AI for Bioinformatics & Drug Discovery

Two briefs. Real-shaped work.

1. Project 1

   Protein-structure exploration on a marine-biotech target

   Choose one target from a shortlist of marine-microbe enzymes with industrial or environmental interest (for example, alginate lyases, chitinases or cold-active proteases). Starting from its UniProt sequence, find and characterise its homologues with BLAST, obtain an experimental structure if one exists and an AlphaFold model, read confidence with pLDDT and PAE, compare predicted and experimental structures where possible, and identify conserved residues that could form the active site. End with a set of claims you can defend, claims you cannot make, and two concrete wet-lab experiments that would test your best hypothesis. All work must be reproducible from a notebook, with every AI-assisted step marked and checked.

   You hand in

   • A Colab notebook that runs top to bottom with header, provenance and toy tests
   • BLAST hit table with an audited summary (E-value, identity, coverage for every claim)
   • Structure confidence checklist and annotated screenshots (pLDDT, PAE, superposition)
   • RMSD results on all residues and on high-confidence residues, with numbering fixed
   • Conserved-residue analysis mapped onto the structure
   • A two-page report: claims, non-claims and two proposed wet-lab experiments

   Data: Public data only: UniProt entries, NCBI BLAST results, RCSB PDB experimental structures and AlphaFold Protein Structure Database models for a Saarthi-provided shortlist of marine-microbe enzymes. No unpublished or proprietary data.. Reviewed by Saarthi against a published rubric, signed off by a mentor.

2. Project 2

   Validated AI literature review on a cancer-biology question

   Pick a focused cancer-biology question from the provided list (for example, the evidence linking a specific gene's expression to prognosis in oral squamous cell carcinoma). Use an assistant to design a reproducible PubMed search, run it yourself, screen records with recorded reasons, and extract findings with an assistant under a quote-or-'not stated' rule. Verify every reference and every quote: identifier resolves to the same title, the claim appears in the text, and no retraction applies. Report how many assistant errors you caught and of which type. Deliver a synthesis that weighs evidence by study type and states its limitations plainly.

   You hand in

   • Search log: query strings, dates, result counts and any changes with reasons
   • Screening record: includes and excludes with reasons
   • Evidence table with PMIDs, study type, n, findings and verified quotes
   • Assistant error log classified as fabricated, misattributed or outdated
   • A 1,000–1,500 word synthesis separated by study type, with limitations
   • AI-use disclosure paragraph written to a named journal's current policy

   Data: Public PubMed records and open-access full texts retrieved by the student's own logged search. No patient-level data.. Reviewed by Saarthi against a published rubric, signed off by a mentor.