WEBVTT 00:00.000 --> 00:16.240 Thank you very much for inviting me and I'm going to talk not so much about COVID but 00:16.240 --> 00:22.880 more about doing artificial intelligence models for biomedical research or machine learning 00:22.880 --> 00:25.160 as we call them actually. 00:25.160 --> 00:29.920 First of all, let me introduce myself, I'm a professor for drug mathematics at the University 00:29.920 --> 00:34.960 of Sarland, I hold the Master Mathematics and Peace Dean Molecular Biology and according 00:34.960 --> 00:38.360 to German law, I'm allowed to teach bioinformatics. 00:38.360 --> 00:46.800 I also have a semi-active blue sky account as it became the plot platform of choice these days 00:46.800 --> 00:47.800 in the community. 00:47.800 --> 00:52.720 First of all, let me tell you a little bit about where I am. 00:52.720 --> 00:59.200 Sarland is a little state in Germany on the French border and geographically it's actually 00:59.200 --> 01:02.320 a closer to Brussels or to Paris than to Berlin. 01:02.320 --> 01:11.440 In Sarland we have nice nature, we have nice food, we have some not-so-nice food and we have 01:11.440 --> 01:15.600 the hand-wants Institute for Pharmaceutical Research and the University with the Center for 01:15.600 --> 01:18.600 Bioinformatics which are my home institutions. 01:18.600 --> 01:25.840 So I will take one minute to make some advertising for the Center for Bioinformatics, we have 01:25.920 --> 01:31.440 the full study program in Bioinformatics in Bethlehem, so if you have someone who wants 01:31.440 --> 01:39.080 to study Bioinformatics, consider sending them there, the Bachelor is taught in German, but 01:39.080 --> 01:44.240 Master is fully in English and we have 90% of international students there. 01:44.240 --> 01:51.040 We have five full professorships and two junior professor groups and the spread of topics 01:51.120 --> 01:58.160 is very broad from everything from algorithms to clinical bioinformatics and to a driven 01:58.160 --> 01:59.160 drug discovery. 01:59.160 --> 02:04.000 Enough of that, let's talk about machine learning and biomedical research. 02:04.000 --> 02:11.200 I stole this figure from a paper, it actually reports models that are used in medicinal 02:11.200 --> 02:16.640 imaging, but it doesn't matter because the trend is the same in every field. 02:16.720 --> 02:24.480 Approximately from mid-2010s, the number of papers reporting some machine learning or AI 02:24.480 --> 02:31.680 model in some biomedical domain exploded and there is no stopping to it, so I think it's growing 02:31.680 --> 02:36.160 and growing uncontrollably. 02:36.160 --> 02:41.840 But as we know AI models are black box models and probably some people have seen this picture 02:41.840 --> 02:45.760 before, razor hand, yes of course. 02:45.760 --> 02:51.440 So you know what the trick is, we don't really know how model learns what it learns. 02:51.440 --> 03:00.880 So here we have a classifier for images, huskies, dogs or wolves and the classifier seems 03:00.880 --> 03:07.360 to be very good on the one error, but if you look closer at what the model actually uses, 03:07.360 --> 03:15.200 what information it uses to make predictions, you figure that dog is not in the image at all 03:15.200 --> 03:22.960 and what you actually trained is a perfect snow detector. The trick is yes, dogs are usually photographed 03:22.960 --> 03:30.400 on some neutral background wolves are predominantly photographed in winter on the snow background 03:30.400 --> 03:39.120 and yeah that's how it worked. It's all very funny, until we actually enter some biomedical 03:39.120 --> 03:46.560 domain. And here COVID comes into play when the epidemic started, of course, the number of 03:47.360 --> 03:54.640 machinery in tools to classify CT scans into COVID and non-COVID also exploded and after 03:56.640 --> 04:02.640 some time into the pandemic, some people actually went into trouble of analyzing some of these 04:02.640 --> 04:10.000 tools and what they discovered is that the models just like the huskies wolf classifier didn't 04:10.000 --> 04:18.480 really learn on relevant information. What it learned was some auxiliary in this case for example labels 04:19.120 --> 04:27.840 on the radiograms. The reason for this was that all COVID samples, so images of COVID 04:27.840 --> 04:36.720 CT scans came from one hospital and the images for healthy, long scans came from a different hospital 04:36.720 --> 04:44.240 which were labeling their scans differently. So basically what the model learned was to detect 04:44.240 --> 04:50.560 these labels and unless you really do analysis, in this case it was some feature important 04:50.560 --> 04:56.640 analysis you never know. And yes what they showed here very nicely if you swap labels 04:57.520 --> 05:06.640 the prediction swaps. So the model didn't learn anything. So if we are thinking about how our models work 05:06.640 --> 05:16.880 in general terms, we as developers have control over the blue area in this image but we don't 05:16.880 --> 05:24.480 have much control about the red area where your model is going to be used, where it will be applied 05:24.480 --> 05:30.240 on some real data. So what we have during the development cycle we have trained validation 05:30.240 --> 05:39.840 and test data and in the deployment time we have some new data inference data. So what does 05:39.840 --> 05:46.240 information leakage have to do with this? Let's first define what is information leakage. By definition 05:46.240 --> 05:52.000 it is use of information during model training process that would not be available during inference. 05:53.120 --> 05:59.760 You can think of your training data as a table. You have features in columns and you have samples 05:59.760 --> 06:06.000 in rows and you can have information leakage across both directions. Feature leakage is very easy. 06:06.000 --> 06:10.160 You have a feature that is highly correlated with the variable that you're trying to predict. For 06:10.160 --> 06:15.920 example you are trying to predict yearly salary of some employees and you have a column that 06:15.920 --> 06:22.480 reports their monthly salary. Very useful but not really what you want to your model to learn. 06:23.840 --> 06:30.960 Another type of leakage is across samples where for example you have samples in the training 06:30.960 --> 06:36.000 set and in the test set that are just identical. You have not cleaned your data properly. You 06:36.560 --> 06:41.440 have duplicates and accidentally they end up in training data and test data. 06:42.720 --> 06:50.080 Which can be theoretical taking care of relatively easily. But there are other more subtle ways 06:50.080 --> 06:57.680 we how this information from the training data can leak into your test data. For example 06:57.680 --> 07:03.520 yeah, parametric visualization, non-identical distribution data when your positive and negative 07:03.680 --> 07:10.640 samples are differently distributed. So what you can try to do since you don't have control over 07:10.640 --> 07:18.720 report. You don't know what will happen during inference time. You can a little bit pay attention 07:18.720 --> 07:32.080 to your training process to try to stop leakage from happening. We published a paper recently where 07:32.080 --> 07:37.760 we have formulated seven questions that every prudent researcher should ask themselves 07:38.640 --> 07:46.320 when training AI models. But I'm going to focus on two of them that are related to a specific type 07:46.320 --> 07:53.840 of data leakage sample similarity. So let's talk about splitting the data while you're training your 07:54.080 --> 08:01.360 model. You have a data set and you need to partition it into a typically into training 08:01.360 --> 08:07.200 validation and test set. Here they are shown in different colors and this is one way of doing this 08:07.200 --> 08:14.880 and this is a different way of doing this. And you see that they are different but you probably 08:14.880 --> 08:25.120 don't know what for. So the idea is that if your partitioning puts the training and test data apart 08:26.320 --> 08:33.520 this makes life of the model kind of hard. You train on one space of data and you're testing 08:33.520 --> 08:43.360 on a different space of the data space and in this way you challenge your model. And hopefully 08:43.360 --> 08:53.680 when it comes to inference the data that will be used for inference will most likely it will 08:53.680 --> 08:59.040 come from a different part of the data space. Hopefully once you challenge your model with 08:59.040 --> 09:03.680 this part of the data space it will also do well on a different part of the data space. That's the 09:03.680 --> 09:13.120 hope. But what you can do without really knowing the inference data. So if we 09:13.120 --> 09:20.320 consider typical biological data sets they can be one-dimensional suppose you have a list of 09:20.320 --> 09:25.760 proteins that you want to predict something for or they can be two-dimensional if you have 09:25.760 --> 09:33.360 protein targets and drugs that can interact with them and then you can split these data sets 09:33.360 --> 09:42.960 from easiest splits to hardest splits. So for one-dimensional data set what you can do to make life 09:42.960 --> 09:51.760 of the model harder is to look for similarities in your data and then try to put clusters of similar 09:51.760 --> 09:58.320 data points into one split. So either training or validation or test split. For two-dimensional 09:58.640 --> 10:05.840 data it's harder because you have similarities across two dimensions. So first thing that you 10:05.840 --> 10:10.880 can do to make life of the model a little bit more difficult than just random misleading the 10:10.880 --> 10:17.360 metrics you can split it by column or by row. In this way some drugs for example will be never seen 10:17.360 --> 10:22.880 by the model during training and then in tests you will see how well it performs on unseen drugs 10:22.880 --> 10:31.440 or similar to some targets and then you can try to account also for similarities and create a split 10:31.440 --> 10:40.080 that will first of all put all similar drugs and targets together. In this way you will inevitably 10:40.080 --> 10:45.600 lose some data because some points you will not be able to classify. So what we are presenting 10:45.600 --> 10:54.720 a tool for controlling data leakage that instead of making models to memory allowing models to 10:54.720 --> 11:07.600 memorize their data are to make models generalized better on unseen data. So this tool takes your 11:07.600 --> 11:15.360 data set and if it is a normal molecules molecular data set it automatically creates all possible 11:15.360 --> 11:21.600 splits of different targets. So now if we consider again this drug target interaction example where 11:21.600 --> 11:26.960 you have some molecules and some drugs and you want to predict whether they bind to each other 11:28.000 --> 11:35.440 we can create all these different splits and here are some metrics that I'm not going to go 11:35.440 --> 11:41.920 into very much detail but what our tool can do it can measure the amount of leakage based on 11:41.920 --> 11:48.560 the similarity between the data points and what we have trained several standards machine learning 11:48.560 --> 11:58.240 models on these splits and what we can show is that the less leakage between the splits the 11:58.640 --> 12:07.840 worse performance of the models and also we could show that our tool makes a life of models harder 12:07.840 --> 12:15.040 than any other tool that was available on the market at the time of publication. So yeah to conclude 12:15.040 --> 12:21.680 that data leakage is a problem in predictive models and it should be addressed before you even start 12:21.680 --> 12:28.000 training your model. It should be addressed when you are looking at your data set and 12:28.720 --> 12:36.640 what is also a bit counterintuitive is that better models that can be can better generalize 12:37.680 --> 12:45.280 to unsend data often perform worse at least in the reported benchmarks. Of course in an independent 12:45.280 --> 12:50.960 benchmark the truth will come out because you would probably challenge models with some hard splits 12:51.040 --> 13:00.640 but if you read papers you often see the benchmarks on random splits and this is a bit 13:00.640 --> 13:07.840 disappointing to people who try to develop really good models. So right that's all from my side 13:07.840 --> 13:11.520 for today thank you very much 13:21.040 --> 13:32.080 Thank you so much for listening to this. The question was is there a way to visualize 13:34.160 --> 13:41.280 yes you can is there a way to visualize information leakage we can't think of a way to visualize it 13:41.280 --> 13:49.280 but we can measure it and we have developed a formula that measures similarity between all data 13:49.280 --> 13:54.480 points in the data sets and between the data sets and in a way this is the objective function 13:54.480 --> 13:58.480 of the optimization problem that we are solving in our tool. 14:09.760 --> 14:14.080 So the question was how many GPUs we need to train the models. 14:15.040 --> 14:22.320 The point of the tool that I was talking about and all the analysis that I was talking about 14:22.320 --> 14:28.160 it was not training the models it was actually before it was data analysis and splitting the 14:28.160 --> 14:37.760 data is also a hard task. So what we did we implemented an optimization problem with 14:38.080 --> 14:43.760 solving and optimization problem using integer linear programming. Now we are doing this with 14:43.760 --> 14:52.080 standard solvers so for bigger data set it's not an issue of time time is usually in hours 14:52.080 --> 14:59.120 but it's often an issue of memory. So one of the next questions that we need to solve is to develop 14:59.440 --> 15:03.120 especially a solver for this. 15:11.280 --> 15:15.040 Thank you very much Olga.