reference/paper_experiment_plan.md

# Experiment Plan for a Paper

## Proposed Paper Claim

This project can be written as a predictor-guided study of controllable
enhancer sequence generation:

> We compare autoregressive and masked-diffusion genomic foundation models
> for bucket-conditioned enhancer generation on DeepSTARR, evaluating generated
> sequences by activity prediction, sequence diversity, distributional fidelity,
> and diffusion likelihood.

Avoid claiming that diffusion is better until diffusion samples are evaluated
with the same activity predictor used for autoregressive samples.

## Data Still Needed

Required:

1. DeepSTARR train/valid/test parquet splits.
2. Generated sequence JSONL for each method under the same protocol:
   - autoregressive unconditional SFT
   - autoregressive bucket-conditioned SFT
   - masked discrete diffusion bucket-conditioned
3. Predictor scoring output for every generated and reference sequence.

Strongly recommended:

1. Training-set sequence index for novelty / nearest-neighbor analysis.
2. Motif database for regulatory grammar analysis:
   - JASPAR insect/human TF motifs, depending on organism framing
   - or the motif compendium used by DeepSTARR, if available
3. Optional external annotations:
   - GC-matched random sequences
   - shuffled reference enhancers
   - negative / low-activity DeepSTARR controls

Not required unless doing biological validation:

1. New MPRA/STARR-seq wet-lab data.
2. Cell-type-specific chromatin accessibility data.

## Models to Train or Re-run

Minimum model set:

1. Activity predictor:
   - input: enhancer DNA sequence
   - output: two-dimensional DeepSTARR activity label
   - report: Pearson, R2, MAE, MSE for label_0, label_1, and sum

2. Autoregressive unconditional generator:
   - base: GENERator-eukaryote model
   - input: prefix sequence
   - output: continuation sequence
   - purpose: baseline for DNA validity and diversity

3. Autoregressive bucket-conditioned generator:
   - condition tokens: `<sp0>` low, `<sp1>` mid, `<sp2>` high
   - output: enhancer sequence conditioned on activity bucket
   - purpose: test whether simple token conditioning controls activity

4. Masked discrete diffusion bucket-conditioned generator:
   - base: GENERanno masked LM
   - input at sampling: condition token + fully masked DNA sequence
   - output: full fixed-length DNA sequence
   - purpose: compare global fixed-length generation against autoregressive continuation

Important re-run:

Run diffusion evaluation with `--predictor_model ./results/deepstarr_regression/best_model`.
The current diffusion results only contain `diffusion_pll`; they do not prove
activity control.

## Metrics to Test

### Predictor Reliability

- Pearson per label and average
- R2 per label and average
- MAE / MSE
- true-vs-predicted scatter
- residual distribution

### Functional Activity and Controllability

- predicted label_0
- predicted label_1
- predicted activity sum
- bucket ordering: high > mid > low
- target-hit rate:
  - high samples above high threshold
  - low samples below low threshold
- positive_delta_rate against matched references
- on-target/off-target specificity:
  - whether improving label_0 harms label_1
  - whether high activity is specific rather than globally shifted

### Sequence Validity and Diversity

- valid_dna_rate
- unique_rate
- duplicate count
- homopolymer length distribution
- entropy per position
- pairwise edit distance among generated samples
- nearest-neighbor edit distance to training sequences

### Distributional Fidelity

- GC content distribution
- k-mer frequency distance, e.g. 3-mer/4-mer Jensen-Shannon divergence
- generated vs reference PCA/UMAP on k-mer vectors
- diffusion PLL for generated and reference sequences

### Regulatory Grammar

- motif enrichment by bucket
- motif count distribution
- motif positional distribution
- co-occurrence of activating motifs
- comparison of motif grammar between generated and reference sequences

## Figures to Make

1. Pipeline figure:
   - DeepSTARR data -> predictor -> AR generator / diffusion generator -> evaluation.

2. Predictor validation:
   - scatter plot for label_0 and label_1
   - metric bar chart for Pearson/R2/MAE

3. Generation quality:
   - valid rate, unique rate, bp accuracy
   - GC distribution
   - homopolymer distribution
   - nearest-neighbor distance

4. Controllability:
   - violin/box plot of predicted activity by bucket and method
   - delta plot against reference
   - 2D label_0 vs label_1 scatter colored by bucket

5. AR vs diffusion tradeoff:
   - x-axis: diversity or k-mer distance
   - y-axis: predicted activity or target-hit rate
   - marker size/color: valid rate or PLL

6. Regulatory grammar:
   - motif enrichment heatmap
   - motif positional density by bucket

## Tables to Include

1. Dataset and split statistics.
2. Predictor performance.
3. Generator comparison under unified sampling:
   - method
   - condition
   - valid rate
   - unique rate
   - activity score
   - target-hit rate
   - k-mer JS distance
   - nearest-neighbor distance
   - diffusion PLL
4. Ablation table:
   - no condition
   - bucket token
   - label_0-specific condition
   - label_1-specific condition
   - optional predictor filtering or reward optimization

## Minimal Experiments Needed Before Writing the Final Paper

1. Score diffusion samples with the activity predictor.
2. Generate AR and diffusion samples using the same number of samples per bucket.
3. Compute GC, k-mer, homopolymer, uniqueness, and nearest-neighbor metrics.
4. Plot label_0 vs label_1 to show whether bucket conditioning controls both labels.
5. Add at least one negative baseline:
   - shuffled reference enhancer
   - random DNA with matched GC
   - unconditional generator