WEBVTT Kind: captions Language: en 00:00:00.400 --> 00:00:05.920 Hello, everyone my name is Ho-Joon Lee from  Yale School of Medicine and I am going to talk   00:00:05.920 --> 00:00:11.920 about An interactome landscape of SARS-CoV-2  virus-human protein-protein interactions   00:00:11.920 --> 00:00:16.800 by machine learning. There are two objectives.   00:00:19.280 --> 00:00:25.440 The first is to develop the protein sequence based  multi-class machine learning or deep learning   00:00:25.440 --> 00:00:33.040 classifiers for evidence or confidence level  prediction using the Viruses.STRING database.   00:00:34.000 --> 00:00:41.040 The second is to - using those classifiers we  want to create a draft interactive landscape   00:00:41.040 --> 00:00:44.800 of cytoskeletal virus human  protein-protein interactions.  00:00:46.320 --> 00:00:53.200 So here is an overview of our machine  learning and deep learning workflow. So we use   00:00:53.760 --> 00:00:59.280 the Viruses.STRING database, which did not  include the SARS-CoV-2 at the time of the   00:00:59.280 --> 00:01:06.560 analysis. This is the network of PPI  virus-human PPIs which contain more than 80,000   00:01:07.760 --> 00:01:18.240 interactions between about 1,200 virus proteins  from 102 virus species and about 8,500 human   00:01:18.240 --> 00:01:27.040 proteins. And each interaction has a combined  score ranging from zero to one thousand which   00:01:27.040 --> 00:01:33.520 we convert into five evidence classes pieces.  And this is the distribution of the number PPIs   00:01:34.480 --> 00:01:42.720 for evidence classes. And we are going to focus  on the experimental PPIs which belong to evidence   00:01:42.720 --> 00:01:53.840 class 3 or 2 based on the zero index here. And  based on the data, we first extract node features,   00:01:54.400 --> 00:02:01.200 another protein features which are fractional  compositions of 20 amino acids. And at this point   00:02:01.200 --> 00:02:08.800 we are developing two different models  - one is more canonical [inaudible]   00:02:09.440 --> 00:02:14.960 models like Random Forests and XGBoost, in this  case. And another one is based on deep learning.   00:02:14.960 --> 00:02:21.040 We specifically use graph neural networks like  GraphSAGE or datalized version of HinSAGE..   00:02:23.120 --> 00:02:30.320 For connected machine learning, we also extract  Edge features which are 72 distance or similarity   00:02:30.320 --> 00:02:36.560 measures between amino acid composition profiles  between virus proteins and human proteins.   00:02:37.360 --> 00:02:40.880 And based on the features, we developed  the Random Forests and XGBoost.   00:02:41.600 --> 00:02:47.920 For Random Forests, we optimize 36 models  by research with temp request regulation   00:02:48.480 --> 00:02:55.520 and 432 models for executive space which has the  same contemporary transplantation. And, in short,   00:02:55.520 --> 00:03:02.640 we obtain up to 67% accuracy and 37%  accuracy for Random Forests cases   00:03:02.640 --> 00:03:10.240 and 74% accuracy and 67% accuracy for XGBoost  cases. And this work, this part, has been   00:03:10.240 --> 00:03:16.000 published as a preprint recently. So  you can refer to the paper in detail   00:03:16.000 --> 00:03:19.840 [https://www.biorxiv.org/content/10.1101/2021.11.07.467640v2].  And for GraphSAGE here, still in   00:03:19.840 --> 00:03:24.800 advanced reading and preparation, but I'm going  to show you, briefly show you, the results from   00:03:24.800 --> 00:03:31.840 GraphSAGE as well. Because this shows more than  70% accuracy which is plenty promising as well.  00:03:34.720 --> 00:03:41.040 And here I'm going to show you a performance  example for the best models for 20% of [inaudible]   00:03:41.040 --> 00:03:48.960 with this random seed. We see, in this case,  when the Forest shows 60% accuracy, XBG was   00:03:48.960 --> 00:03:56.160 67.7% accuracy. And if you look at computer  metrics, again, I'm going to focus on this [EC3?]   00:03:56.160 --> 00:04:03.520 which implies mostly expanded PPIs. And  if we look at the individual classes,   00:04:03.520 --> 00:04:09.440 focusing on the f1-score, the extra booster shows  higher f1 scores across four individual classes.  00:04:12.640 --> 00:04:19.920 Based on, based on this to [inaudible]  boost model. The important features   00:04:19.920 --> 00:04:25.760 were identified using two alternating methods  here. One by Gini index and the other by SHAP   00:04:26.320 --> 00:04:30.000 analysis, which based on SHAP,  came through at the SHAP values.   00:04:30.800 --> 00:04:38.800 And interestingly enough, we see that cysteine and  histidine are most - two most important features.   00:04:39.360 --> 00:04:42.480 Where this minus [C_minus and  H_minus] means that the fraction   00:04:42.480 --> 00:04:49.200 of cysteine between virus and human. And the ratio  means the ratio between the fractions cysteine and   00:04:49.200 --> 00:04:58.640 histidine reactions between virus and humans. One control experiment we performed is to   00:04:58.640 --> 00:05:04.880 compare prediction of experimental PPIs  and - with a prediction of text mining   00:05:04.880 --> 00:05:11.760 PPIs in the virus [inaudible]. Because the  data size, the difference is pretty big here,   00:05:11.760 --> 00:05:18.640 six squared difference, but what we observed here  is that XGBoost, in fact, shows higher accuracy.   00:05:18.640 --> 00:05:26.240 With 94% accuracy compared to 90% accuracy for  the text mining case. So despite the data size   00:05:26.240 --> 00:05:34.160 difference, XGBoost it shows a good prediction  performance. And this is the agreement between   00:05:34.160 --> 00:05:43.280 random force activities for ec3 and test  binding as we expect shows mostly ec1 or ec2.  00:05:46.880 --> 00:05:54.560 So based on those encouraging results, we  applied those classifiers to SARS-CoV-2   00:05:55.280 --> 00:06:02.960 for our second objective in two ways. So  first, we apply that to IntAct database,   00:06:02.960 --> 00:06:10.960 which are a collection of experimental  PPIs. And here I’m showing you the network   00:06:11.680 --> 00:06:18.080 by XGBoost with [inaudible] predicted  evidence. So ec3 for blue, ec4, red.   00:06:18.720 --> 00:06:26.800 So this can be viewed as prioritizing a network.  So although these links would be about 2,000 links   00:06:26.800 --> 00:06:35.040 from experimental data are equally meaningful,  we can also prioritize those links based on this   00:06:35.680 --> 00:06:42.400 [inaudible] class predicted XGBoost in  this case. Secondly, we also apply that   00:06:42.400 --> 00:06:48.160 to protein-wide interaction [inaudible]  the old pairs of more than half a million   00:06:49.360 --> 00:06:58.000 between 27 SARS-CoV-2 proteins and about more  than 20,000 human proteins. And here I'm showing   00:06:58.000 --> 00:07:09.120 you the subset of 22,000 PPIs with evidence class  of at least 2. I either actually use XGBoost or   00:07:09.120 --> 00:07:17.440 Random Forest. And this is the another subset -  140 PPIs with the highest evidence class, 5, by   00:07:18.480 --> 00:07:27.520 XGBoost. And based on this interaction network we  observed that many human proteins are enriching   00:07:27.520 --> 00:07:32.720 vascular smooth muscle contraction and  the targets and also H2A components.  00:07:36.800 --> 00:07:42.320 There are a few more applications of this  work that have been found in the past month,   00:07:42.320 --> 00:07:50.640 actually. So Giuseppe Novelli, who is the renowned  geneticist in Rome, in Italy, he reached out to me   00:07:50.640 --> 00:07:56.000 by email and by surprise, last month. He had  read my preprint telling me about his quality   00:07:56.000 --> 00:08:06.160 therapeutics publication for HECT E3 ligases and  his idea of using the results from this important   00:08:06.160 --> 00:08:12.960 work through this ongoing research. And we  immediately realized that we can help each other   00:08:15.200 --> 00:08:23.600 based on my results on interactive network  results. And we found that HECT-domain protein   00:08:23.600 --> 00:08:28.400 tend to interact with the SARS-CoV-2 proteins  with an evidence class of greater than 2 with   00:08:28.400 --> 00:08:33.760 statistical significance. In other words,  HECT-domain proteins are favored by SARS-CoV-2.   00:08:35.040 --> 00:08:41.200 Based on that observation you're asking whether  there are other protein families favored by   00:08:41.200 --> 00:08:49.600 SARS-CoV-2. In addition we can also extend that  to other virus species like human metapneumovirus,   00:08:49.600 --> 00:08:53.840 which Dr. Novelli is also working on as well.  00:08:55.680 --> 00:08:59.280 So finally, I'm going to show you  briefly about the graph neural networks   00:08:59.280 --> 00:09:07.840 using GraphSAGE and HinSAGE architecture. On  the left are the accuracies by 15 different   00:09:07.840 --> 00:09:13.840 models using three different Java weights in the  columns and five different edge embedding methods.   00:09:15.040 --> 00:09:22.000 As you see, without dropout rates, in fact, we  see more than 70% accuracy values and accuracies   00:09:22.000 --> 00:09:27.440 which are very promising. This is a based on  Viruses.STRING and if we apply that to SARS-CoV-2   00:09:27.440 --> 00:09:34.080 IntAct database, you see the prediction is  enriched with evidence class 2, or 3, in fact,   00:09:34.720 --> 00:09:41.520 which are mostly [inaudible] PPIs. And  this consensus is number of agreements   00:09:42.080 --> 00:09:50.720 among these 15 different models. We see  more consensus of, like, 8-9 for against   00:09:50.720 --> 00:09:58.320 two compared to 6-7 against 1 but I think this  is also very important because - we will see.  00:09:59.840 --> 00:10:03.600 Okay, with that I'd like to thank  my collaborators for very helpful   00:10:03.600 --> 00:10:09.920 discussions and feedback and support. And the Yale  Center for Research Computing for computational   00:10:09.920 --> 00:10:15.760 resources. And COVID HASTE community from Yale’s  School of Engineering & Applied Science. And   00:10:15.760 --> 00:10:23.280 finally the Northeast Big Data Hub Seed Fund  for supporting this work. Thank you so much.