ADMET predictions

tdudgeon · April 18, 2020, 10:49am

I am coordinating generating ADMET prediction for compounds being proposed for purchase or synthesis. This is largely being done with help from the OpenRiskNet and NanoCommons communities.

Initial work is focused on the non-covalent inhibitors but there is no reason why it can’t be extended to covalent ones as well.

We have a range of models accessible being provided by a number of communities. I will post examples separately as this forum only allows me to include 2 links
in a post
For some models 10,000’s of molecules can be tested, for others we need to prioritise a smaller number.

Also important is that we are trying to coordinate amongst those parties to provide an expert way to assess all this information so that we can condense down a large number of predictive models into a simple traffic light classification to assist compound selection.

This is all work in progress and if anyone has data or expertise that they would like to make available then please let me know.

tdudgeon · April 18, 2020, 10:50am

And here are some examples:

github.com

tdudgeon/jupyter_mpro/blob/master/enalos_analysis.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 18,
   "metadata": {},
   "outputs": [],
   "source": [
    "import pandas as pd\n",
    "import numpy as np\n",
    "\n",
    "import plotly\n",
    "import plotly.graph_objects as go\n",
    "import plotly.express as px\n",
    "import plotly.io as pio\n",
    "\n",
    "from rdkit.Chem import PandasTools"
   ]
  },
  {

This file has been truncated. show original

github.com

tdudgeon/jupyter_mpro/blob/master/prosilico_analysis.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 1,
   "metadata": {},
   "outputs": [
    {
     "name": "stderr",
     "output_type": "stream",
     "text": [
      "RDKit WARNING: [11:31:40] Enabling RDKit 2019.09.3 jupyter extensions\n"
     ]
    }
   ],
   "source": [
    "import pandas as pd\n",
    "import numpy as np\n",
    "\n",
    "import plotly\n",

This file has been truncated. show original

tdudgeon · April 18, 2020, 10:51am

And one more example:

github.com

tdudgeon/jupyter_mpro/blob/master/marionegri_analysis.ipynb

{
 "cells": [
  {
   "cell_type": "code",
   "execution_count": 22,
   "metadata": {},
   "outputs": [],
   "source": [
    "import pandas as pd\n",
    "import numpy as np\n",
    "\n",
    "import plotly\n",
    "import plotly.graph_objects as go\n",
    "import plotly.express as px\n",
    "import plotly.io as pio\n",
    "\n",
    "from rdkit.Chem import PandasTools"
   ]
  },
  {

This file has been truncated. show original

mc-robinson · April 19, 2020, 10:20am

Thanks @tdudgeon, this is great, and seems like it will be really helpful moving forward.

Just wondering, do we have any idea what the expected accuracy/error are for these predictions on each endpoint? I see the enalos data has a flag for “reliable” and “unreliable” predictions, which seems helpful along these lines.

And I am devoting my day tomorrow to getting all of the submission data tidy for analyses like this. Will alert you when I finish!

tdudgeon · April 19, 2020, 10:38am

This depends totally on the type of model that is being run. Some have an estimate of accuracy, others don’t.
We will try to include that aspect into the way the predictions get summarised.

Tim