AI 在药物研发中的应用

Response to: AI has dispointed on Covid

Geeks piled in when the pandemic broke but great promise has not been fulfilled.

Need:

• Data & Method

Result

• CS: ChEMBL

• Wet lab: Examining the data from the ChEMBL SARS-CoV-2 drug repurposing screens

Sources of data: Instrumental to the development of ML algorithms

• Frank et al., In vitro screening of a FDA approved chemical library reveals potial inhibitors of SARS-CoV-2 replication

• Tesia et al., Discovery of synergistic and antagonistic drug combinations against SARS-CoV-2 in vitro

• Bernhard et al., Indentification of inhibitors of SARS-CoV-2 in-vitro cellular toxicity in human (Caco-2) cells using a large scale drug repurposing collection

ML methods: COVID-motivated trends

• De-novo design: not relevant due to time constraints

• Virtual screening: limited to safe compounds

• Combinations are crucially important

• Tight integration of biology and chemistry

Presentations

Learning the language of viral evolution and escape – Bonnie Berger (MIT, Science)

Problem: Viral escape

Alergy: Grammar & sematic change

Novel approach: learn both “grammaticality” and “semantics” using lanaguage model

Results:

Learn viral sequence patterns

Learn “semantic embeddings” of viral sequences

Zero-shot prediction of escape mutations

Model not trained on known escape mutations

State-of-the-art preformance across viruses

Towards therapeutics and vaccines: identify target regions less prone to escape

Language modelling can shed light on those parts of a protein not prone to escape

Software: viral language learning

https://github.com/brianhie/viral-mutation

Casual network models of SARS-CoV-2 and ageing for drug repurposing – Caroline Uhler (MIT, Nature communication)

Graph neural netwoks for COVID-19 drug repurposing

Problem: Never-before-seen disease

Problem: How to represent COVID-19? Map SARS-CoV-2 targets to the human interactome

COVID-19 subgraph

Gordon et al. expressed 26 of the 29 sars-cov-2 proteins and use ap-ms to identify 332 human proteins to which viral proteins bind.

https://www.nature.com/articles/s41586-020-2286-9

Key insight: subgraphs

Idea: Use the pradigm of embeddings to operationalize the concept of closeness in pharmacological space.

Problem: Why are subgraphs challenging?

Problem formulation: Predict links between drug and disease subgraphs

Method: SubGNN – Subgraph Neural networks

Results

Embedding space of COVID-19

https://www.ncbi.nlm.nih.gov/pmc/articles/PMC7280907/

Experimental validations of predictions

Network drugs

Key ML lessions