Project Plan: HTS_IL17_Psoriasis

Objective

Build a FAIR, locally runnable, AWS-adaptable Nextflow workflow and Streamlit app that integrates public HTS/qHTS-style assay data, psoriasis disease omics, proteomics, protein model features, optional structure prediction evidence, and grounded evidence summaries.

The project should be reviewed as a staff-scientist-level portfolio project: scientifically useful, transparent about limitations, and structured so it could be extended to real screening data.

Scientific Question

Which Th17/IL-17 psoriasis pathway candidates are most compelling when ranked by:

• HTS-style activity and counterscreen selectivity
• psoriasis disease expression evidence
• protein-level validation
• disease-relevant cell-type context
• protein model features
• optional AlphaFold/ESMFold structure confidence
• transparent evidence summaries

Implementation Phases

Phase 1: Local Demo

• Use bundled compact example datasets with public-data-shaped columns.
• Run nextflow run main.nf -profile test.
• Generate tables, evidence cards, report, and app data.
• Launch streamlit run app/streamlit_app.py.

Phase 2: Public Data Expansion

• Add real PubChem PUG-REST retrieval for AID 2604 and AID 2546.
• Add ChEMBL API retrieval for RORC/IL-17 pathway assay records.
• Add processed GSE54456 import and DESeq2-style differential expression.
• Add PRIDE/PXD021673 processed protein quantification import.
• Add GSE162183 processed marker/cell-type summaries.

Phase 3: Protein Models And Structure

• Add ESM-2, ProtBERT, or ProtTrans embedding generation for top candidates.
• Cache embeddings with accession, model name, and model version.
• Add AlphaFold DB lookup or optional ESMFold prediction for top candidates only.
• Keep GPU and large model calls optional.

Phase 4: Grounded Evidence Summaries

• Add a grounded summarization module that consumes workflow tables only.
• Require accession and table citations in every evidence card.
• Emit limitations and next-experiment suggestions.
• Allow local template mode when no model API key is available.

Scoring Model

The first implementation uses a transparent weighted score:

This should stay configurable in future versions.

Actionable Biological Insights To Surface

• Does the candidate align with Th17/IL-17 biology, keratinocyte activation, cytokine signaling, or inflammatory skin proteomics?
• Is the candidate supported at both RNA and protein levels?
• Is the signal specific to disease-relevant cell types, or broadly expressed?
• Does the HTS signal survive counterscreen filtering?
• Is structure evidence high-confidence enough to support follow-up modeling?
• What validation experiment would reduce uncertainty fastest?

Critical Review Notes

• The HTS data are pathway-proximal small-molecule assays, not peptide screens.
• ROR gamma is upstream of IL-17 biology; this supports relevance but not direct IL-17 binding.
• RNA-protein concordance can be imperfect; discordance should be flagged, not hidden.
• Protein language model embeddings are features, not explanations by themselves.
• Structure prediction confidence is not interaction validation.
• Evidence-summary output must be evidence-grounded and citation-backed.

Things To Do

• Replace compact example tables with full public dataset downloads.
• Add raw FASTQ mode for transcriptomics as an optional cloud workflow.
• Add richer cheminformatics for PubChem compounds.
• Add peptide-specific public assay data if a suitable source is found.
• Add model-card-style documentation for protein-model and evidence-summary components.
• Add CI using nextflow run main.nf -profile test.
• Add screenshots after the app is visually reviewed.