A note of a virtual conference on AI & CoViD-19 - Section 4
Full Agenda: HAI Stanford
Full video: Youtube
Session IV: Treatments & Vaccines - 4:55:00
Rapid analysis of SARS-CoV-2 Genomic Content Using the Function Genomic Platform
- Structure: What IBM research lab has done in the past several weeks in response to COVID-19.
- COVID-19 data is rapidly growing and dynamic with two examples about Genome Sequence Data (thousands) and its mutation twice per month. Application must scale to keep pace with automation.
- This information is used for developing antivirals, vaccines and rapid diagnostic tests.
- IBM work (Functional Genomics Platform) from Genomes, Genes, Protein Functional Domains. Put it together to automate this process with an API and web browser.
- A demonstration of IBM tech works: processed 25K viral genomes in days. Open access to researchers.
- A description of data in the system: Genomes, Genes and Protein in Domains.
- Link to data: ibm.biz/functional-genomics
COVID-19 Machine Learning Challenges
- Automated literature review: NLP on 45K academic paper updated weekly;
- Forecasting challenge: cases and fatalities in city with countries and policies;
- Dataset challenge: sharing dataset useful for making decision on aspect of the pandemic; (joining data set is hard)
Machine Learning Enabled Systems for Delivering Care to Critical Ill Patients
- Resources drained, pressure and stress, heath work eventually get ill. So, the solution needed to consider complexity and interdependencies.
- Healthcare is a COMPLEX system. Importantly, it helps us define AI solutions from machine learning models to intelligent care delivery systems. (New structure, processes and pattern for delivering care enabled by ML models)
- Example of Stanford Hospital. First started with no AI to learn about the process and understand problems. Then an example of a particular problem statement: how can we identify sick patients early before ICU.
- Understand the current state and barriers -> identify pain points to prioritize by on frequencies and impact in workflow and outcome -> deriving the components for a new system with lessons about a need to have an objective -> Identify a problem which can be solved by machine learning -> Shown an impacted by a vendor -> Visualization.
- Design a human-centric system by an example of patient’s need with COVID-19
AI-Assisted Elderly Care for Acute Infection and Chronic Disease
- How AI can help seniors automatically in life, especially in COVID-19.
- Studies show that COVID-19 hit elder harder than other age groups.
- Various reasons put seniors at a higher risk: age, health condition.
- Question of how to take care of seniors at home, early symptom detection, manage chronic disease management.
- Solution: AI health with camera depth sensors with four steps.
- Privacy-preserving data collection.
- A notice of ethics, privacy and security.
Identifying COVID-19 Vaccine Candidate with ML
- How the human immune system responds to vaccines.
- A description of biological processes with B-cell and T-cells.
- B-Cell: Linear regression structure features. (Discotope2)
- T-Cell: DNN sequence features. (MARIA, NetMHCpan4)
- How these computation tools are used.
- An update from ShangHai Research group.
Repurposing Existing Drugs to Fight COVID-19
- We don’t have enough time (5-10 years normally) for developing new drugs.
- Fastest way is to repurpose existing drugs already on the market and clinical trials.
- Advantage pre-existing drugs: know about effectiveness.
- Apply NLP, protein structure predictions and biophysic to identify drugs.
- Define working hypothesis by mining literature for therapeutic targets: what, why and what evidence? As a content indexing system that let you drill down into document and individual sense.
- Once we have a promising target and mechanism of action, find an existing drug that can find and block an active side of the target by using a biological model.
- Use a biophysic stimulator for docking which can predict a perfect binding.
- Use a visualization to see which one is effective, then find 6 candidates with a profile.
- Computational model is OK to have an experimental validation on cells and animals, but not human.
Tagged #summary, #note, #book.