How nirmalkumar31/rbp-protocol-calibration works
This repository is a research pipeline for measuring how negative-example construction changes apparent RNA-binding-protein sequence-model performance, with an earlier variant-scoring study retained alongside it. The main workflow assembles ENCODE peak data and sequence windows, constructs negative sets and chromosome-blocked folds, trains models and composition baselines, then estimates and compares their contributions. Cloud scripts support ingestion and GPU sweeps; local analysis assembles evidence, while verification checks published values. The graph combines sampled source references with README-documented capabilities; unsampled implementation details are not inferred.
Main entry points: cloud_analysis.py Cloud analysis · registry.py Model registry · windows.py Sequence windows · trainer.py Model trainer
Components 24
Data preparation
- ENCODE inputs
src/rbp/data/encode.py - Sequence windows
src/rbp/data/windows.py - Negative sets
src/rbp/data/negatives.py - Fold construction
src/rbp/data/splits.py - Study panel
src/rbp/utils/panel.py - Configuration
src/rbp/utils/config.py
Modeling
- Training data
src/rbp/train/data.py - Model registry
src/rbp/models/registry.py - CNN model
src/rbp/models/cnn.py - Language models
src/rbp/models/lm.py - Model trainer
src/rbp/train/trainer.py
Evaluation and evidence
- Nested estimands
src/rbp/eval/nested.py - Statistical methods
src/rbp/stats.py - AUROC comparison
src/rbp/eval/delong.py - Cloud analysis
scripts/cloud_analysis.py - Evidence verification
scripts/verify.py
Cloud execution
- Cloud ingestion
scripts/cloud_ingest.py - Cloud storage access
src/rbp/utils/cloud.py - Local artifact store
src/rbp/utils/localstore.py - GPU sweeps
cloud/modal/modal_sweep.py
Variant study
- Variant assignment
src/rbp/variants/assign.py - Conservation control
src/rbp/variants/conservation.py
Other components
- Researcher
- ENCODE and references
Connections 25
- Researcher → Cloud ingestion: starts ingestion
README.md - ENCODE and references → Cloud ingestion: provides inputs
scripts/cloud_ingest.py - Cloud ingestion → Cloud storage access: uploads data
scripts/cloud_ingest.py - ENCODE inputs → Sequence windows: supplies peaks
- Negative sets → Sequence windows: uses windows
src/rbp/data/negatives.py - Fold construction → Sequence windows: uses windows
src/rbp/data/splits.py - Researcher → GPU sweeps: launches sweeps
README.md - GPU sweeps → Model trainer: runs training
- Training data → Fold construction: loads folds
src/rbp/train/data.py - Model registry → Configuration: reads settings
src/rbp/models/registry.py - Model registry → CNN model: builds CNN
src/rbp/models/registry.py - Model registry → Language models: builds language model
src/rbp/models/registry.py - Model trainer → Model registry: uses model handle
- Model trainer → Training data: consumes loaders
- Cloud analysis → Nested estimands: computes estimands
- Nested estimands → Statistical methods: uses methods
src/rbp/eval/nested.py - Nested estimands → AUROC comparison: uses AUROC tests
src/rbp/eval/nested.py - Cloud analysis → Cloud storage access: accesses cloud
scripts/cloud_analysis.py - Cloud analysis → Conservation control: uses variant analysis
scripts/cloud_analysis.py - Evidence verification → Cloud storage access: checks evidence
scripts/verify.py - Researcher → Evidence verification: checks results
README.md - Variant assignment → Sequence windows: uses sequence windows
src/rbp/variants/assign.py - Conservation control → Statistical methods: uses estimators
src/rbp/variants/conservation.py - Cloud analysis → Local artifact store: returns scores locally
README.md - Study panel → Training data: selects datasets