June 2019 - Candle-aesp - mailman3.cels.anl.gov

Re: [Candle-aesp] priorities for the AESP project
by Yoo, Hyunseung 01 Jul '19

01 Jul '19

I have a simple, master-slave model, parallel Mordred generator (https://github.com/hyoo/parallel_mordred_generator) and tested on a linux box. As I mentioned during the conference, there was a configuration issue; Theta wasn’t happy with conda-installed mpi4py, so I moved to mpi4py module and then the RDKit wasn’t compiled correctly. I will work with ALCF support team to resolve the issue and will try a couple of test runs to understand the throughput and will report back. Harry From: Candle-aesp <candle-aesp-bounces(a)lists.cels.anl.gov> on behalf of "Stevens, Rick L. via Candle-aesp" <candle-aesp(a)lists.cels.anl.gov> Reply-To: "Stevens, Rick L." <stevens(a)anl.gov> Date: Sunday, June 16, 2019 at 11:03 PM To: "Emily M. via Candle-aesp Dietrich" <candle-aesp(a)lists.cels.anl.gov> Subject: [Candle-aesp] priorities for the AESP project Hi, here are some things I’d like to make sure are moving forward. 1. Demonstration of our ability to compute drug descriptors for a large set of candidates in parallel. Ideally we would set up some runs that would compute mordred descriptors for each of the compound in the lists we have in this set of files. Is this something we can get running on Theta? 2. Prototype of our generative model that can be run in parallel to sweep through latent space and generate large stream of SMILEs examples. I’d like to get a version of our VAE models that can run in inference model and do some runs to generate 1M, 10M, 100M example compounds by sampling the latent space in some clever fashion. Ideally I’d like to take the set of generated compounds and search against those in the datasets in #1 to figure out how many are duplicates and how many are novel. For anything novel I would like to compute the Mordred descriptors and store them along with the smiles. Perhaps the right way to do this is to have the inference model directly call mordred so that what comes out of the model is SMILES + Descriptors .. and round robin joined to the RNAseq and SNP features for the Tumors (both CL and PDX) 3. Prototype of our single drug prediction model with UQ in regression mode on growth that is optimized for inference and that can take input from streams or a file. The idea here is that we we will have some number of these models that are running in parallel with #2 With UQ running we want to get out of the model both an estimate of the growth for the given drug/tumor combination but also an estimate of the error on that estimate. This error estimate will need to come from the calibration curves from the UQ. 4. A way to capture the prediction for each generated drug of all the cell lines and PDX models (think of this as a vector of predictions of order ~1K values and error) that we would want to associate with the drug candidate. 5. Some comparison of the throughput of running #1 coupled to #3 vs #2 coupled to #3. Perhaps we can iterate on this via email this week —rick

3 3

I think I forwarded this..
by Stevens, Rick L. 22 Jun '19

22 Jun '19

The component property prediction models are something we need To look at in the context of ATOM workflow. Also this RL based approach looks interesting in general. https://arxiv.org/pdf/1711.10907.pdf

1 0

Re: [Candle-aesp] priorities for the AESP project
by Zhu, Yitan 21 Jun '19

21 Jun '19

Hi, I want to ask who is working on the single drug model (item 3) and whether any help is needed on it. I did some single drug analysis on CTRP data before, using gene expression data and COXEN approach for selecting genes. I don’t know much about UQ though. But I think there is potential to improve our prediction performance of single drug model. Thanks Yitan From: Candle-aesp <candle-aesp-bounces(a)lists.cels.anl.gov> on behalf of "Stevens, Rick L. via Candle-aesp" <candle-aesp(a)lists.cels.anl.gov> Reply-To: "Stevens, Rick L." <stevens(a)anl.gov> Date: Sunday, June 16, 2019 at 11:03 PM To: "Emily M. via Candle-aesp Dietrich" <candle-aesp(a)lists.cels.anl.gov> Subject: [Candle-aesp] priorities for the AESP project Hi, here are some things I’d like to make sure are moving forward. 1. Demonstration of our ability to compute drug descriptors for a large set of candidates in parallel. Ideally we would set up some runs that would compute mordred descriptors for each of the compound in the lists we have in this set of files. Is this something we can get running on Theta? 2. Prototype of our generative model that can be run in parallel to sweep through latent space and generate large stream of SMILEs examples. I’d like to get a version of our VAE models that can run in inference model and do some runs to generate 1M, 10M, 100M example compounds by sampling the latent space in some clever fashion. Ideally I’d like to take the set of generated compounds and search against those in the datasets in #1 to figure out how many are duplicates and how many are novel. For anything novel I would like to compute the Mordred descriptors and store them along with the smiles. Perhaps the right way to do this is to have the inference model directly call mordred so that what comes out of the model is SMILES + Descriptors .. and round robin joined to the RNAseq and SNP features for the Tumors (both CL and PDX) 3. Prototype of our single drug prediction model with UQ in regression mode on growth that is optimized for inference and that can take input from streams or a file. The idea here is that we we will have some number of these models that are running in parallel with #2 With UQ running we want to get out of the model both an estimate of the growth for the given drug/tumor combination but also an estimate of the error on that estimate. This error estimate will need to come from the calibration curves from the UQ. 4. A way to capture the prediction for each generated drug of all the cell lines and PDX models (think of this as a vector of predictions of order ~1K values and error) that we would want to associate with the drug candidate. 5. Some comparison of the throughput of running #1 coupled to #3 vs #2 coupled to #3. Perhaps we can iterate on this via email this week —rick

3 11

Fwd: REAL Webinar July 1st 2019 (working registration link)
by Stevens, Rick L. 21 Jun '19

21 Jun '19

Sent from my iPhone Begin forwarded message: From: Enamine Subscribe News <noreply(a)enamine.net<mailto:[email protected]>> Date: June 21, 2019 at 18:59:45 GMT+3 To: Rick Stevens <stevens(a)anl.gov<mailto:[email protected]>> Subject: REAL Webinar July 1st 2019 (working registration link) Reply-To: Enamine Subscribe News <noreply(a)enamine.net<mailto:[email protected]>> [https://gallery.mailchimp.com/1e7facecbc2f104e4a73e9c76/images/f9fd50f0-9d1…] In a joint venture, Enamine and BioSolveIT built the world's largest chemical space and made it ultra-fast searchable. And now using REAL Space Navigator you can access 11 billion compounds from Enamine REAL space and do efficient hit exploration, from finding previously unknown analogs to scaffold hopping. The REAL space encoded with more than 100 Enamine synthesis protocols and in-stock building blocks, provides an escape from availability bias of current stock screening collections towards IP free areas. The REAL compounds will be synthesized in 3-4 weeks with an exceptional success rate of 80% and above. In this webinar, we explain the genesis and composition of the REAL space, the updates on the REAL database, the enumerated option to surf the REAL space, as well as the search technology to access this vast resource in minutes only. We also demonstrate the user-friendly graphical interface of REAL Space Navigator. Afterward, you may simply download → install & execute to explore this huge resource for free within less than 5 minutes. Join us to learn more about this exciting endeavor. Over 11 billion virtual molecules that become real on demand are a resource that is simply too valuable to miss... guest speaker: Franca Klingler, Principal Scientist Consulting and Services at BioSolveIT GmbH date/time: Monday, July 1st, 22:00 CST (Shanghai), 16:00 CEST (Berlin), 10:00 am EDT (New York) Join Webinar<https://enamine.us11.list-manage.com/track/click?u=1e7facecbc2f104e4a73e9c7…> You receive this e-mail for one of the following reasons: you are our former or existing customer; you opted-in into our newsletters; we have other legitimate reasons to believe this information might be relevant to you. Want to change how you receive these emails? Unsubscribe from this list<https://enamine.us11.list-manage.com/unsubscribe?u=1e7facecbc2f104e4a73e9c7…> | Update your preferences<https://enamine.us11.list-manage.com/profile?u=1e7facecbc2f104e4a73e9c76&id…> | View email in browser<https://mailchi.mp/7a60aa690f82/real-webinar-july-1st-2218265?e=008a372c89>

1 0

revised flow chart
by Stevens, Rick L. 18 Jun '19

18 Jun '19

1 0

prototype flow chart for AESP challenge problem
by Stevens, Rick L. 18 Jun '19

18 Jun '19

1 1

RL-VAE paper
by Stevens, Rick L. 18 Jun '19

18 Jun '19

Worth looking at this.. A few interesting aspects relevant to our Generator problem. https://arxiv.org/pdf/1904.08915.pdf

1 0

Re: [Candle-aesp] priorities for the AESP project
by Brettin, Thomas S. 17 Jun '19

17 Jun '19

For the response model running in inference mode, is this an uno model? I’d like to look at some methods for compressing. From: Candle-aesp <candle-aesp-bounces(a)lists.cels.anl.gov> on behalf of "Stevens, Rick L. via Candle-aesp" <candle-aesp(a)lists.cels.anl.gov> Reply-To: "Stevens, Rick L." <stevens(a)anl.gov> Date: Monday, June 17, 2019 at 6:03 AM To: "Emily M. via Candle-aesp Dietrich" <candle-aesp(a)lists.cels.anl.gov> Subject: [Candle-aesp] priorities for the AESP project Hi, here are some things I’d like to make sure are moving forward. 1. Demonstration of our ability to compute drug descriptors for a large set of candidates in parallel. Ideally we would set up some runs that would compute mordred descriptors for each of the compound in the lists we have in this set of files. Is this something we can get running on Theta? 2. Prototype of our generative model that can be run in parallel to sweep through latent space and generate large stream of SMILEs examples. I’d like to get a version of our VAE models that can run in inference model and do some runs to generate 1M, 10M, 100M example compounds by sampling the latent space in some clever fashion. Ideally I’d like to take the set of generated compounds and search against those in the datasets in #1 to figure out how many are duplicates and how many are novel. For anything novel I would like to compute the Mordred descriptors and store them along with the smiles. Perhaps the right way to do this is to have the inference model directly call mordred so that what comes out of the model is SMILES + Descriptors .. and round robin joined to the RNAseq and SNP features for the Tumors (both CL and PDX) 3. Prototype of our single drug prediction model with UQ in regression mode on growth that is optimized for inference and that can take input from streams or a file. The idea here is that we we will have some number of these models that are running in parallel with #2 With UQ running we want to get out of the model both an estimate of the growth for the given drug/tumor combination but also an estimate of the error on that estimate. This error estimate will need to come from the calibration curves from the UQ. 4. A way to capture the prediction for each generated drug of all the cell lines and PDX models (think of this as a vector of predictions of order ~1K values and error) that we would want to associate with the drug candidate. 5. Some comparison of the throughput of running #1 coupled to #3 vs #2 coupled to #3. Perhaps we can iterate on this via email this week —rick

3 2

priorities for the AESP project
by Stevens, Rick L. 16 Jun '19

16 Jun '19

Hi, here are some things I’d like to make sure are moving forward. 1. Demonstration of our ability to compute drug descriptors for a large set of candidates in parallel. Ideally we would set up some runs that would compute mordred descriptors for each of the compound in the lists we have in this set of files. Is this something we can get running on Theta? 2. Prototype of our generative model that can be run in parallel to sweep through latent space and generate large stream of SMILEs examples. I’d like to get a version of our VAE models that can run in inference model and do some runs to generate 1M, 10M, 100M example compounds by sampling the latent space in some clever fashion. Ideally I’d like to take the set of generated compounds and search against those in the datasets in #1 to figure out how many are duplicates and how many are novel. For anything novel I would like to compute the Mordred descriptors and store them along with the smiles. Perhaps the right way to do this is to have the inference model directly call mordred so that what comes out of the model is SMILES + Descriptors .. and round robin joined to the RNAseq and SNP features for the Tumors (both CL and PDX) 3. Prototype of our single drug prediction model with UQ in regression mode on growth that is optimized for inference and that can take input from streams or a file. The idea here is that we we will have some number of these models that are running in parallel with #2 With UQ running we want to get out of the model both an estimate of the growth for the given drug/tumor combination but also an estimate of the error on that estimate. This error estimate will need to come from the calibration curves from the UQ. 4. A way to capture the prediction for each generated drug of all the cell lines and PDX models (think of this as a vector of predictions of order ~1K values and error) that we would want to associate with the drug candidate. 5. Some comparison of the throughput of running #1 coupled to #3 vs #2 coupled to #3. Perhaps we can iterate on this via email this week —rick

1 0

Population based data augmentation
by Ramanathan, Arvind 11 Jun '19

11 Jun '19

https://arxiv.org/abs/1905.05393 [1905.05393] Population Based Augmentation: Efficient Learning of Augmentation Policy Schedules - arxiv.org<https://arxiv.org/abs/1905.05393> If you have a disability and are having trouble accessing information on this website or need materials in an alternate format, contact web-accessibility(a)cornell.edu for assistance.web-accessibility(a)cornell.edu for assistance. arxiv.org I thought this was a useful approach - given that we already implement population based hyper parameter search, data augmentation with this could be quite helpful. Best, Arvind

1 0