The sheer power of today’s computers is allowing umpteen years of NMR milk records to be used as predictors of not only pregnancy but disease as well, including TB. Peter Hollinshead talks to NMR managing director Andy Warne about this breakthrough.
Now you are part of the HappyMoo project, which looks like it could revolutionise farming by being able to predict the characteristics of a particular cow right down to identifying disease from its milk alone, how big a breakthrough do you personally see this as being? Is it a once-in-a-lifetime quantum leap?
In simple terms, yes it is. Although we have started to use milk recording samples for other things, such as testing for Johne’s disease, the actual way fat, protein and somatic cells are measured in milk recording has not really changed in the last 10 years.
However, the use of spectral data is the next development which allows us to unlock additional data from milk samples without doing any extra testing on milk.
Before we get onto the nitty gritty of the mechanics of it all, can you give me some idea of its scope, particularly as regards which diseases we may be able to predict?
This will depend on what data we have recorded, but if we have a trait already recorded, we will be able to put that against the current milk sample and predict those characteristics.
Can you give me one or two of those diseases which may be in your sights at the moment?
It’s not only diseases, but it is fertility, energy, TB or anything we have the data for.
Without the TB bit, you would already be looking at energy balances and the like. So what new diseases are you looking at?
Well, lameness and mastitis are two as we already have 20,000 lameness records to work with.
Just as a measure of its impact though, could it predict ketotic cows, for example, whereby you already have determinants of negative energy balance and mobilisation of body fats which are a precursor to ketosis?
Yes, we have the energy balance data of the cow in early lactation, so we can use that to reflect that sort of trait. However, ketosis is not part of the HappyMoo project, although we will be looking at that.
As I understand it, what you are doing is correlating the physical attributes or health status of a group of cows at a particular time with the milk profiles of those animals as determined by Mid Infrared Spectroscopy. Once you have millions of these, you can take any cow’s milk in the future and study its profile and predict what physiological condition or health state that cow is in. Or have I missed the point somewhere?
Well, a large part of this project is actually using the data we hold historically. We have been storing spectral data for the last eight years, so we have a very large phenotypic database.
Any data we have collected historically can be compared to the spectra of those milk samples and that allows us to use the model for future predictions.
Coming back to that ketosis element; it may be something you have not recorded directly in the past and therefore possibly is not in your eight years of records, but how could you get that data to know whether a cow was going to be ketotic from its current milk profile?
We would have some records of ketosis recorded on farm software and we could go back and dig out those records and relate them back to the spectra going back in history.
Alternatively, we could for six months work with say 100 farms and produce some good ketosis records, then we would be able to train the model to look for ketosis.
This is new technology and something we have developed with a number of milk recording bodies around Europe and it is a really exciting tool.
To a certain extent we are asking the industry what it would like us to find.
Perhaps one of the biggest pressures now is greenhouse gas emission and we could use this tool to predict which animals produce high yields of milk but low levels of methane.
So effectively, you are saying within a year we could determine which cows are high greenhouse gas emitters if we had the technology on some reference farms to measure this?
Yes, but as a commercial organisation working on behalf of our customers, we are more interested in developing tools which give our farmers a reward for what they produce today. At the moment, they do not get one for low methane cows, although that may come in time.
So right now we are looking at diseases, efficiency in terms of feed and fertility, with predictions of whether the animal is pregnant or not.
So you and other milk recorders are in a particularly powerful position with this sort of technology, aren’t you?
Yes, NMR is taking the traditional milk recording process and adding additional value from the same milk samples. We believe we are leapfrogging the idea of on-farm testing, for example where increasingly complex pieces of lab equipment are being used.
Agriculture is actually slow to catch on to the power of machine learning, but this is the technology used for driverless cars and predictive text on our phones.
We believe, working with Scotland’s Rural College (SRUC), we are actually leading the way in agriculture, not just in the UK or Europe, but on a global scale.
We don’t see any other milk recording organisation anywhere in the world as far ahead with this as we are.
Moving on to practicalities if we may. To get the magnitude of the database you would require, you are working on this HappyMoo project with France, Germany, Belgium and Ireland, where you have the same comparable readings. How far has that got?
Yes, we standardise our milk testing machines in all the labs in the European milk recording group and that allows us to compare our milk spectra with that from the other countries.
So are you confident that the Germans, for example, won’t use this to jump ahead of you?
Well, they might if we don’t act quickly to stay ahead. But we believe through our work with the SRUC, particularly with TB, we have added value to our own milk spectra in a way some of the other milk recording organisations within that group have not been able to do.
So with our very large phenotypic database and the research and machine learning capabilities of the SRUC, we think that is a partnership which will help the UK step ahead of other members of the HappyMoo project.
I believe you even collect hair samples to measure cortisol as a means of assessing how stressed that cow is?
This is all part of the project and we are trying to identify whether an animal is in a stressful state of not, and one of those indicators is cortisol in hair.
We will be taking these hair samples and analysing them and comparing them to the milk spectra to see whether we can see that signature in the milk.
I am not quite clear what sort of stress it is telling you about. Is it an ephemeral stress through movement around the yard or nutritional stress, for example?
We are monitoring long-term stress and not just a movement with a group or collecting yard. The benefit of this approach in fact is you can relate any phenotypical characteristic to the milk spectra.
I suppose one weakness is cow status or condition is mostly subjective, which will be dependent on the practitioner’s assessment at the time, would it not?
We are taking a trait we recognise and feeding it through the modelling processes. If there are animals which have not been recorded as accurately as we would like, they should be few and far between and should not really affect the modelling process because of the large quantities of data we are using.
I believe the SRUC is looking at identifying TB in cattle more accurately than any present day tests can. This is a big claim in itself, but it would have been able to draw on a confirmed TB status from APHA records, then relate it back to milk records of cows prior to being culled to develop an accurate algorithm. But the fact it would know which particular cows had TB gives it a firmer basis on which to work, don’t you think?
Yes, absolutely. The research the SRUC is doing with TB is a project NMR is sponsoring, and the work it recently reported is a project we are very much involved with. We are using SRUC’s expertise in machine learning and combining that with our very large dataset.
Are you predicting you can detect Johne’s in a cow from first lactation or even earlier?
We don’t know, but it is something we are looking at. Fortunately, we have that spectral data stored and we have the phenotypic records of those Johne’s animals, and it is more than possible Johne’s animals have a signature earlier in their life than you would pick up with an ELISA test, where you are looking at antibodies produced in reaction to an infection.
Moving on to breed and diet. Would your results in any way be influenced by breed, as some breeds presumably produce different constituents and levels of fatty acids, etc., and could results be influenced by feeding of say protected fat in their diet?
We would expect to see a difference in the spectra of the milk, but because we feed all this information into the modelling process, we would still be able to identify the trait we were looking for.
Would you need more regular information to predict disease in a cow?
I think with mastitis and indeed lameness management in the herd, you are trying to build up individual cow records to see whether an animal has a tendency to have mastitis or be lame.
It is not necessarily a question of whether this cow is high or low in a particular month and in what way I am going to treat the animal, but it is the picture of the animal and whether it is a chronic mastitis animal and whether it has had incidences before.
It is very much about building up a history of the cow, which may determine how a producer would view that animal.
The HappyMoo project is financed by European Development Funding. Will this cease as we complete our Brexit transition by the end of this year, and what is NMR’s investment cash-wise in the project?
It is all protected and we have access to it until 2022 when the project completes, and NMR is in it for €300,000.
We know this will give us a quantum leap over our competitors and, although the energy balance and fatty acid work is already available, we believe the work we are doing in pregnancy will give us a big leap forward as we will be able to predict whether an animal is pregnant simply from milk records.
Going back one stage if I may. When do you think you will have that TB prediction available commercially?
TB is one of the more difficult areas for NMR to approach as it is a notifiable disease, so you have to tread very carefully.
We are working with Defra to see how we might deploy this technology in the future.
Are you thinking this may be the ultimate test for TB and possibly replace the skin test?
Clearly, the animal needs to be milking, so it could not replace the skin test completely, as you couldn’t use it in beef and youngstock.
We believe farmers want to take back control of the way TB testing is done, as at the moment the only TB being done is on behalf of Defra and that leaves farmers with few tools to allow them to really help themselves.
Is the HappyMoo branding from the European group or was it dreamed up in the UK? Like with eggs, should we expect to see ‘HappyMilk’ therefore ‘HappyCows’ branding in future? Do you think that will be used by retailers to overtly display a reassurance to help satisfy customers’ demands?
The name came from the European group, but it could be a name for buyers to help sell their products. We already work very closely with major retailers on provenance schemes and, in working with both our farmers and retailers, we have to ensure there is a balance of data between those two groups.
The key thing here is this is a non-invasive technique which requires no additional data collection but will give a quantitative measure of aspects of animal health and well-being.
When this new service comes into being with a wider remit possibly than it has now, will it cost producers more?
It is impossible to predict what the traits will be and what each will cost, but we see this relatively holistically in terms of adding value to the milk recording service overall, and therefore we don’t see a wholesale change in the price of milk recording.
The particular benefit to NMR will be to add value to our service and improve our market share. So we would see this as a tool to encourage people who are not milk recording to start, or people who are currently recording with other organisations to milk record with NMR.
Ultimately, farmers will have choices, but I think other recording organisations will struggle to keep up with this additional value we are adding to our service.
Finally, in the wider picture you are managing director and must be doing predictions looking ahead. Where do you think this approach will take us in the next five to 10 years?
The pressure on farmers is increasing all the time with retailers seeking additional provenance data, and additionally with growing pressure from the environmental lobby.
At NMR, we believe we have a big role to play to enable farmers to produce good quality milk in an efficient way, and in five years we expect to see all producers in the UK milk recording. Hopefully a big proportion of them with NMR.