DePristo: Layer and layer of matrices, of different dimensions. Ray-loop multiplies and adds; shows basket of animals - cat or dog? #AACR17
1:54pm April 2nd 2017 via Hootsuite
DePristo: Recognizing a cat: learning features from raw, heterogeneous data; no explicit feature engineering req'd #AACR17
1:53pm April 2nd 2017 via Hootsuite
DePristo: Organized in layers, working together to solve complex tasks. Layered network architecture, new training math, scale #AACR17
1:52pm April 2nd 2017 via Hootsuite
DePristo: Deep learning revolution: modern reincarnation of artificial neural networks. Collection of simple trainable math units #AACR17
1:51pm April 2nd 2017 via Hootsuite
DePristo: Trad machine learning - filtering spam via giant rule-based system; new way: build filters that reduce errors and improve #AACR17
1:50pm April 2nd 2017 via Hootsuite
DePristo: AI=making computers smarter; machine learning-making computers that learn #AACR17
Mark DePristo (Google) Deep learning in medicine - an intro and applications to NGS and disease diagnostics #AACR17
1:49pm April 2nd 2017 via Hootsuite
Cohen: Invites others to participate in a Principal Investigators meeting next week in Arlington VA #AACR17
Cohen: From language in paper, to representation, to PySB (Python System Biology) to model to simulations #AACR17
1:47pm April 2nd 2017 via Hootsuite
RT @itsthebunk: .@AACR Summary re #ProjectGenie #AACR17 https://t.co/zhC6XMmpor
1:44pm April 2nd 2017 via Hootsuite
Cohen: Shows example for drug repurposing, from groups in Chicago, Manchester. Another example for Vemurafenib resistence #AACR17
1:43pm April 2nd 2017 via Hootsuite
Cohen: 3 curators, of 10 papers, had 34 interactions across all 3. Machines - found 17/34. Sensitivity 50%; lots of FPs. #AACR17
1:39pm April 2nd 2017 via Hootsuite
Cohen: But that info is distributed all across the paper. Human curators are the ground truth (via MITRE) 10 papers example #AACR17
1:38pm April 2nd 2017 via Hootsuite
Cohen: Reading 'has never been easy'; NLP has faced many challenges. Shows a single paper - wants MAPK1-P phosphorylates SMAD2 #AACR17
1:37pm April 2nd 2017 via Hootsuite
Cohen: Machines read journal articles, 300K at 5s/paper. Assembled, multiple models are managed, then used for testing, reasoning #AACR17
Cohen: Underlying slide available from NCI website here https://t.co/LzUrHxJkjq 22K interaction w/Ras pathway #AACR17
1:35pm April 2nd 2017 via Hootsuite
Cohen: Eventually get to causal understand in complicated systems, where currently correlation exists. Started with Ras, McCormick #AACR17
1:34pm April 2nd 2017 via Hootsuite
Paul Cohen (DARPA) Big mechanism - how machines read the cancer literature and build cell-signalling models #AACR17
1:33pm April 2nd 2017 via Hootsuite
Chin: Training requires right kind of data - real world data is messy and noisy. Yet critical the right kind of training data avail #AACR17
Chin: AI systems are taught, not programmed, by examples of 'right decisions'; ground truth needs to be true. #AACR17
1:32pm April 2nd 2017 via Hootsuite
Chin: AI systems are human/machine hybrids; designed for assisting human decisions; algorithms are tools. #AACR17
1:31pm April 2nd 2017 via Hootsuite
Chin: 3rd challenge transparency. 4th: workforce, who has ability to develop a system that makes sense. 'Not a job for the SW devs' #AACR17
1:30pm April 2nd 2017 via Hootsuite
Chin: EHR, kiosks, devices/wearables. Challenges: training sets. Ground truth 'is very hard - very rare black/white in medicine' #AACR17
1:29pm April 2nd 2017 via Hootsuite
Chin: Building data from clinical setting, real-world data (home, retail sites, 3rd party, pt-generated wearables) #AACR17
Chin: Who to get data into AI system? Understanding the Patient as a person, not just a diagnosis. #AACR17
1:28pm April 2nd 2017 via Hootsuite
Chin: New feature requests were 3%; clinicians asking for something already asked for. Adverse event tracking, and cohort them #AACR17
1:27pm April 2nd 2017 via Hootsuite
Chin: A failure of humans to update the data; 24$ derived data - redundant negation w/in 'exclusion criteria' double-negative use #AACR17
1:26pm April 2nd 2017 via Hootsuite
Chin: Root cause analysis; 47% issues were UI/UX; 23% were conflicting Dx in documentation (conflicting 'definitive diagnosis') #AACR17
1:25pm April 2nd 2017 via Hootsuite
Chin:Training the user, quality improvement/safety; 40 users and feedback from leukemia study. Summaries useful; Rx rec's useful #AACR17
1:24pm April 2nd 2017 via Hootsuite
Chin: (For those watching at home - WSJ backstory a few weeks ago in WSJ https://t.co/WopphGQ1BW #AACR17 )
1:23pm April 2nd 2017 via Hootsuite
Chin: And initial exposure to https://t.co/OAQvbLpb5H data, initial accuracy 60% but rises over subsequent iterations. #AACR17
1:22pm April 2nd 2017 via Hootsuite
Chin: Shows training and test set methodology, and comparison to SME (subject matter expert - experienced oncologist) #AACR17
1:21pm April 2nd 2017 via Hootsuite
Lynda Chin (MD Anderson TX) Big data and AI: developing cognitive applications in medicine #AACR17
1:20pm April 2nd 2017 via Hootsuite
Sawyer: Looked at transcription factors in Mu et al paper, SOX2 reprogramming factor. Across models, SOX2 changed after knock-down #AACR17
12:24pm April 2nd 2017 via Hootsuite
@Maelli14 Yes many fields are merging, some of the most exciting work is being done at the intersection between bio… https://t.co/CdBhUnqqOy
12:23pm April 2nd 2017 via Hootsuite in reply to Maelli14
Sawyer: Lineage plasticity in mPrCa - TP53/RB1 knockdown causes shift to base lineage gene exp '17 Science https://t.co/Z4hWJvB72J #AACR17
12:22pm April 2nd 2017 via Hootsuite
Sawyer: Not onto TP53 alterations in castration-resistant, mPrCa '15 Cell https://t.co/beuDMlNvJM #AACR17
12:20pm April 2nd 2017 via Hootsuite
Sawyer: GR expression is reversible; Enz +/- may drive GR expression; suggests 2-step model of GR regulation; H3K27me3 mark on prom #AACR17
12:15pm April 2nd 2017 via Hootsuite
Sawyer: Two mech - AR bypass, lineage plasticity. AR bypass is via exp of glucocorticoid recept GR '13 Cell https://t.co/PnM92UJEI1 #AACR17
12:12pm April 2nd 2017 via Hootsuite
Sawyer: For mPrCa - 1st-gen AR Rx, relaps AR over-expressed/ampl after Rx; enzalutamide Rx. Resist 2nd time 50% AR mut/amp #AACR17
12:10pm April 2nd 2017 via Hootsuite
Sawyer: Four TKI's now on the market, BCR-ABL. Reminds a sequential Rx may need combination approaches https://t.co/2FvHPsSu7K #AACR17
12:07pm April 2nd 2017 via Hootsuite
Charles Sawyer (Memorial Sloan NY) New twists on cancer drug resistance #AACR17
12:04pm April 2nd 2017 via Hootsuite
Vogelstein: Shows clip from movie Elysium, a Dx 'just like that'. But 50y ago, and MRI machine was at the same place. #AACR17
11:59am April 2nd 2017 via Hootsuite
Vogelstein: Same idea with saliva, and blood. 96% in saliva, larynx 52%. Combined >90% for larynx. #AACR17
11:57am April 2nd 2017 via Hootsuite
Vogelstein: Wang, Papadopoulus manuscript 2017: endometrial 93% detection n=224, Ovarian 43% n=182. Combined: Ov inc. to 71% #AACR17
11:56am April 2nd 2017 via Hootsuite
Vogelstein: Another idea - not only ctDNA but also released gDNA from the localized site. Isaac Kinde, Luis Diaz: use pap smears #AACR17
11:55am April 2nd 2017 via Hootsuite
Vogelstein: Work to be published later this year - combination blood test for pancreatic - >75% of Stage IA/IB. Lennon, Kinzler #AACR17
11:54am April 2nd 2017 via Hootsuite
Vogelstein: So liquid biopsies can use combination of mutations, markers: methylation, transcripts, proteins, metabolites. #AACR17
11:53am April 2nd 2017 via Hootsuite
Vogelstein: Prevention of 30% is 'why isn't it 100%?' Ideas to increase sensitivity of ctDNA - Rx mainstays are in combinations #AACR17
11:52am April 2nd 2017 via Hootsuite
Vogelstein: Perception of success of prevention vs success of Rx is 'dramatically different'. New Rx at 30% = miraculous. #AACR17
11:51am April 2nd 2017 via Hootsuite
