 Yeah, thank you for the introduction and thanks a lot to the organizers to invite me. It was really funny to see my name in between all the famous people, so most of you probably won't know me. I'm Paul Cisca. Before I start my talk, I organized a guided snorkeling tour for tomorrow at two o'clock. It's 120 euros. We can be maximum 12 people. If you are interested in join, please send me an email with your name and telephone number and perhaps you could already now raise your hand if you want to join so that we can look forward it perhaps. Who wants to join? Okay. No, no, no, the entire group and they will provide snorkeling equipment and everything and it's Antonio told me it's a beautiful spot. It's around the corner. So yes, I'm looking forward to this and now my talk. So I'm dealing or I'm interested in how insects deal with these turbulent odor plumes. I think it's relevant also for mice because this is a normal way odorants distribute in the air. So animals, they encounter highly intermittent odorants and they encounter odor filaments or odor stimuli, which are very short and which rise from zero concentration to 100% of concentration within no time. Imagine you are a human walking through such a bloom or a honey bee which flies with up to 25 kilometers per hour. That means that you increase even more temporal intermittency and turbo lenses. I will talk about how insects, what kind of information insects can use to separate odorants from different sources. This behavioral data constrains the coding scheme. So it poses temporal constraints on how odor information can be encoded in the olfactory system and I will show you what are the physiological constraints of an odorant code by showing you the temporal precision of insect or factory receptor neurons. I like to, so I always show in fact this two odor plume picture from Antonio and Massimo's work which nicely shows how odorants from different sources intermingle or not intermingle. So what you see is that the thread and the screen odorants which are released from two nearby sources, they show correlated fluctuations but also uncorrelated fluctuations. They overlap to a certain degree but they also stay separated. So if you want to know whether this plume originates from one or from two sources, you have two different types of information available. One is you just sample the space and you know that here was the red odor and there was the green odor. That would help you to solve this odor segregation problem. Or you use temporal information. Their insects with their short antennae, in particular Osofila which have antennae which are I think 100 micrometer long, they don't have a spatial resolution. So they rely on temporal resolution and this is what I will tell you today. And Aya, please remember the introduction of Andreas Schäfer because he nicely explained this concept behind using temporal odorant structure to solve this object segregation task. And as Andreas, I also read this paper by Hopfeldt and I found it very inspiring. And so in particular the sentence. So I have to read it again. The variable nature of turbulent airflow makes such a remote sensing problem solvable if the animal can make use of the information conveyed by the fluctuations with time of the mixture of odor sources. Good. So if you are an insect with a small antenna with no spatial resolution, you need to have a good temporal resolution in order to do the source segregation. And here also comes the olfactory cocktail party or the cocktail party analogon in play. So when we want to solve this cocktail party problem, so when you start all talking at the same time now to me, I can use different mechanisms to separate your voices. So I can focus my attention to one voice. I hear from which directions the voices come, but there's additional information, and that is if you start talking, then all the sounds you produce, they will arrive in synchrony, they will fluctuate in synchrony, whereas when you are talking, your sounds, they will arrive with a different temporal signature. And in fact, our auditory system uses this information, so it can use odorant sounds and sound onset synchrony to segregate concurrent sounds. So here's one auditory example now. We have 24 sounds, they all start at the same time, you hear this. So you have heard one sound. Now we delay one of the frequencies, 600 thirds by 160 milliseconds. Now you hear this. Did you hear a difference? So you should have heard two sounds. I will play it again. Please listen. Or again. So is this a good analogy? Stimulus wise, I think it's a very good analogy, because you could think of the sounds as being odorants, all these which start synchronously come from one source, and this one odorant here comes from another source. But is it a good analogy for an olfactory segregation mechanism? And now you have to imagine you are an insect. And so how do insects smell? The olfactory organs of insects, they are exposed to the air. So the receptor neurons, they are situated on the antennae in the senzilla. And they are continuously sampling the air. So it's not like sniffing. Insects continuously sample the odorant stimuli, and it is in fact a bit like as our hearing. We also continuously hear there is no intermittency due to sniffing or so. So in theory, insects have good access to the temporal structure of the stimulus. So when we want to investigate how insects perceive and process these temporal stimulus information, we need to precisely control the odorant stimuli. And this we do with an odorant delivery device which is capable to produce arbitrary binary mixtures where we can let fluctuate the odorants with frequencies up to 70 hertz. So these coherence plots show the temporal resolution of the odorant stimuli. And one important aspect is, or property, is that you can take different odorants which differ in volatility and differ in their stickiness to all the internal materials of an olfactory meter, which low pass to a different degree. These odorant stimuli, and if you apply these fluctuating odorants, so these traces show six different odorants, the temporal resolution is very similar. So this odor delivery device, I brought the heart of this. This is a polyether ketone body, which we equipped with fast-switching valves. And everybody can easily build this thing. So we supply in this paper, we provide all the plans to cut middle this. And it allows highly precise and from set up to set up highly reproducible odorant stimuli. Okay. So first we wanted to know whether insects can use odorant onset asynchrony to segregate odorants. And we did this experiment in honey bees. Because it's very easy to ask a honey bee whether she smells a certain odorant. And the logic or the experiment was as follows. We trained the honey bees to an odorant A to recognize this odorant. And then we tested the same honey bee with three mixtures of this odorant. So we mixed this odorant A with a novel odorant B. And we either presented this mixture as a synchronous mixture where the onset was simultaneously of these two odorants or we shifted the onset of the two odorants by just six milliseconds. And perhaps you see this little shift, so this is true to scale. The odorant stimulus is 800 milliseconds long. And this little shift in the onset is six milliseconds. And how do we ask a bee? Do you smell odorant A? We do a classical conditioning. And this is shown here in this movie. So the odor is presented to the bee. Then we feature some sugar. And bees learn to respond to an odor within one trial. So very quickly, if you want to make them discriminate different odorants, you need just three trials. And then 60% or so of the bees would discriminate. And this is then the performance in the test after they learn that odor A is good. 50% of the bees responded to the synchronous mixture. And 70% roughly responded to these asynchronous mixtures. And one important finding was that it doesn't matter whether this mixture started with a rewarded odor A to which they responded or whether it started with a novel odorant. And so this data looks as if bees can use this asynchrony in the onset of the odorant to segregate these two odorants. Yes, please. Yeah. Yeah. Yeah. Exactly. So this F3B was stimulated with each of these three stimuli, one. So each bee received first three panic trials and then three test trials. But each bee got a different. I mean, the sequence was permeated as was the identity of the odorant. So this experiment was controlled with respect to the odorant and the sequence. And so bees show extinction. If you present these odorants over and over, they will learn that these odorants are not associated with sugar, but this extinction is rather slow. So it could be, but we don't see the learning. So this extinction learning, yeah, we don't see it in this data. Because we randomized the sequence, yeah, but if I would present one odor again and again, the response probability would go down to zero after a while. I didn't test it. So I used, this was like a preliminary experiment. This was the only, I did another experiment where I varied something else, but not the concentration. Yeah. So these concentrations are in the dynamic range of receptor-neuron activity. So it's diluted one to 1,000 in liquid and we know that we are not saturating the receptor neurons. Yeah. So we tested this in drosophila, how mixtures interact at the receptor level and also within one-cent-syllum, these different neurons which are housed in one-cent-syllum, how they interact with each other and we don't find a strong effect, but I can give you more details to this later. More questions? Okay. So then we proceeded, and now first I want to say that it's not only bees who can do it. So there are several invertebrate species which are capable to separate odorants either through spatial cues or temporal cues. All these animals, they have exposed olfactory organs, yeah, which are in direct contact with the air. What I found extremely exciting here that two days ago we learned from Tim Holy that even mice probably could solve this task because the spatial arrangement of odorants is mapped onto a spatial map in the nasal epithelium. So that probably would allow mice to use the spatial information to do an odor segregation. And we learned from Andreas Schäfer that mice are also sensitive for high-frequency temporal stimulus information. So I think these results are, I mean, at least for me, super exciting. I couldn't even get a better start on these ones. No, but I saw how sad you were when this disappeared, so I thought I saved it for your children. Okay, but now we want to know how much of this temporal information is actually available for the insect. So we recorded the temporal position of insect olfactory receptor neurons. Before I show you the data, I want to make a claim. I know first, so receptor neurons and insects, I have to introduce perhaps them quickly. So this is trosophila, and the following data which I will show you is all trosophila data. So trosophila, smell with their antennae, these are the antennae, they are covered with senzilla, and each senzillum houses two or three or four receptor neurons. The good thing is you can identify them all. It's very easy to find the same receptor neuron again and again. And this is the way how odor coding is normally quantified. You look at a population of receptor neurons, so this is worked by Marine de Bruyne. You give an odorant, and then you calculate the spike rate over a window of 500 milliseconds or any window. And then you see that each odorant activates a particular pattern of spike rates across those receptor neurons. These patterns allow us very well to identify any odorant. So one could think that these spike rate differences across receptor neurons encode the odorant identity. The disadvantage of such a rate code is that it takes a lot of time to be read out. So an odorant, two types of receptor neurons, odorant comes, receptor neurons change their rate. In order to read out the rate code, you have to integrate over at least two spikes. So the rate and also the precision with which these neurons produce these changes in spike rates then limits the temporal resolution of the odor code. But there are faster ways to encode information in the nervous system. You could just look or integrate over a small time window so that you, so me, I'm the post-anaptic neuron now and look which neuron populations fire in synchrony within a given time window. Such a synchrony code is limited by the jitter with which these receptor neurons generate their spikes. Another way to encode odorants is by looking at these relative patterns of first spike latencies and Andreas Schaefer modeled this and showed that this is a feasible coding scheme even for mice. Again, the temporal resolution of such a coding scheme is limited by the temporal position. So in fact, all coding schemes are limited by the temporal position of spike time generation. So what is the temporal position? That has never been directly measured in insects. One could think that insect receptor neurons are super fast because first they are exposed to this is one, one syncyllum of a drosophila. So you see these little holes, perhaps they are in direct contact with, with the air odorant molecules diffuse through these holes. They diffuse through the dendrites of the receptor neuron and then they directly open channels. So this is the transduction process in insects that could potentially be very fast. And so we wanted to know how fast is this process. So previous studies indicated that the minimum response latency of an insect receptor neuron is 10 milliseconds. Some other claim that it's 30 milliseconds, so not very fast. Okay, how did we address this question? So what is the speed of odor transduction and what is the temporal position of odor transduction? We did paired recordings from the same receptor neuron type in different syncylums by poking in tungsten electrodes into syncylums of the same type. Then you get these action potential traces. So you can identify receptor neurons based on the heights of their action potentials and then we presented a very fast odor pulse in order to mimic the situation that an insect is flying through a narrow odorant filament. So our odorant stimulus had a rise time of 3.6 milliseconds. So rise time from 5 to 95 percent of the signal of the stimulus. So this is the method. This is some data. So these raster plots show action potentials recorded in the one neuron and in the other neuron. So same receptor neuron type in one fly to 10 repeats of these odorant stimuli. Here we used the thielbutyrate at a concentration of 10 to the minus 5. So diluted in a solvent mineral oil, 10 to the minus 4, 10 to minus 3. A bar is 100 millisecond. So what you see is as you increase the concentration, these spikes, they tightly face lock or tightly lock to the stimulus. Okay. From this data, we first measured the response latency. So the time between arrival of the odorant and the generation of the first spike. On the x-axis, you see the concentrations. So we started with blank air and then represented the odorant concentrations, yeah, different odorant concentrations, so increasing concentrations. So these are lock orders of concentration. Note that the y-scale shows the response latency in milliseconds at the lock scale. So if you use very high concentrations, you get very short response latencies. So the median response latencies are depicted here. So 2.5 milliseconds, for example, at a high concentration of the thielbutyrate. If you record in these AB2A neurons, for example, these different points show the data from the single neuron types, neurons. Okay. This shows that odor transduction can be very, very fast. Then we wanted to know how precise is it. And in order to quantify the precision, we took the standard deviation of the first spike latencies. And we call this jitter. So this is also given in milliseconds. And again, these numbers here, they show the median jitter values over different animals, different neurons. And we find that as you increase the concentration, the jitter goes down to below a millisecond. So odor transduction is not only fast, it's also temporarily very precise. And this is, in fact, in the range of the jitter of mechanosensory or visual neurons. So it's not much slower or not much less precise than these temporarily very precise mechanical or even auditory receptor neurons. Yes. And not much. I will show this data for different odorants. It does not differ much. Could you speak up? Yes. I will show that we can use this. Now we want to compare it with the rate changes. In order to make it better comparable with these millisecond time scales, I plot here not the rate but the inverse rate, which means the interval between two consecutive spikes. And what we found is that, so first at highest concentration, we have the shortest interval between two spikes, three milliseconds, that corresponds to 330 spikes per second. And the interesting thing is that these changes in rate, they vary over a lower range of concentration. They vary over three orders of, lock orders of magnitude. Whereas the first spike latency or the jitter varies over at least one or two lock orders larger range. So you are more sensitive if you could use this timing information at the lower concentrations and you don't saturate as quickly at the higher concentrations. So there is more at least odorant concentration information contained in the precise spike timing than in the rates. And so to demonstrate that this precise spike timing is not an artifact of super high odorant concentration, I highlight here a concentration 10 to the minus 4 where we just see a significant response change or change in the rate. And we find a temporal position of 2.1 milliseconds already and a response latency of 10 milliseconds, which, yeah, which is short. It's the maximum spike rate. So we looked during the stimulus and took the maximum spike rate and here then depicted as the minimum inter-spike interval. But we find the same if we take just the interval between the first and the second spike, it gets really worse if we integrate over 100 milliseconds. So take the average rate, then we lose dynamic range massively, yeah. Yes, a little bit. I will show some data perhaps which will answer your question, yes? Population rates, but you mean across different neural populations? I will show it also, yeah, okay. So the cool thing, what you can do in Trozophila, which you cannot do in mice, is the following calculation. So we know exactly how the circuit is, the architecture of the circuit. So thanks to work from the Hansen's and Zilke's Access Lab, we know that there are exactly 23 of those AB2A neurons. So the data I showed you came from these neurons. We know also from their work is that they project onto one projection neuron in those glomerulus. And we know from Kazamas and Wilson's lab, Rachel Wilson's lab, that all receptor neurons may contact to this projection neuron or the other way around, perhaps each projection neuron in the Trozophila glomerulus gets input from all receptor neurons. So we know the conversion rate, which is here, 2321. This now allows us to estimate what is the minimum time window, a readout neuron, so projection neuron, which would be the analog to a midfield cell. What is the minimum time window of integration time needed to discriminate odor evoked spikes from spontaneously generated spikes? We did it by just counting the number of spikes across these 23 neurons in a given time window, let's say one millisecond. We get the distribution. So we have a perhaps mean number of spikes, let's say 70 spikes, and a standard deviation. And we do the same for a one millisecond time window before the stimulus. So we get also a mean spike number and distribution. And then in order to quantify the distance between or the separability between the odor evoked spikes and the spontaneously generated spikes, we just take the difference in these two means between the odorant and spontaneous spikes and normalize it by the standard deviation. So this is a common measure, which is called sensitivity index. Rachel Wilson used it to do exactly such a calculation. She called it detection accuracy. So it's a measure for how well you can separate two signals. These integration time windows, we were sliding over the time in order to see how does integration time or odorant information changes over time. We took the real data, we shuffled it, and we always picked like 23, but it's, yeah, so it's a real data-based simulation. So if we make this time window, for example, 32 milliseconds long, we count the number of spikes in 32 milliseconds and then measure the distance to the spontaneously evoked spikes in a 32 millisecond time window counted before the stimulus. Then we see that, so this is this magenta line that, like, after seven milliseconds or so, we cross a threshold of two, which would mean that our signal is like two times larger than the standard deviation of the noise, yeah, so we can express it like, simplified like this. And then it increases as you slide your window forward. The interesting thing is you can make your integration time windows very short, yeah, so if you look at the black trace, this is this signal detection accuracy if you have just a one millisecond time window. You also reach, within the same time, the threshold of two detection accuracy units. It does not go as high, but I mean I would think we don't need this high certainty as a read-out neuron that there was an odorant, probably two standard deviations above noises, probably enough. The interesting thing now is that in a time window of one millisecond, a maximum of one spikes fit per neuron, yeah, because neurons have a refractory period, so they need at least two milliseconds, let's say, to generate an X spike. So that means that a purely spike timing-based code, so if you just count the number of spikes in a very narrow time window after your odorant stimulus is enough to expect a lot of odorant information. Here we wanted to know what is the minimum time window now, the minimum integration time window or read-out neuron would need in order to cross this two standard deviation threshold. And we found ridiculously short integration time windows, so at the highest concentration you just need 10 microseconds to be certain that there was an odor, yeah. And even at concentrations, 10 to the minus 6, perhaps now I should switch to the next slide, so sorry it looks a bit different, but this is now the same stimulus detection concentration dependency what I just showed on the previous slide, so very short time windows are only needed for high concentrations. And now if we look at a concentration of 10 to the minus 6, where we don't see any signature of an odorant in the rate, if we just look at the rate, we see already that 4.4 milliseconds of integration time is enough to detect this odorant. And then with the next highest concentration you are already in a time window where only one neuron fits in, where you don't, where you can't have a, yeah, there's no need to read out the rates at any concentration actually, yes, yes, exactly, yeah. And I mean, I really like Dema's primacy code, I mean these Missile Booty rate for this is not the best ligand for this receptor neuron, yeah, so if you would take the best ligand, you would get even shorter times, yeah, and then if only six glomeruli would be sufficient to encode any given odorant, odorant detection identity, identification could happen in no time, yes, but I have to show you the circuit, I don't show it in the talk, but later I want to show it, I think it's possible, okay. So odorants are encoded in, yeah, in a combinatorial pattern, so and here I want to give an impression of the spatial or combinatorial pattern of activity if we take timing. So we recorded from two receptor neuron types, this is the AB2A neuron which I showed the data and the AB3A neuron, what I plotted here is, yeah, all the data we had actually for the first spike latencies for two different odorants, so the green is the Missile Booty rate which I just showed, magenta ethyl acetate and the different concentrations and you have to read this plot as follows, so if you look at this box, ethyl acetate at a concentration of 10 to the minus 3, this line here, the horizontal line shows the median first spike latency of the AB2A neuron, so it responds after around 6 milliseconds, yeah, and the boxes show the 5 to 95 percentile range. AB2A responds after 6 milliseconds, AB3A after perhaps 20 milliseconds, so what you see here is that if you like go from up to down, so from low concentration to high concentration, the response latencies become shorter, of course what you also see is that each odor evokes a specific sequence of response latencies, so the AB2A neuron always responds a little bit quicker to ethyl acetate as does the AB3A neuron and for Missile Booty rate it's the other way around, so odorant identity could be encoded in the order of receptor neuron activation, so whenever there's an odorant that first activates AB2A and then AB3A, it's more likely to be a fetal acetate than the tube Booty rate, so this is what this slide shows you, and the order is odorant specific, but it's invariant to concentration, so even though the latencies they vary over two orders of magnitude, the order stays stable, concentration invariant for different odorants, I don't know, I only tested these odorants here, so yeah, but he did calcium imaging, he has no spike data, so from this data I cannot estimate the response latencies, I think I didn't get your question fully, but you have to then also take into account all the other or many other receptor neurons, but first I have to do the experiment, so first I have to present all the other odorants and test with them, okay, so now let's compare this with an rate code performance, so again what is shown here is the minimum inter-spike interval, and one thing which stands out is that these patterns are highly correlated, so a neuron which fires with a high rate to one odorant also has short response latency, so there's a high correlation in these across neuron patterns, but what we see is that a rate code couldn't discriminate so well over a wide concentration range, yes, so this is just a slide which summarizes the work I did with Brian Smith and Rick Gurkin in Tempe where we tried to measure the response latency and the pulse tracking capability of insect antennae, so we shopped off insect antennae of different insect species and found that they could respond to highly concentrated odorants in this case within a few milliseconds, so this fits to the data we got from Trusophila, I didn't show you data showing how well spikes can follow repetitive pulses, but we found that Trusophila receptor neurons they can resolve up to 125 pulses per second as a population, and the same is true for these electroentronogram responses where we record from a huge population of receptor neurons, so in general odor transduction in insects is very fast and has a high temporal resolution, okay, so summary and conclusion, how much time do I have? Ah, super, okay, then time for discussion. Okay, so summary, so we found short first spike latencies of less than three milliseconds, we found that the jitter can be less than a millisecond, we found that stimulus detection could happen within less than 20 microseconds, we have to prove that projection neurons are sensitive actually that they can discriminate or have short integration time, so this will be my next project, so but I don't have data, I don't even have the setup for this yet, so what are the conclusions? First, which is trivial, synchrony across receptor neurons reliably encodes the onset of an odorant stimulus, so yeah when there is synchronous activity across receptor neurons the olfactory system should know that there was an odorant just right now or a few milliseconds before, we found that the latency and also the degree of synchrony or the jitter of the first spike latencies encodes odorant concentration, we found that the order of first spike latencies across receptor neurons encodes odorant identity and importantly is this concentration invariant and you are to conclude further, we think that the olfactory system could use temporal coding scheme for fluctuating odorant stimuli and because it is so highly correlated with these patterns of rates the olfactory system could actually use both, it could use a rate code for more static odorant stimuli and could switch between temporal or rate codes depending on the environment in which it just is and then what I at least find exciting is that the first spike latency code where the sequence or the order of receptor neuron activities encode odor identity is very well suited to do rapid odorant detection which would allow odorant source separation just because it's fast, yeah yes that's it, I want to acknowledge first of all Alfa Renner and Alex Egea Weiss two bachelor students and all these double recordings and data analysis were their bachelor thesis so they came into the lab completely untrained they got prosophala and set up and this is again a beauty of prosophala that you can do these experiments very easily together with Georg Reiser we developed this odor delivery device which I want to advertise Giovanni Galizia houses me so I'm an independent scientist but a little bit dependent because he provides me lab space and yeah he's a great colleague for discussion Christoph Klein-Eidem was also involved in these single-sensitum recordings and then yeah thanks to my HFSP grant team Thomas Novotny, Brian Smith and Riora Hekansaki we are trying to develop robots which can do odor background segregation which are equipped with moth antennae which are genetically modified so that they are responding to single odorants only and yeah thanks for your attention