WEBVTT Kind: captions Language: en 00:00:01.980 --> 00:00:07.620 Hi, my name is Hong Qin and I'm  faculty at the University of Tennessee,   00:00:07.620 --> 00:00:17.760 Chattanooga. This talk is about a project  I was fortunate enough to work on with a   00:00:17.760 --> 00:00:26.940 large team of colleagues from North Carolina  A&T, Spelman College, Catholic University of   00:00:26.940 --> 00:00:33.780 America. These are the collaborators on paper,  but in reality, there are many many more. In   00:00:34.440 --> 00:00:41.100 our collaboration we probably had more than a  dozen team members that have been working on this.  00:00:41.940 --> 00:00:50.340 So this is a National Science Foundation supported  project. It's called PIPP Phase One. My goal is   00:00:50.340 --> 00:00:59.820 to develop an AI based framework to predict  and prevent future coronavirus pandemics.   00:00:59.820 --> 00:01:07.560 Although we say it's coronavirus, the model  now has the possibility to be generalizable to   00:01:09.240 --> 00:01:27.420 presumably many other viral pandemics. So this is the - ok, it's automatically   00:01:27.420 --> 00:01:36.240 forwarding - my apologies. Ok, so this is the  overview of we proposed in the beginning, although   00:01:36.240 --> 00:01:46.080 we we have made a lot of modifications right now.  The idea is still the same. We propose how do we   00:01:46.080 --> 00:01:57.780 predict the virus - a new virus pandemic? Our idea  is first we generate all the possible SARS-CoV-X   00:01:57.780 --> 00:02:03.480 from that are predictable based on recombination  patterns or habitat change. We then   00:02:04.560 --> 00:02:10.860 predict those potential SARS-CoV-X reactions.  Here's the challenge - it's easy to generate   00:02:10.860 --> 00:02:17.340 those sequences, but how do we know which one  actually will become virulent or transmit in   00:02:17.340 --> 00:02:27.000 a human population? That's really the key here.  So we turn to AI. When I say AI is a black box,   00:02:27.000 --> 00:02:32.460 well in this case that's probably also a  blessing because we really don't know how   00:02:33.060 --> 00:02:41.580 a virus jumps from host, from an animal, to a  human to become virulent. So AI is a black box but   00:02:41.580 --> 00:02:47.280 it's probably a great tool we can use even though  it may be a black box. That's my argument here.  00:02:47.280 --> 00:02:54.060 And my apology, I had to black out certain things  because my University is applying for a patent   00:02:54.060 --> 00:03:03.900 based on the function of this. So the goal is here  we have input sequences - this is just primary   00:03:03.900 --> 00:03:10.500 nucleotide sequences. From this sequence, we will  do feature engineering to generate some helpful   00:03:10.500 --> 00:03:15.900 features. Then we use the sequence feature  to feed into our AI model. Based on our AI   00:03:15.900 --> 00:03:25.260 model we will predict the finish. Given the recent  advance of AI, we can also use the [inaudible]   00:03:25.260 --> 00:03:32.100 dynamic simulation or even experimental methods  to verify some candidate genes as the potential   00:03:32.100 --> 00:03:42.840 drivers for the virulence change. Given the AI  model, we also generalize this - basically use   00:03:42.840 --> 00:03:52.380 transfer learning from coronaviruses, presumably  to other viruses such as influenza or HIV,   00:03:55.680 --> 00:04:03.840 HPV. There's many viruses. I'm not actually really  not expert in biology, I'm just computer science.  00:04:03.840 --> 00:04:11.040 And so the challenges we are really dealing with -  my apologies, it's automatically forwarding again.   00:04:13.620 --> 00:04:19.920 The challenge is really to predict  pathogenicity from sequences. We can   00:04:19.920 --> 00:04:24.540 also found a potential rule and test the  general ability by transforming predict   00:04:24.540 --> 00:04:32.940 the - future viruses and predict the outcome.  It will be AI enabled, the early warning system.   00:04:36.540 --> 00:04:43.380 We want to use AI to predict vaccine development.  Given that we know how the virulence variance   00:04:43.380 --> 00:04:50.940 changes. So that's the big picture of this. So here are some results we have   00:04:51.660 --> 00:05:03.960 developed - that we have accomplished. This  is an estimation of the viral fitness based   00:05:03.960 --> 00:05:09.720 on the information - the data we have. This  is actually for the Omicron sub-variant   00:05:11.940 --> 00:05:19.200 and we can compare the Omicron variant using  a method with development pairwise comparison.   00:05:19.200 --> 00:05:23.700 Then we can estimate the relative  difference between those variants.  00:05:24.360 --> 00:05:29.280 Based on this data we can generate  a so-called viral fitness landscape.   00:05:30.240 --> 00:05:36.480 For those aware of the theory of evolution - in  evolution there is a so-called fitness landscape   00:05:36.480 --> 00:05:41.700 and there's really a lot of arguments about that,  but there's also a lot of theory based on the   00:05:41.700 --> 00:05:52.140 evolution landscape. With the fitness landscape of  a potential virus, we could in theory predict its   00:05:52.140 --> 00:06:04.680 evolutionary trajectory. This - we can - we have  a method to do this. This is the estimated result   00:06:04.680 --> 00:06:13.320 for the alpha beta delta omicron subvariants in  the USA. We are also in the process of applying   00:06:13.320 --> 00:06:21.720 this to influenza or smallpox. We don't have  enough data for smallpox or other viruses,   00:06:21.720 --> 00:06:32.460 but influenza seem to be promising. So this is a  significant finding of this - our current project.  00:06:35.100 --> 00:06:43.560 Here's some model results. Without divulging too  much detail into the AI model we have developed.   00:06:44.160 --> 00:06:50.400 There's really a lot of parameters, but the  accuracy of current model can achieve quite a   00:06:50.400 --> 00:07:00.360 high accuracy. Of course, this is based on what we  know. The challenge is really about the unknown.   00:07:03.720 --> 00:07:10.500 Even though we are predicting the unknown, the  data is already known. In machine learning we   00:07:10.500 --> 00:07:16.500 train it with input training and testing.  The challenge about predicting the unknown   00:07:19.440 --> 00:07:31.740 still remains an open question. Some people may argue, well why do we have to   00:07:31.740 --> 00:07:37.140 do deep learning? There is a lot of statistical,  genetics, genomics, quantitative treatment - there   00:07:37.140 --> 00:07:44.340 are so many other methods and why we do we even  have to turn into deep learning? Here, we have   00:07:44.340 --> 00:07:50.340 some evidence that show deep learning, at the  minimum, it detects different signatures from the   00:07:50.340 --> 00:07:56.700 standard genome-wide association studies. On the  left, is a result from our deep learning result.   00:07:58.020 --> 00:08:09.120 This is actually based on the WHO labeled  variances Alpha, Beta, Delta, Omicron,   00:08:09.120 --> 00:08:16.320 and others. We want to see what signature in the  SARS-CoV-2 genome are important to contribute to   00:08:16.320 --> 00:08:27.780 this increase of virulence. Those high signals  you can see at the end of the SARS-CoV-2 genome.   00:08:29.220 --> 00:08:36.060 And around this region I'm highlighting at  about 20,000 to 25,000, that's the spike   00:08:36.060 --> 00:08:45.900 gene which WHO [has identified] and most of the  immune response will be reacted to. Based on the   00:08:45.900 --> 00:08:52.260 conventional knowledge, that should be in fact,  that's how WHO classified the SARS-CoV-2 variants.   00:08:54.300 --> 00:09:00.600 Surprisingly AI models pick up those signals,  but those are not the strongest signal,   00:09:00.600 --> 00:09:05.400 and they pick up many other signals. Some  of the stronger signals are not there.   00:09:06.960 --> 00:09:15.300 By using conventional statistical genomic  methods like Genome Wide Association Study - that   00:09:17.460 --> 00:09:22.620 still picks the spike gene as the  stronger signal. There are some others,   00:09:22.620 --> 00:09:30.840 but not very strong. So in this case, the  AI and the conventional statistical method   00:09:30.840 --> 00:09:35.280 pick, or at minimum, give a  different weight to those signals.  00:09:37.440 --> 00:09:48.780 This is surprising and also reassuring, in  a way. So we picked the conventional wisdom,   00:09:48.780 --> 00:09:53.820 the spiking, but we also picked some  other signals which may or may not be   00:09:55.320 --> 00:10:01.620 verified by the experimental method. But how do  we predict and how do we verify this, right? So we   00:10:01.620 --> 00:10:08.700 use AI to predict many things and we know AI can  generate a lot of false predictions. In this case   00:10:08.700 --> 00:10:18.060 how can we verify it? That's quite challenging. We are trying, in general, two different ways.   00:10:18.060 --> 00:10:29.160 One is to generate a model experimental system  to verify the findings. Our team is using a   00:10:32.280 --> 00:10:39.240 biological model this is the [inaudible]  we also have a cell line system to   00:10:40.320 --> 00:10:46.620 convert those genes into the life cycle  and measure their relative activities.  00:10:49.320 --> 00:10:56.700 Then we also have a computational person to  perform molecular dynamics to simulate how   00:10:56.700 --> 00:11:02.700 those mutations affect [inaudible]  of their activity. In this case,   00:11:02.700 --> 00:11:09.720 ACE2 and RBD, but we we also have other  [inaudible] in which predict in different   00:11:09.720 --> 00:11:17.520 regions how they react with the human genes -  potential targeting the human immune systems.  00:11:19.680 --> 00:11:32.220 We also organize a workshop and expert panels to  discuss how to design trustworthy AI to promote   00:11:32.220 --> 00:11:45.240 social trust in the AI - equitable AI. The COVID  pandemic has shown there is high disparity in   00:11:45.240 --> 00:11:53.880 our current system so if we use AI to predict the  future pandemics, it can it can easily amplify the   00:11:53.880 --> 00:12:03.660 hidden disparities in our current system, probably  in our current data as well. If we recognize this   00:12:03.660 --> 00:12:11.820 and we organize a quite diverse panel, and  including people from South Africa, Kenya,   00:12:13.500 --> 00:12:22.560 there's a person from Europe and across the - with  different background like lawyers, physicians,   00:12:23.280 --> 00:12:35.100 policy makers, governmental - NIST - governmental  agencies to have all kinds of perspective on how   00:12:35.100 --> 00:12:42.480 to promote AI how and to promote trust in  AI in historically marginalized communities.  00:12:45.360 --> 00:12:53.280 We also emphasize a lot of workforce  development given our collaborations with   00:12:55.560 --> 00:12:59.940 Minority Serving Institutions, including  Spelman College, North Carolina A&T,   00:12:59.940 --> 00:13:08.400 and Catholic University of America, which  contain a lot of Hispanic students, and my   00:13:08.400 --> 00:13:12.960 home institution University of Tennessee,  Chattanooga. We also collaborate with a   00:13:12.960 --> 00:13:18.180 hospital system Global South. Lastly I want to thank NSF,   00:13:18.180 --> 00:13:29.820 UTC and the AI Tennessee Initiative for the  support of this. And lastly, I didn't put in   00:13:29.820 --> 00:13:36.720 the slides, NSF has released call for the phase  2 application for a National Center, to be at   00:13:36.720 --> 00:13:43.980 a scale probably 10 times bigger, if not a 100  times bigger so I'm looking for collaborators,   00:13:43.980 --> 00:13:49.560 especially with community engagement, public  policy, I heard someone working policy. So   00:13:49.560 --> 00:13:59.760 I wish to connect after this. So I'll just stop here, thank you.