[M5][DRUG-DISCOVERY] Integrate Multiple Myeloma pharmacogenomics and dependency data

[andreazedda]

Objective

Connect bmyCure4MM drug-discovery capabilities to Multiple Myeloma-specific biological dependencies, drug-response data, molecular states, and evidence rather than relying on generic molecular similarity or drug-likeness alone.

This issue owns the MM-specific evidence/data layer consumed by the end-to-end mechanism-to-molecule pipeline in #77.

Initial data sources

• DepMap CRISPR dependencies and omics;
• GDSC drug-response data;
• PRISM drug-repurposing data;
• CoMMpass genomics, transcriptomics, treatment, and outcomes;
• curated MM cell-line metadata and relevant public functional screens;
• ChEMBL, PubChem, PDB, and other compound/target resources as supporting layers.

Scope

• resolve MM cell-line and sample identities across sources;
• construct gene-dependency, drug-sensitivity, target, pathway, and biomarker matrices;
• record tissue, subtype, treatment context, assay, concentration, and quality;
• define train/test splits that avoid cell-line, compound, or scaffold leakage where applicable;
• establish transparent RF/linear/elastic-net baselines before deep models;
• add interpretable genomic/transcriptomic response-prediction baselines;
• connect candidates to the resistance and evidence graphs;
• separate target validation, compound activity, ADME, and clinical plausibility.

Relationship to #77

#57 = MM-specific evidence and dependency substrate
#77 = mechanism → target → modality → molecule → developability → PK/PD → virtual-patient counterfactual pipeline

#57 therefore provides governed evidence for the Mechanism & Pathway Atlas, patient/subtype mechanistic profiles, target prioritization, MM-specific candidate relevance and drug-response validation in #77.

A generic molecular score cannot satisfy #57 or #77. Candidate-level outputs must keep distinct:

binding / docking
molecular similarity
MM dependency
functional response
selectivity
off-target liability
clinical / translational evidence

The three discovery lanes in #77—REPURPOSING, OPTIMIZATION, DE_NOVO_DESIGN—may consume this evidence differently, but all MM-relevance claims require an explicit #57 evidence record or an explicit INSUFFICIENT_MM_SPECIFIC_EVIDENCE state.

Acceptance criteria

• MM-specific evidence is required before a candidate is presented as MM-relevant.
• Cell-line identity and cross-dataset mapping are validated.
• Drug-response units, assay conditions, and censored values are explicit.
• Baseline models and leakage-resistant splits are implemented.
• External validation or held-out domain evaluation is reported.
• Molecular similarity, binding, dependency and clinical evidence remain separate scores.
• Candidate ranking exposes evidence, uncertainty, contradictions and failure reasons.
• GNN or generative models are added only after strong baselines and task definition.
• [M5][DRUG-DISCOVERY] Build mechanism-to-molecule discovery and virtual-patient counterfactual validation pipeline #77 can consume [M5][DRUG-DISCOVERY] Integrate Multiple Myeloma pharmacogenomics and dependency data #57 outputs through versioned, provenance-preserving contracts without duplicating dataset authority.
• A candidate with no supported MM-specific dependency remains explicitly unsupported even when docking, ADME or generic drug-likeness scores are favorable.

Non-goals

This issue does not claim that an in-vitro response or computational score is a clinically effective treatment. It does not own molecule generation, retrosynthesis, developability orchestration, candidate PK/PD or virtual-patient counterfactual comparison; those are integrated under #77.