Experiments
Switch language: 实验设计
Experiment philosophy
The project is designed as a structured benchmark rather than a single-model
implementation. The experiment board in TASK.md organizes work into a model
matrix, fusion ablations, and cross-cell-type generalization analysis.
Primary experiment matrix
The current plan tracks experiments E01 through E16:
E01: M1 CNN single-branch baselineE02: M2 CNN dual-branch baselineE03: M2 withconcat_sub_mulfusionE04: M3 k-mer + CNNE05: M4 BiLSTME06: M6 standard TransformerE07: M11 CNN + BiLSTME08: M12 CNN + TransformerE09: M13 DNABERT-2 frozenE10: M13 HyenaDNA fine-tuneE11: M9 MambaE12: M7 Linear TransformerE13: M8 iTransformerE14: M10 RWKVE15: M5 mLSTME16: M14 MAE pretrain + fine-tune
Fusion ablation
The fusion strategy itself is a research axis. The documented ablation plan includes:
F01: concatF02: addF03: subtractF04: multiplyF05: bilinearF06: concat_sub_mul
Why this comparison matters
Enhancer-promoter prediction is naturally a pairwise learning problem. A strong encoder can still underperform if the interaction representation is weak. By explicitly exposing multiple fusion operators, the project separates two questions:
how expressive the sequence encoder is;
how effectively enhancer and promoter features are combined.
Seed strategy and reproducibility
Each experiment is intended to run on five random seeds:
seed
0seed
1seed
2seed
3seed
4
This reduces the risk of over-interpreting a favorable or unfavorable single run and gives a more stable estimate of expected behavior.
Expected outputs per experiment
Every seed directory stores:
configuration snapshot;
checkpoints;
training history;
test metrics;
ROC curve;
PR curve.
At the experiment level, summary statistics are intended to report mean and standard deviation across seeds.
Showcase value
This experiment design demonstrates more than model implementation. It shows:
baseline preservation;
controlled architectural expansion;
reproducible training organization;
explicit ablation thinking;
practical readiness for comparative sequence modeling studies.