While I of course never have stupid mistakes...ahem...I have many "friends" who:
but I'm sure there are some other very common pitfalls that are unique to bioinformatics programming. What are your favorites?
Invent a new weakly defined, internally redundant, ambiguous, bulky fruit salad of a data format. Again.
It made me laugh when I read the MIQE guidelines not long after this comic came out:
The nomenclature describing the fractional PCR cycle used for quantification is inconsistent, with threshold cycle (Ct), crossing point (Cp),and take-off point (TOP) currently used in the literature ... we propose the use of quantification cycle (Cq)
I truncated many fasta files this way when trying to see which headers it contained:
grep > some.fasta
I also see a lot of off-by-one errors due to switching between formats
and switching between languages:
Not only is the bed format 0-based, it's also "half-open", meaning the start position is inclusive, but the end position is not.
So if your region starts at position 100 and ends at 101 using standard 1-based coordinates with both start and end inclusive (ie it's two bases long), when you convert it to 0-based half-open coords for bed format the region now starts at 99 but it still ends at 101!
Gene annotation stored in an excel file and find out that some HUGO gene names have been hacked by Excel. SEPT9 become sept-9. Conclusion Do not use the .xls format to store your data.
Listen people saying this eternal mistake "Hey these two sequences are 50% homologs"
This is a popular one, dec1 is another well known example. But you can actually tell Excel not to do that auto correction. Since you most often get the data from biologist who may have treated the data in Excel already better use an another ID column and not the gene name column if that is available (you often receive both anyway). These errors can even occur in databases that you download data from or which are used for annotation, so it is good to check.
Reminds me of that old joke...
There are only two hard things in computer science: naming things, cache invalidation and off-by-one errors
I feel like a lot of "stupid mistakes" revolve around betrayed trust and false assumptions
For example:
If you forgive an attempt to be somewhat provocative, my two favorite mistakes are:
Letting academics build software
Academics are in the need to publish papers, and one easy way to do that is to implement an algorithm, demonstrate that it works (more or less), and type it up in a manuscript. BT,DT. But robust and useful software requires a bit more than that, as evidenced by the sad state of affairs in typical bioinformatics software (I think I've managed to crash every de novo assembler I've tried, for instance. Not to mention countless hours spent trying - often in vain - to get software to compile and run). Unfortunately, you don't get a lot of academic credit for improved installation proceedures, testing, software manuals, or especially, debugging of complicated errors. Much better and productive to move on to the next publishable implementation.
Letting academics build infrastructure
Same argument as above, really. Academics are eager to apply to construct research infrastructures, but of course they aren't all that interested in doing old and boring stuff. So although today's needs might be satisfied by a $300 FTP server, they will usually start conjecturing about tomorrow's needs instead, and embark on ambitious, blue sky stuff that might result in papers, but not in actually useful tools. And even if you get a useful database or web application up and running (and published), there is little incentive to update or improve it, and it is usually left to bitrot, while the authors go off in search of the next publication.
I am fine with point 2, but I have to disagree with 1. Your de novo assembler example is actually not a good one. De novo assembly is very complicated and highly data dependent. I doubt any assemblers work for any data sets, no matter developed by academia or by professional programmers.
@Jeremy: I'm not so sure industry is much better, and it's possible that academia is the democracy of development - worst, except for the others. Also, a lot of industry software are add-ons, designed to sell something else. FWIW, Newbler seems to be one of the better assemblers out there, and CLC is at least half-decent as an analysis platform for non-informaticians.
What about citations of your paper describing the software? The more people are able to use your software, the more cited it will be. That sounds like an incentive to me, and I try to see it that way when I write code (but I don't publish much code, it's mostly quite ad-hoc stuff).
Ketil repeatedly says in the post that academics don't have incentive to improve their tools, as it's more beneficial to move on to other publications. I say the incentive are citations of your method. Not to mention I regularly see academic software where they care about the things he says academics don't care about. So why would somebody say there are no incentives for improvements, when I can clearly see them and point them out to you?
You do have a point. However, citations are quite rare. One common problem with bioinformatics tools is that people use but seldom cite them properly. Imagine something like bioawk - it would be day to day usage for quite a few people, but it's not seen as something cite-worthy. Good tools have this habit of melding into the background that they don't stand out. Journals are changing this practice now though, with initiatives like STAR methods.
One mistake not unique to bioinformatics is: while editing one source file, compile and run another file.
Using grep to find sequence (or other) IDs without using the -w switch: grep 'seq12' will also find seq121, seq122 and so on.
IMHO being off by one is the emperor of all bioinformatics mistakes - it rules them all - and probably causes tens of millions of dollars in wasted effort
trying to solve any problem with BioPerl :-)
but the
grep > some.fastaare my favorite mistakes.
Re-inventing the wheel. So often did I have to debug (or just replace) a bad implementation of a fasta-parser when BioPython/BioPerl have perfect implementations, I don't understand why no-one bothers to use them. 10 minutes in Google can save you 2 days of work and other people a week of work (you save 2 days of programming, they save a week of understanding your program to find the bug)
As the nim library docs say, "The great thing about re-inventing the wheel is that you can get a round one."
My main reason for reinventing the wheel is that I want to use much more powerful and general language: Python instead of R. Of course, if the stuff I needed was already in Python/Pandas it would be a different thing entirely.
I fully agree, re-inventing the wheel is so tempting. We are way too eager to write a few lines of code each time. Plus, because you may have convinced yourself that you can resolve the code in 15 minutes, you don't bother about writing any documentation. In short, there is a very large tendency to re-invent the wheel... many, many times!
Using other peoples code is all fun and games, until you realise that your package has 106 dependencies, and you only use 1 function from each. Each of those dependencies has its own dependencies, depends on a particular version of gcc (but not the same one as the others), doesn't play nice with some common system used on other peoples systems...
Try to open microarray or, worse, NGS datafiles with excel or word...
Forget to do 'dos2unix' and then spend a lot of time trying to figure out why there is no OUTPUT
Well I have couple:
-o something.out option. Then switching off the monitor and coming the next day to see a bunch of characters in my terminaltar -zxvf without checking the tar file before, I have decompressed thousands of files in my current directory assuming they came in their own folder.
I gave my Amazon EC2 password to someone in my group who wanted to run something quickly (estimated cost, $2). I received the bill 2 months later: $156. This person forgot to close the instance. This is a 8 months story and I'm still waiting for my reimbursement... Conclusion: don't trust colleagues!
I'll offer this one, which is a bit on the general side: Deletion of data that appear to serve no relevance from the computational side, but which have importance to the biology/biologist. Often, this arises from a lack of clear communication between the two individuals/teams as to what everything means, what it exactly means and why it is relevant to the process being developed.
Not keeping an adequate notebook.
I often encounter problems related to the fact the computer scientists index their arrays starting with 0, while biologists index their sequences starting with 1. Simple concept that drives the noobs mad and even trips up more experienced scientists every once in a while.
Do pathways statistics or gene set enrichment statistics and then represent the list of gene sets as a valuable result, instead using that statistics just as a means to decide which pathways need to be evaluated.
(This is bad for many reasons for instance because the statistical contribution of a key regulatory gene in a pathway is equal to that of 1 out 7 iso-enzymes that catalyze a non-relevant side reaction, and because the significance of a pathway changes when you add a few non-relevant genes, and also because we have many overlapping pathways).
Another typical mistake is to solve problems that nobody has.
No, I think it is actually wrong to publish a list of pathways without further judgement. I think not doing the judgement is a mistake. But I have to admit that I don't really understand your examples. So maybe my English is not good enough to understand the finesses of the difference between poor judgement and stupid mistakes.
One mistake: not looking to see that the 0x4 bit in the bitflag column of a SAM (or BAM) file indicates the entry is mapped. RNAME, CIGAR, and POS may be set to something non-null (an actual string!) but these are not meaningful if the 0x4 flag says the read is unmapped.
using rm -rf * .fasta in unintended directory; especially if within the home directory...
I too have had that moment of dread when I realized I typed rm * /folder versus rm /folder/*! Check out some of the solutions on this forum page, specifically trash-cli. You can set up a trash folder so after deleting files they are not completely gone and can be restored if needed. You would have to manually empty the trash folder or set up a cron job to do so on a regular basis, but this may help circumvent the nightmares listed here!
I was just deleting some unnecessary files from a dir and managed to have a space and an asterisk at the end of the rm command. As soon as I realized what was happening I hit ctrl-c, but important files without backups were already gone. Oh well, it will only take like 2-3 weeks to reproduce them. Also time to edit .bashrc following Philipp's post..
Masking out sequence in a FASTA file (e.g. s/TAAT/NNNN/ig) where the sequence is formatted, i.e. split onto multiple lines.
This will miss TAAT that is split over the end of one line and the start of the next!
The classic mistake (also mentioned above by Casey) is not being aware the genome assembly effect coordinates.
What kills me the most is the hand editing of data sets.
If you're reading this and do it, please stop -- and start using automated builds -- with clear documentation.
Having separate files for each sequence.
Of a 454 run.
I wouldn't say it's stupid, but I think a very common mistake is to not correct for batch effects in high-throughput data.
Batch effects can (best-case) hide the real effect that you're looking for, or (worst-case) make it look like your variable of interest is contributing to your findings when it's actually an artifact.
Leek + Irizarry et al. have a sobering review on this here.
Generate Multiple Sequence Alignment direct from fasta or other file and ask: why there are no gaps & deletion in the MSA viewer?
developing algorithms and software around one single piece of low quality data with no prior knowledge while being ignorant about the entire problem.
Easy one: you wait for 4 hours downloading a big DNA file (e.g. bam file) and you mistakenly delete it when trying to move it with a good old rm.
Creating a new set of (public) identifiers for your database for entities that already have identifiers in a widely-user public db.
Having unstable identifiers that are publicly available.
Our paper, out this week, is a decent round up of how to avoid identifier-specific issues with your data and that of others. http://dx.doi.org/10.1371/journal.pbio.2001414
tacking on another command line argument without looking through the rest of them
novoalign -a ATCTCGTATGCCGTCTTCTGCTTG -d genome.ndx -F ILMFQ -f query.fq -a -m -l 17 -h 60 -t 65 -o sam -o FullNW
the first adapter argument (-a ATCTCGTATGCCGTCTTCTGCTTG) is negated by the empty second one
Running the bwa/GATK pipeline with a corrupt/incompletely generated bwa index of hg19. Everything still aligned, but one of 2 mates would have its strand set incorrectly. Other than the insert size distribution, everything seemed normal, until the TableRecalibration step downshifted all quality scores significantly and then UnifiedGenotyper called 0 SNPs. 1st time I've seen a problem with step 1 of a pipeline not become obvious until step 5+.
Scripting an hour to do something you could have done in half an hour manually, and then never needing to repeat it again.
I've just made one, which cost me a good headache trying to figure out the biology underlying my strange results!
POS field to be the leftmost position of my mapped read on the '+' strand, and the rightmost position on the '-' strandNote to self : "Read the manual..."
that only applies if the sequence contains only matches or mismatches, this means edit strings that are composed of a number followed by M (like 76M) . For all other alignments you will need to parse the CIGAR string and build the end coordinate from the start + numbers in the edit string.
Thanks a lot! I'll keep that in mind! Or maybe you can share it here as a Tool? or as an answer to this thread : Mapping Reads With Bwa And Bowtie ?
I made one a few months ago. I launched a heavy process in a pay-per-use cluster, it was running for one week. I thought, 6 pennies/hr cannot be too much money. I received a bill for $832 usd. I'm not using this cluster again unless I estimate the total cost of the process.
edit: the price is per core
Possibly: implementing methods that magically generate p-values from non-replicated RNA-seq experiments, possibly as a result to pressure from 'experimentalists'. I really would like to know the history behind their implementation (where they forced by reviewers, or by other groups?). Now we have to explain why those p-values are bogus, and why there are so little significantly differentially expressed genes detected in a non-replicated analysis.
When using scp, I did this:
scp -r somename@somehost:$home ~/
instead of $home, should use ~/
this created a file named $home in my other account.
used rm $home
and kabum!
should pay attention and look how remove this kind of file before.
I spent hours to implement R parallel script to fully exploit 64 cores and 250 GB RAM of the server lab. Thus, really proud of myself, I run it on my 4 cores and 8 GB RAM desktop pc. Boom.
I remember once I put a coma in the wrong bracket in some matrix assignment on genotype data and after running the code my laptop froze so hard I had to shut it down by holding the power button. I remember the laptop needed two reboots to get back to normal. After going through the code to see what happened, I realized I assigned a matrix to each cell of itself (or something like that), basically creating an object somewhere around 250 GB in size. My laptop had 8 GB of RAM and about the same of swap (older Ubuntu) and it run out of both within an eye-blink :D
Losing an hour to learn that some files saved on a Mac have a strange 'beginning of file' like character...
Normalize file standards already! x_o
My common one cat aVeryBigFile.whatever &
I had another good one recently. I was executing an untested bash script for generation of output from two input files on our cluster. I just let it run over night. When I checked in the morning, it had generated about 40 TB worth of output (expectation was about 20 MB). There was a tiny spelling mistake that led to an infinite loop. Oops. I was lucky to check it when I did because there was still a few TB space left so at least other jobs didn't get killed because of it..
I am not joking,
how about forgetting to have a full loaded battery for your wireless keyboard and/or mouse near you, and the current one inside is running out of power?
I have spent hours, in repeated occasions, looking for a mysterious error in a perl script that at the end was simply a = instead of a == within an IF statement.
Another recurrent mistake: not documenting what I did and what those scripts do in the belief that everything is so intuitive, organised, simple and natural that it won't be necessary. Then, sometime after, I have to spend hours trying to guess what all that mess was.
rm foo *
instead of
rm foo*
and there goes all the content.
I'm sure someone has mentioned this, but running something like:
bcftools view -S something.txt in.bcf > in.bcf
Goodbye in.bcf, I hardly knew ye.
Also memory management for Java programs on an HPC (beagle I'm looking at you). I love spending hours trying to guess exactly how much memory the JVM is going to use and toggling -Xmx/-Xms flags.
I feel your pain. A handy way to avoid this is to set -o nocolobber, which prevents you from redirecting into existing files. See: https://mywiki.wooledge.org/NoClobber
What about building an interface not aimed at a community of (fellow) Biologists?
s/foo/bar/g
without the g at the end
as I just proved in the field
Making claims without experimental validation. Especially involving studies utilizing multiplexed technologies such as microarrays and high-throughput sequencing.
Just read this article: "How Not To Be A Bioinformatician" Thought it would be interesting to post here....
Some really great comments here, nice to know that such things happen to all genii ;). I have to say my most painful moments relate to my assumption that data obtained elsewhere is correct in every way. I also remember early in my career, using PDB files and realising that sometimes, chains are represented more than once, thus when manually checking calculations involving atomic coordinates, being utterly perplexed and wanting to break my computer. Oh the joys of Bioinformatics.
Using a statistical test on data that does not satisfy the assumptions of that test.
For example, finding differentially-expressed genes by doing ANOVA on log2(FPKM).
Assuming that the gene IDs in "knownGenes.gtf" from UCSC are actually gene IDs. Instead they just put the transcript ID as the gene ID.
This just caused me a bit of pain when doing read counting at the gene level. Basically, any constitutive exon in a gene with multiple splice forms was ignored because all the reads in that exon were treated as ambiguous.
Using the same output (> oh.shit) for multiple commands.
I found myself guilty iterating through the loop and storing data let's say every 100 iterations... but not storing the very last bit of the data (ie lines 10001 to 10026) at the very end.
These are just a few, but thought worth sharing:
How about writing a tool and being convinced it works perfectly, so you start testing it on a complete dataset instead of testing it first on a subset and finding out after it ran for an hour or so that you made a tiny mistake somewhere. Sooo much time wasted that I'll never get back :P
Spending a few months to find an interesting correlation in data and then presenting to the lab that sequenced to find out they changed coverage on the last few samples!
Another fun one is when you develop a pipeline with a small test set thinking speed over all and then you increase your test set size and realize that you're creating TBs of temp data and utilizing hundreds of GBs of RAM :)
Chromosome Identifiers - UCSC "chr1" - Ensembl "1"
Produces very confusing outputs ...
Over estimating future me when I was younger or present me cussing younger me :). Commenting is imp as old me understands that.
Since no one mentioned.
The uppercase/lowercase error.
Sometimes, you need to distinguish between "atgc" and "ATGC". But in most cases, they mean the same thing. So always convert any string you met to uppercase if you do not need to distinguish.
I feel spending my whole Ph.D. debugging this. Hope it helps for others.
Running statistical analyses with no understanding of the tests your using, why you're using them, when it's appropriate to use them, how your data needs to be formated, normalize, or scaled to use them... but hey, you made a volcano plot... so it must be publication quality. right? read the damn paper...
drawingbiological insights into a dataset solely from GO enrichment analysis
using sudo, dd, parted or any combination thereof ... like a moron.
A double mistake combo, 1 - use tar to compress a single file and, 2 - inverting the command arguments
tar cvfz file file.tgz
instead of
tar cvfz file.tgz file
Bye bye file!
It happened to me so many times, that I was considering doing an imagery brain check up.
You also occasionally see IUPAC codes pop up in fasta sequences.
-n in numerical sorting of positions in a bed file.-k flag in sort for a specific column, instead sorting all possible stuff in it.
Mixing between hg18, hg19 and yes new one GrCh 38 too !! :)
A very general one:
Using packages without trying to understand what they actually do.
A more specific one:
Assume that the human mitochondrial reference sequence is the ancestral sequence.
Opening a .clc file in terminal only to find out it was a binary! #facepalm
I have a bit of a different take:
The most pervasive and damaging problem afflicting bioinformatics studies is the failure to curate covariate information, resulting in problems of separation and/or complete separation.
This problem is particularly sinister because this frequently occurs long before the bioinformatician is consulted on the project, meaning the project already has unnecessary limitations before the data scientist/statistician/what-have-you receives the data.
At minimum, this will limit statistical power, in certain cases, it means there is a confound that can't be mitigated without meta-analytics or extreme ends. In my experience this affects, oh I don't know, >70% of labs, including world-class, well-known labs, who don't retain a statistician.
my personal favorite is irreparably breaking your xsession, or breaking a mount point so you can't boot past initramfs
Points 1 and 2 are not really mistakes, I never delete older versions. Instead, I create files with timestamps. Unless you know you'll never need the older version (and still have code/documentation to create the older version if necessary), deleting it is not a smart thing to do. I'd rather delete unzipped files after using them than delete the zip archive.
Points 1-4 are all about space saving. Of these, only #3 is problematic in all scenarios as no information is lost. #4 is probably specific to your way of using GATK.
having someone in your group inadvertently save data (rather than code/readmes) to github. Have fun getting that off!
Use of this site constitutes acceptance of our User Agreement and Privacy Policy.
version 2.3.6
Staying in the office all day
good way to boost reputation.
meta: should this Q be community-wiki?
what's the mean by said generate random genomic sites without avoiding masked (NNN) gaps? more detailed? do not understand.
see this?
http://genome.ucsc.edu/cgi-bin/das/hg19/dna?segment=chr1:1,10000
you wouldn't want to sample from it