How to make an indentification machine

Successful automation often involves machines that carry out tasks differently than persons. For example, a Coulter counter (developed by Wallace H. Coulter, an American engineer), analyzes blood cells by electrical charge, producing a detailed report of red and white cell types faster and more cheaply than does a technician examining a blood smear under a light microscope.  As another case, machine identification of commercial products is enabled by a UPC bar code, which represents a product name in a digital format that can be “read” almost instantaneously by a laser scanner. In a similar way, DNA barcoding “reads” the digital code of DNA, associating that with species names in a reference database, opening the door to fully or partly automated identifications. In 9 September 2010 Nature, scientists from London Natural History Museum, Louisiana State University, and University of Plymouth, UK, propose a different route to automate taxonomic identification, namely, teaching computers to do morphologic pattern recognition. Now that we are on the threshold of “anyone, anywhere, anything” identification with DNA barcoding, this seems a step backward.

I see three major challenges that limit any morphology-based identification system: naming an organism from bits and pieces, recognizing look alikes and life stages, and the diversity of diagnostic features requiring specialized equipment. On the other hand, DNA is the same whether from an intact specimen or an unrecognizable stomach fragment, readily distinguishes look alikes in any life stage, and can be analyzed using the same equipment regardless of specimen. More generally, at the end of the day, little scientific insight will have been gained from a system that distinguishes life forms by the multitidinous particulars of appearance, whereas a library of DNA barcodes linked to named specimens offers a broad view of species-level differences across the diversity of life.

According to MacLeod and colleagues, “a [DNA] bar code isn’t useful until the reference species has been identified by experts”. This makes no sense to me. All large barcode surveys of animals, from ants to fish, have revealed hidden genetic divergences, in many cases leading to recognition of new species.  In fact, DNA barcoding is fast way of screening existing collections for unrecognized species. In this same section, as part of discounting a DNA approach, they state “researchers frequently need to identify non-living objects as well as living ones”. I don’t understand how this is an objection, since, for example, DNA barcodes from ancient bone fragments have been used to define species of extinct flightless Moa (Lambert et al J Heredity 2005).

I know from iPhoto’s remarkable ability to recognize individuals that computers are getting better at pattern recognition. Further development focused on taxonomic specimens may lead to useful tools. However, this seems unlikely to lead to a widely applicable automated system. In a study cited by the authors, phytoplankton identifications by 16 marine ecologists were compared to those with DiCANN, a machine learning system (Culverhouse et al Marine Ecol Prog Series 2003). The authors of that study conclude what is likely to be generally true about morphology based identification:   “In general, neither human nor machine can be expected to give highly accurate or repeatable labeling of specimens”.

5 thoughts on “How to make an indentification machine

  1. I wouldn’t be so radical about morphology-based taxonomies. In the first instance, they precede DNA identification. If it weren’t for them no one would have had the idea of DNA-based taxonomies. In the second instance, morphology lies at the base of whatever identification comes in the future. Third, probably DNA-based taxonomies are additional tools for species identification, but I guess will never make it further from identifying organisms we have never seen in the flesh. I take the stand that we must first SEE the organism, and if not, we might get lost in a barage of information without getting to know the whole organism, not a small part of it. And last, I guess that such radical attitudes have taxonomists like myself out of a job, simply because morphology based taxonomies are dubbed as ‘backward’ or ‘outdated’ and therefore receive no funding because the fad is DNA. I myself adore doing DNA identifications for parasites, but first I collect the little worms, study them thoroughly, and then identify them with DNA techniques.

  2. Hugo, Thanks for your comments. We experience the natural world primarily through sight, so morphology will always be the primary way we recognize species. On the other hand, if the goal is have a machine that can recognize species, then DNA is a better modality. In addition, a standardized DNA approach is likely to provide insights that computerized morphologic keys do not. Best, Mark

  3. Species can go through many profound “life stages”. For example we use DNA barcoding in our business to identify the species of solid wood products and match back to their origin.

    DNA allows us not only to identify species but also the location that the timber was harvested. This is very useful given legality and sustainability issues associated with tropical hardwoods.

    Okay – I recognize that teak furniture or merbau decking are not exactly what one would consider when thinking about these issues, but I offer it here to demonstrate a really important application for barcoding that is having a profound effect on illegal logging and conservation programmes such as REDD.

    I can imagine morphology will remain important, but I clearly see a future for DNA barcoding to generate very cost effective species identification in numerous circumstances.

Leave a Reply