
Figure 7 | Evidence of smORF evolution. a | Evolution of tarsal-less (tal) small open reading frames (smORFs) in
Drosophila melanogaster. tal is a polycistronic protein-coding gene, which had been initially
Nature annotated
Reviews as a long
| Molecular Cell Biology
non-coding RNA (lncRNA) but which contains four short coding sequences (short CDSs). Three of these short CDSs are
11 and 12 codons in length (Tal‑1A, Tal‑2A and Tal‑3A) and include a conserved functional heptapeptide LDPTGXY,
indicating that they were formed through tandem duplications, a process corroborated by the tal phylogenetic tree12.
The smORF tal‑AA (32 codons in length) is the most evolutionarily recent Tal short CDS12 and encodes a sequence with
two heptapeptides, separated by sequence that is 17 amino acids in length; it includes degenerate STOP (orange shading)
and START codons (green shading), which is consistent with the elongation and fusion of two shorter tal‑1A‑like smORFs
by STOP codon loss. * denotes extant STOP codons. b | The sarcolamban (Scl) family of short CDSs show low amino acid
sequence conservation, but their structure and molecular function are conserved13,20. The family includes two fly peptides
(only SclA is shown) and four mammalian peptides. They all repress the activity of SERCA (FIG. 4d), except DWORF, which,
on the contrary, promotes SERCA activity by acting as a competitive inhibitor of the other Scl family smORF-encoded
peptides40. Both myoregulin (MRLN) and phospholamban (PLN) show amino‑terminal extension of their sequences.
c | The evolutionary conservation of canonical ORFs and short CDSs, indicating the proportion of ORFs that are conserved
at each taxonomic level (data extracted from Ensembl and Flybase). Solid graph lines indicate the mean and coloured
areas indicate the standard error of the mean (SEM), for data from fruit flies, mice and humans. One-half of canonical ORFs
are conserved across the Kingdom, whereas short CDSs display a 50% conservation at the Class level. However, note that
approximately 75 (8%) of ancient short CDSs are conserved across the domain Eukarya. Conservation in the following
species was used to indicate conservation at each taxonomic level. D. melanogaster ORFs: Kingdom if conserved in
Mus musculus or Homo sapiens; Phylum if in Daphnia pulex or Ixodes scapularis; Class if in Tribolium castaneum;
Order if in Anopheles gambiae. M. musculus ORFs: Kingdom conservation if in D. melanogaster; Phylum if in Danio rerio;
Class if in H. sapiens; Order if in Rattus norvegicus. H. sapiens ORFs: Kingdom if in D. melanogaster; Phylum if in D. rerio;
Class if in M. musculus; Order if in Pan troglodytes. For all ORFs, conservation in Saccharomyces cerevisiae indicated
conservation across the domain Eukarya. TMH, transmembrane α‑helix.

NATURE REVIEWS | MOLECULAR CELL BIOLOGY VOLUME 18 | SEPTEMBER 2017 | 587

©
2
0
1
7
M
a
c
m
i
l
l
a
n
P
u
b
l
i
s
h
e
r
s
L
i
m
i
t
e
d
,
p
a
r
t
o
f
S
p
r
i
n
g
e
r
N
a
t
u
r
e
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.
A N A LY S I S

smORFs across animal species can be classified
according to sequence length and transcript structure.
These features correlate with other characteristics, such
as their evolutionary conservation and amino acid
usage (FIG. 1). Furthermore, these smORF classes seem
to have specific cellular and molecular functions, facili­
tating more detailed studies and an understanding of
the genomic role of smORFs. Non-transcribed inter-
genic ORFs probably have no function; uORFs act as
cis-regulators of the translation of downstream canon-
ical ORFs; lncORFs can give rise to novel biologically
active peptides; short CDSs produce peptide regulators
of canonical proteins in the cytoplasm and membranes
(FIG. 4); and short isoforms can produce peptides that
interfere with homologous proteins (FIG. 4c). Altogether,
smORFs may generate new protein sequences during
evolution, with the different smORF classes representing
steps in the evolution of proteins from inert intergenic
sequences (FIG. 6).
Understanding the origin, evolution and function
of smORFs would be needed to clarify this crucial and,
so far, underappreciated function of the genome; a far
more dynamic and living genome than we currently
contemplate. In addition, considering the current inter-
est in artificial peptides as new clinical agents, smORFs
could provide an unexplored reservoir of peptides that
either already function as new drugs, delivery vectors
and antimicrobial agents (such as short CDS), or are cur-
rently inactive but naturally primed for such functions
by virtue of their amino acid composition (such as uORF
and lncORF peptides). The conservation of individual
smORFs and smORF classes across animal species,
which we discuss in this Analysis, offers experimental
model systems to explore this fascinating field.

1. Gerstein, M. B. et al. What is a gene, post-ENCODE?
History and updated definition. Genome Res. 17,
669–681 (2007).
2. Derrien, T. et al. The GENCODE v7 catalog of human
long noncoding RNAs: analysis of their gene structure,
evolution, and expression. Genome Res. 22,
1775–1789 (2012).
3. Guttman, M. & Rinn, J. L. Modular regulatory
principles of large non-coding RNAs. Nature 482,
339–346 (2012).
4. Basrai, M. A., Hieter, P. & Boeke, J. D. Small open
reading frames: beautiful needles in the haystack.
Genome Res. 7, 768–771 (1997).
This seminal work effectively establishes the field
of smORF studies by arguing that smORFs exist in
large numbers and can encode functional peptides.
5. Kastenmayer, J. P. et al. Functional genomics of genes
with small open reading frames (sORFs) in
S. cerevisiae. Genome Res. 16, 365–373 (2006).
The only genome-wide assessment of smORF
function, demonstrating smORF function in
approximately 5% of baker’s yeast genes.
6. Ladoukakis, E., Pereira, V., Magny, E. G., Eyre‑Walker, A.
& Couso, J. P. Hundreds of putatively functional small
open reading frames in Drosophila. Genome Biol. 12,
R118 (2011).
7. Frith, M. C. et al. The abundance of short proteins in
the mammalian proteome. Plos Genet. 2, 515–528
(2006).
8. Aspden, J. L. et al. Extensive translation of small
open≈reading frames revealed by Poly-Ribo-Seq.
eLife 3, e03528 (2014).
9. Bazzini, A. A. et al. Identification of small ORFs
in vertebrates using ribosome footprinting and
evolutionary conservation EMBO J. 33, 981–993
(2014).
References 8 and 9 represent the first two studies
of smORF translation using ribosome profiling in
animals. Reference 8 introduces the concept that
smORFs can be divided into different categories
according to sequence features and translation
efficiency.
10. Ruiz-Orera, J., Messeguer, X., Subirana, J. A.
& Alba, M. M. Long non-coding RNAs as a source
of new peptides. eLife 3, e03523 (2014).
This computational study shows that the
conservation and translation metrics of lncORFs
resemble those of evolutionarily young proteins.
11. Smith, J. E. et al. Translation of small open reading
frames within unannotated RNA transcripts in
Saccharomyces cerevisiae. Cell Rep. 7, 1858–1866
(2014).
12. Galindo, M. I., Pueyo, J. I., Fouix, S., Bishop, S. A.
& Couso, J. P. Peptides encoded by short ORFs
control development and define a new eukaryotic
gene family. PLoS Biol. 5, 1052–1062 (2007).
13. Magny, E. G. et al. Conserved regulation of cardiac
calcium uptake by peptides encoded in small open
reading frames. Science 341, 1116–1120 (2013).
This study finds that smORFs can be conserved
across hundreds of millions of years of evolution
at the levels of peptide structure and function.
14. Andrews, S. J. & Rothnagel, J. A. Emerging evidence
for functional peptides encoded by short open
reading frames. Nat. Rev. Genet. 15, 193–204
(2014).
This work further confirms the existence of
functional smORFs and reviews smORF functions
and current testing techniques.
15. Pueyo, J. I., Magny, E. G. & Couso, J. P. New peptides
under the s(ORF)ace of the genome. Trends Biochem.
Sci. 41, 665–678 (2016).
16. Saghatelian, A. & Couso, J. P. Discovery and
characterization of smORF-encoded bioactive
polypeptides. Nat. Chem. Biol. 11, 909–916 (2015).
17. Hemm, M. R., Paul, B. J., Schneider, T. D., Storz, G.
& Rudd, K. E. Small membrane proteins found by
comparative genomics and ribosome binding site
models. Mol. Microbiol. 70, 1487–1501 (2008).
18. Hanada, K. et al. Small open reading frames
associated with morphogenesis are hidden in plant
genomes. Proc. Natl Acad. Sci. USA 110, 2395–2400
(2013).
19. Ma, J. et al. Discovery of human sORF-encoded
polypeptides (SEPs) in cell lines and tissue.
J. Proteome Res. 13, 1757–1765 (2014).
20. Anderson, D. M. et al. A micropeptide encoded by
a putative long noncoding RNA regulates muscle
performance. Cell 160, 595–606 (2015).
21. Mackowiak, S. D. et al. Extensive identification
and  analysis of conserved small ORFs in animals.
Genome Biol. 16, 179 (2015).
This study uses a stringent computational
approach to identify hundreds of conserved
smORFs in lncRNAs and UTRs in several
model animals, and re‑evaluates previous
computational studies.
22. Lemaitre, B., Reichhart, J. M. & Hoffmann, J. A.
Drosophila host defense: differential induction of
antimicrobial peptide genes after infection by various
classes of microorganisms. Proc. Natl Acad. Sci. USA
94, 14614–14619 (1997).
23. Pueyo, J. I. & Couso, J. P. The 11‑aminoacid
long Tarsal-less peptides trigger a cell signal in
Drosophila leg development. Dev. Biol. 324,
192–201 (2008).
24. Djakovic, S., Dyachok, J., Burke, M., Frank, M. J.
& Smith, L. G. BRICK1/HSPC300 functions with SCAR
and the ARP2/3 complex to regulate epidermal cell
shape in Arabidopsis. Development 133, 1091–1100
(2006).
25. Hanyu-Nakamura, K., Sonobe-Nojima, H.,
Tanigawa, A., Lasko, P. & Nakamura, A. Drosophila
Pgc protein inhibits P‑TEFb recruitment to chromatin
in primordial germ cells. Nature 451, 730–733
(2008).
26. FlyBase Consortium. The FlyBase database of
the Drosophila genome projects and community
literature. Nucleic Acids Res. 27, 85–88 (1999).
27. Ingolia, N. T. et al. Ribosome profiling reveals
pervasive translation outside of annotated protein-
coding genes. Cell Rep. 8, 1365–1379 (2014).
28. Ingolia, N. T., Lareau, L. F. & Weissman, J. S.
Ribosome profiling of mouse embryonic stem cells
reveals the complexity and dynamics of mammalian
proteomes. Cell 147, 789–802 (2011).
This work uses ribosome profiling in mouse
embryonic stem cells to show pervasive translation
from alternative start sites, non-canonical start
codon usage, and uORF and lncRNA translation.
29. Pueyo, J. I. et al. Hemotin, a regulator of phagocytosis
encoded by a small ORF and conserved across
metazoans. PLoS Biol. 14, e1002395 (2016).
30. Johnstone, T. G., Bazzini, A. A. & Giraldez, A. J.
Upstream ORFs are prevalent translational
repressors in vertebrates. EMBO J. 35, 706–723
(2016).
31. Wang, X. Q. & Rothnagel, J. A. 5ʹ‑untranslated regions
with multiple upstream AUG codons can support low-
level translation via leaky scanning and reinitiation.
Nucleic Acids Res. 32, 1382–1391 (2004).
32. Calvo, S. E., Pagliarini, D. J. & Mootha, V. K.
Upstream open reading frames cause widespread
reduction of protein expression and are polymorphic
among humans. Proc. Natl Acad. Sci. USA 106,
7507–7512 (2009).
33. Fritsch, C. et al. Genome-wide search for novel human
uORFs and N‑terminal protein extensions using
ribosomal footprinting. Genome Res. 22, 2208–2218
(2012).
34. Ji, Z., Song, R., Regev, A. & Struhl, K. Many lncRNAs,
5’UTRs, and pseudogenes are translated and some
are likely to express functional proteins. eLife 4,
e08890 (2015).
35. Duncan, C. D. & Mata, J. The translational landscape
of fission-yeast meiosis and sporulation. Nat. Struct.
Mol. Biol. 21, 641–647 (2014).
36. Pegueroles, C. & Gabaldon, T. Secondary structure
impacts patterns of selection in human lncRNAs.
BMC Biol. 14, 60 (2016).
37. Guttman, M., Russell, P., Ingolia, N. T., Weissman, J. S.
& Lander, E. S. Ribosome profiling provides evidence
that large noncoding RNAs do not encode proteins.
Cell 154, 240–251 (2013).
38. Banfai, B. et al. Long noncoding RNAs are rarely
translated in two human cell lines. Genome Res. 22,
1646–1657 (2012).
39. Chew, G. L. et al. Ribosome profiling reveals
resemblance between long non-coding RNAs
and 5ʹ leaders of coding RNAs. Development 140,
2828–2834 (2013).
40. Nelson, B. R. et al. A peptide encoded by a transcript
annotated as long noncoding RNA enhances SERCA
activity in muscle. Science 351, 271–275 (2016).
41. Yates, A. et al. Ensembl 2016. Nucleic Acids Res. 44,
D710–D716 (2016).
42. Rubin, G. M. et al. Comparative genomics of the
eukaryotes. Science 287, 2204–2215 (2000).
43. Li, Z. et al. Detection of intergenic non-coding RNAs
expressed in the main developmental stages in
Drosophila melanogaster. Nucleic Acids Res. 37,
4308–4314 (2009).
44. van Heesch, S. et al. Extensive localization of long
noncoding RNAs to the cytosol and mono- and
polyribosomal complexes. Genome Biol. 15, R6
(2014).

588 | SEPTEMBER 2017 | VOLUME 18 www.nature.com/nrm

©
2
0
1
7
M
a
c
m
i
l
l
a
n
P
u
b
l
i
s
h
e
r
s
L
i
m
i
t
e
d
,
p
a
r
t
o
f
S
p
r
i
n
g
e
r
N
a
t
u
r
e
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.
 A N A LY S I S

45. Wang, H., Wang, Y., Xie, S., Liu, Y. & Xie, Z.
Global and cell-type specific properties of lincRNAs
with ribosome occupancy. Nucleic Acids Res. 45,
2786–2796 (2017).
46. Slavoff, S. A. et al. Peptidomic discovery of short open
reading frame-encoded peptides in human cells.
Nat. Chem. Biol. 9, 59–64 (2013).
This work applies a novel proteomics approach
to discover SEPs in human cells.
47. Miklos, G. L. G. & Rubin, G. M. The role of the
genome project in determining gene function:
insights from model organisms. Cell 86, 521–529
(1996).
48. Mumtaz, M. A. & Couso, J. P. Ribosomal profiling
adds new coding sequences to the proteome.
Biochem. Soc. Trans. 43, 1271–1276 (2015).
49. Crappe, J. et al. Combining in silico prediction and
ribosome profiling in a genome-wide search for novel
putatively coding sORFs. BMC Genomics 14, 648
(2013).
50. Fatica, A. & Bozzoni, I. Long non-coding RNAs:
new players in cell differentiation and development.
Nat. Rev. Genet. 15, 7–21 (2013).
51. Kronja, I. et al. Widespread changes in the
posttranscriptional landscape at the Drosophila
oocyte-to‑embryo transition. Cell Rep. 7, 1495–1508
(2014).
52. Zanet, J. et al. Pri sORF peptides induce selective
proteasome-mediated protein processing. Science
349, 1356–1358 (2015).
53. Kessler, M. M. et al. Systematic discovery of new
genes in the Saccharomyces cerevisiae genome.
Genome Res. 13, 264–271 (2003).
54. Baggerman, G., Cerstiaens, A., De Loof, A.
& Schoofs, L. Peptidomics of the larval Drosophila
melanogaster central nervous system. J. Biol. Chem.
277, 40368–40374 (2002).
55. Loose, C. R., Langer, R. S. & Stephanopoulos, G. N.
Optimization of protein fusion partner length for
maximizing in vitro translation of peptides. Biotechnol.
Prog. 23, 444–451 (2007).
56. Lauressergues, D. et al. Primary transcripts of
microRNAs encode regulatory peptides. Nature 520,
90–93 (2015).
57. Paharkova, V., Alvarez, G., Nakamura, H., Cohen, P.
& Lee, K. W. Rat Humanin is encoded and translated
in mitochondria and is localized to the mitochondrial
compartment where it regulates ROS production.
Mol. Cell. Endocrinol. 413, 96–100 (2015).
58. Lee, C. et al. The mitochondrial-derived peptide
MOTS‑c promotes metabolic homeostasis and reduces
obesity and insulin resistance. Cell Metab. 21,
443–454 (2015).
59. Kozak, M. Regulation of translation via mRNA
structure in prokaryotes and eukaryotes. Gene 361,
13–37 (2005).
60. Szamecz, B. et al. eIF3a cooperates with sequences
5ʹ of uORF1 to promote resumption of scanning by
post-termination ribosomes for reinitiation on GCN4
mRNA. Genes Dev. 22, 2414–2425 (2008).
61. Ebina, I. et al. Identification of novel Arabidopsis
thaliana upstream open reading frames that control
expression of the main coding sequences in a peptide
sequence-dependent manner. Nucleic Acids Res. 43,
1562–1576 (2015).
62. Combier, J. P., de Billy, F., Gamas, P., Niebel, A.
& Rivas, S. Trans-regulation of the expression
of the transcription factor MtHAP2‑1 by a uORF
controls root nodule development. Genes Dev. 22,
1549–1559 (2008).
63. Zhang, Z. & Dietrich, F. Identification and
characterization of upstream open reading frames
(uORF) in the 5ʹ untranslated regions (UTR) of genes
in Saccharomyces cerevisiae. Curr. Genet. 48, 77–87
(2005).
64. Abrusan, G. Integration of new genes into cellular
networks, and their structural maturation. Genetics
195, 1407–1417 (2013).
65. Kelley, L. A. & Sternberg, M. J. Partial protein
domains: evolutionary insights and bioinformatics
challenges. Genome Biol. 16, 100 (2015).
66. Joliot, A. & Prochiantz, A. Transduction peptides: from
technology to physiology. Nat. Cell Biol. 6, 189–196
(2004).
67. Schulman, B. A. & Harper, J. W. Ubiquitin-like protein
activation by E1 enzymes: the apex for downstream
signalling pathways. Nat. Rev. Mol. Cell Biol. 10,
319–331 (2009).
68. Alonso, J. & Santaren, J. F. Characterization of
the Drosophila melanogaster ribosomal proteome.
J. Proteome Res. 5, 2025–2032 (2006).
69. Ghiglione, C., Perrimon, N. & Perkins, L. A.
Quantitative variations in the level of MAPK activity
control patterning of the embryonic termini in
Drosophila. J. Dev. Biol. 205, 181–193 (1999).
70. Vaux, D. L. & Korsmeyer, S. J. Cell death in
development. Cell 96, 245–254 (1999).
71. Itoh, K., Nakamura, K., Iijima, M. & Sesaki, H.
Mitochondrial dynamics in neurodegeneration.
Trends Cell Biol. 23, 64–71 (2012).
72. Staudt, A. C. & Wenkel, S. Regulation of protein
function by ‘microProteins’. EMBO Rep. 12, 35–42
(2011).
73. Seo, P. J., Hong, S. Y., Kim, S. G. & Park, C. M.
Competitive inhibition of transcription factors by small
interfering peptides. Trends Plant Sci. 16, 541–549
(2011).
74. Graeff, M. et al. Microprotein-mediated recruitment
of CONSTANS into a TOPLESS trimeric complex
represses flowering in Arabidopsis. PLoS Genet. 12,
e1005959 (2016).
75. Ling, F., Kang, B. & Sun, X.‑H. Id proteins: small
molecules, mighty regulators. Curr. Top. Dev. Biol.
110, 189–216 (2014).
76. Au, Y. The muscle ultrastructure: a structural
perspective of the sarcomere. Cell. Mol. Life Sci. 61,
3016–3033 (2004).
77. Gawlin´ski, P. et al. The Drosophila mitotic inhibitor
Frühstart specifically binds to the hydrophobic patch
of cyclins. EMBO Rep. 8, 490–496 (2007).
78. Guo, B. et al. Humanin peptide suppresses apoptosis
by interfering with Bax activation. Nature 423,
456–461 (2003).
79. Weinmaster, G. & Fischer, J. A. Notch ligand
ubiquitylation: what is it good for? Dev. Cell 21,
134–144 (2011).
80. Ching, K. H., Kisailus, A. E. & Burbelo, P. D.
Biochemical characterization of distinct regions
of SPEC molecules and their role in phagocytosis.
Exp. Cell Res. 313, 10–21 (2007).
81. Zasloff, M. Antimicrobial peptides of multicellular
organisms. Nature 415, 389–395 (2002).
82. Brogden, K. A. Antimicrobial peptides: pore formers
or metabolic inhibitors in bacteria? Nat. Rev.
Microbiol. 3, 238–250 (2005).
83. D’Onofrio, G., Mouchiroud, D., Aissani, B.,
Gautier, C. & Bernardi, G. Correlations between
the compositional properties of human genes, codon
usage, and amino acid composition of proteins. J. Mol.
Evol. 32, 504–510 (1991).
This study correlates the overall nucleotide and
amino acid compositions of protein-coding
sequences, highlighting and attempting to explain
the biased nonrandom amino acid usage of
canonical proteins.
84. Hansen, M., Kilk, K. & Langel, U. Predicting
cell‑penetrating peptides. Adv. Drug Deliv. Rev. 60,
572–579 (2008).
85. Jones, S. W. et al. Characterisation of cell-penetrating
peptide-mediated peptide delivery. Br. J. Pharmacol.
145, 1093–1102 (2005).
86. Murphy, M. P. Targeting lipophilic cations to
mitochondria. Biochim. Biophys. Acta 1777,
1028–1031 (2008).
87. Hoffmann, J. A., Kafatos, F. C., Janeway, C. A.
& Ezekowitz, R. A. Phylogenetic perspectives in innate
immunity. Science 284, 1313–1318 (1999).
88. Fan, L. et al. DRAMP: a comprehensive data
repository of antimicrobial peptides. Sci. Rep.
6, 24482 (2016).
89. Wenzel, M. et al. Small cationic antimicrobial peptides
delocalize peripheral membrane proteins. Proc. Natl
Acad. Sci. USA 111, E1409–E1418 (2014).
90. Slavoff, S. A., Heo, J., Budnik, B. A., Hanakahi, L. A.
& Saghatelian, A. A human short open reading frame
(sORF)-encoded polypeptide that stimulates DNA end
joining. J. Biol. Chem. 289, 10950–10957 (2014).
91. Pueyo, J. I. & Couso, J. P. Tarsal-less peptides control
Notch signalling through the Shavenbaby transcription
factor. Dev. Biol. 355, 183–193 (2011).
92. Palm, C., Jayamanne, M., Kjellander, M.
& Hallbrink, M. Peptide degradation is a critical
determinant for cell-penetrating peptide uptake.
Biochim. Biophys. Acta 1768, 1769–1776 (2007).
93. Jaillon, O. et al. Translational control of intron splicing
in eukaryotes. Nature 451, 359–362 (2008).
94. Pimplikar, S. W. Reassessing the amyloid cascade
hypothesis of Alzheimer’s disease. Int. J. Biochem.
Cell Biol. 41, 1261–1268 (2009).
95. Waxman, D. & Peck, J. R. Pleiotropy and the
preservation of perfection. Science 279, 1210–1213
(1998).
96. Billingsley, M. L. et al. Functional and structural
properties of stannin: roles in cellular growth,
selective toxicity, and mitochondrial responses
to injury. J. Cell. Biochem. 98, 243–250 (2006).
97. Chng, S. C., Ho, L., Tian, J. & Reversade, B. ELABELA:
a hormone essential for heart development signals via
the apelin receptor. Dev. Cell 27, 672–680 (2013).
98. Pauli, A. et al. Toddler: an embryonic signal that
promotes cell movement via Apelin receptors. Science
343, 1248636 (2014).
References 97 and 98 characterize the
32‑amino-acid-long SEP toddler, which acts
as a hormone in the zebrafish heart.
99. Dunn, J. G., Foo, C. K., Belletier, N. G., Gavis, E. R.
& Weissman, J. S. Ribosome profiling reveals
pervasive and regulated stop codon readthrough
in Drosophila melanogaster. eLife 2, e01179 (2013).
100. Carvunis, A. R. et al. Proto-genes and de novo gene
birth. Nature 487, 370–374 (2012).
This study proposes a model for the de novo
emergence of protein-coding genes from
proto-genes or sequences, forming a continuum
between noncoding DNA and fully coding genes.
101. McLysaght, A. & Guerzoni, D. New genes from non-
coding sequence: the role of de novo protein-coding
genes in eukaryotic evolutionary innovation.
Phil. Trans. R. Soc. B http://dx.doi.org/10.1098/
rstb.2014.0332 (2015).
102. Zhao, L., Saelao, P., Jones, C. D. & Begun, D. J.
Origin and Spread of de novo genes in Drosophila
melanogaster populations. Science 343, 769–772
(2014).
103. Zhou, Q. et al. On the origin of new genes in
Drosophila. Genome Res. 18, 1446–1455 (2008).
104. Neme, R. & Tautz, D. Phylogenetic patterns of
emergence of new genes support a model of frequent
de novo evolution. BMC Genomics 14, 117 (2013).
105. Reinhardt, J. A. et al. De novo ORFs in Drosophila are
important to organismal fitness and evolved rapidly
from previously non-coding sequences. PLoS Genet. 9,
e1003860 (2013).
106. Moyers, B. A. & Zhang, J. Evaluating phylostratigraphic
evidence for widespread de novo gene birth in genome
evolution. Mol. Biol. Evol. 33, 1245–1256 (2016).
107. Xie, C. et al. Hominoid-specific de novo protein-coding
genes originating from long non-coding RNAs.
PLoS Genet. 8, e1002942 (2012).
108. Schlotterer, C. Genes from scratch — the evolutionary
fate of de novo genes. Trends Genet. 31, 215–219
(2015).
109. Sommer, R. J. The future of evo-devo: model systems
and evolutionary theory. Nat. Rev. Genet. 10,
416–422 (2009).
110. Yang, S. & Bourne, P. E. The evolutionary history
of protein domains viewed by species phylogeny.
PLoS ONE 4, e8378 (2009).
111. Milligan, M. J. et al. Global intersection of long
non‑coding RNAs with processed and unprocessed
pseudogenes in the human genome. Front. Genet. 7,
26 (2016).
Acknowledgements
The authors thank their colleagues J. Pueyo, E. Magny,
S. Bishop and F. Casares for helpful suggestions about
the manuscript. This work was funded by grants from the
Spanish Ministerio de Economía, Industria y Competit­ividad
(MINECO; ref. BFU/2016-77793-P) and the British Bio­
technology and Biological Sciences Research Council (BBSRC;
ref. BB/N001753/1) to J.-P.C.
Competing interests statement
The authors declare no competing interests.
Publisher’s note
Springer Nature remains neutral with regard to jurisdictional
claims in published maps and institutional affiliations.

NATURE REVIEWS | MOLECULAR CELL BIOLOGY VOLUME 18 | SEPTEMBER 2017 | 589

©
2
0
1
7
M
a
c
m
i
l
l
a
n
P
u
b
l
i
s
h
e
r
s
L
i
m
i
t
e
d
,
p
a
r
t
o
f
S
p
r
i
n
g
e
r
N
a
t
u
r
e
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.
 CORRECTION

CORRECTION

Classification and function of small open reading frames

Juan-Pablo Couso & Pedro Patraquim
Nature Reviews Molecular Cell Biology doi:10.1038/nrm.2017.58 (2017)
The original online version of this article contained four errors, which have now been corrected. The corrections included
two typos in the main text, the addition of a missing point in the X axis in Figure 3b, and the exchange of the position of two
column headers in Figure 5c.

NATURE REVIEWS | MOLECULAR CELL BIOLOGY www.nature.com/nrm

©
2
0
1
7
M
a
c
m
i
l
l
a
n
P
u
b
l
i
s
h
e
r
s
L
i
m
i
t
e
d
,
p
a
r
t
o
f
S
p
r
i
n
g
e
r
N
a
t
u
r
e
.
A
l
l
r
i
g
h
t
s
r
e
s
e
r
v
e
d
.

View publication stats