>Medical Compound Literature Optimizer

1. Data ingestion:
>Apply a simple, but robust, scraper for the entirety of ArXiv. Include PubMED, bioRxiv, ChemRxiv and patent databases if possible.

2. Label with BERT-likes:
>Applying BERT-likes, extract named entities such as specific compounds, target protein, cell lines, assays, equipment, ect.
>Identify relations between entities
>Categorize findings
>Categorize qualities such as structure (SMILES, InChl, molecular fingerprints), physiochemical properties (logP, TPSA, molecular weight), solubility, stability
>Methodology, target, disease area

3. Node and cluster creation:
>Each paper is a "node" and labels are applied to the node.
>Similarities are mapped and clustered using GNN techniques.
>Clustering can be sorted by compound similarity, quality, methodology or findings, or all of the above.

4. Optimization and hypothesis generation:
>Within a cluster, identify "promising" and "ineffective" results
>By analysing the differences in methodologies and qualities of the "promising" and "ineffective" within the cluster, potential optimizations can be infered

5. Output:
>This model could provide methological improvements, structural modifications, repurposing, hypothesis generation

6. Benefits:
>Highly accelerated discovery
>Reduced redundancy
>Discovering non-intuitive or overlooked links
>Connecting findings from isolated fields or papers
>Metastudy applications
>Bias reduction
>Pre-emptive research optimization

7. Nuances:
>Utilizing an established general concensus as a dataset bootstrap (I.e. fundamental principles, well-established mechanisms)
>This AI is not to "interpret" subjectivity of what is "effective" or "ineffective", human-in-the-loop is still required to actually test the output. The AI is to process mass sparse data, identify potential correlations and apply pattern recognition. The human is to refine ambiguous labels and ultimately deciding if the causation is present.

>>40620136
Don't be so quick to dismiss my design propsals tyrone.