File size: 13,340 Bytes
db32e07 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 | # Commands to run (Spec-RAG retrieval + sanity check + oracle)
Paths below use your env: `Spec-RAG`, `SpecBridge`, `De-SpecBridge`, `MassSpecGym_test.mgf`, and `out/retrieval/`.
---
## Improving retrieval performance (candidates_A / candidates_B)
- **`--ef-search 512`** (default): Higher FAISS HNSW `ef_search` at query time improves ANN recall. Increase to 1024 if you need better recall and can afford slower search.
- **`--formula-filter`**: When MGF has `FORMULA`/`formula`, restricts candidates to same-formula molecules (over-fetches from FAISS then reranks). Use this so Recall/Tanimoto improve when the library contains same-formula molecules.
- **`--chemberta-model`**: Must match the model used to build the library (`build_library.py --chemberta-model`). Default `Derify/ChemBERTa_augmented_pubchem_13m`. If your SpecBridge checkpoint was trained with a different ChemBERTa, set this and rebuild the library with that model.
- **Formula over-fetch**: `--formula-min-fetch`, `--formula-max-fetch`, `--formula-fetch-multiplier` control how many vectors are fetched when using `--formula-filter`; higher values can improve recall when formula buckets are large.
- **Rebuild FAISS with higher quality**: `python scripts/build_faiss.py --library-dir ... --ef-search 256` (default is now 256) before retrieving.
---
## 1) Build library (once)
```bash
cd /cluster/tufts/liulab/yiwan01/Spec-RAG
python scripts/build_library.py \
--smiles-path /cluster/tufts/liulab/yiwan01/De-SpecBridge/data/pubchem_clean.smi \
--out-dir /cluster/tufts/liulab/yiwan01/Spec-RAG/out/retrieval/library \
--despecbridge-path /cluster/tufts/liulab/yiwan01/De-SpecBridge
```
---
## 2) Build FAISS indices
```bash
python scripts/build_faiss.py \
--library-dir /cluster/tufts/liulab/yiwan01/Spec-RAG/out/retrieval/library
```
---
## 3) Train mappers (only if using Variant A without --smited-mapper-ckpt)
```bash
python scripts/train_mapper.py \
--mgf-path /cluster/tufts/liulab/yiwan01/SpecBridge/data/MassSpecGym_train.mgf \
--specbridge-ckpt /cluster/tufts/liulab/yiwan01/SpecBridge/runs/specbridge_align_chemberta_pub_v3g_msgym_mapper_spec/ckpt_001200.pt \
--dreams-ckpt /cluster/tufts/liulab/yiwan01/SpecBridge/data/ssl_model.ckpt \
--out-dir /cluster/tufts/liulab/yiwan01/Spec-RAG/out/retrieval/mappers \
--despecbridge-path /cluster/tufts/liulab/yiwan01/De-SpecBridge
```
---
## 4) Retrieve (Variants A, B, C)
**Variant A (SMI-TED only; pretrained mapper):**
```bash
python scripts/retrieve_generate.py \
--mgf-path /cluster/tufts/liulab/yiwan01/SpecBridge/data/MassSpecGym_test.mgf \
--library-dir /cluster/tufts/liulab/yiwan01/Spec-RAG/out/retrieval/library \
--specbridge-ckpt /cluster/tufts/liulab/yiwan01/SpecBridge/runs/specbridge_align_chemberta_pub_v3g_msgym_mapper_spec/ckpt_001200.pt \
--dreams-ckpt /cluster/tufts/liulab/yiwan01/SpecBridge/data/ssl_model.ckpt \
--smited-mapper-ckpt /cluster/tufts/liulab/yiwan01/De-SpecBridge/runs/smited_mapper_final/mapper_best.pt \
--despecbridge-path /cluster/tufts/liulab/yiwan01/De-SpecBridge \
--variant A \
--K 100 \
--ef-search 512 \
--formula-filter \
--out-jsonl /cluster/tufts/liulab/yiwan01/Spec-RAG/out/retrieval/candidates_A.jsonl
```
**Variant B (ChemBERTa retrieval; no mapper-dir needed):**
```bash
python scripts/retrieve_generate.py \
--mgf-path /cluster/tufts/liulab/yiwan01/SpecBridge/data/MassSpecGym_test.mgf \
--library-dir /cluster/tufts/liulab/yiwan01/Spec-RAG/out/retrieval/library \
--specbridge-ckpt /cluster/tufts/liulab/yiwan01/SpecBridge/runs/specbridge_align_chemberta_pub_v3g_msgym_mapper_spec/ckpt_001200.pt \
--dreams-ckpt /cluster/tufts/liulab/yiwan01/SpecBridge/data/ssl_model.ckpt \
--variant B \
--K 100 \
--ef-search 512 \
--formula-filter \
--chemberta-model Derify/ChemBERTa_augmented_pubchem_13m \
--out-jsonl /cluster/tufts/liulab/yiwan01/Spec-RAG/out/retrieval/candidates_B.jsonl
```
**Variant B formula-keyed candidate-pool mode (direct exact rerank inside the provided candidate pool; no dependency on library overlap, and no FAISS needed unless a query falls back):**
```bash
python scripts/retrieve_generate.py \
--mgf-path /cluster/tufts/liulab/yiwan01/SpecBridge/data/MassSpecGym_test.mgf \
--library-dir /cluster/tufts/liulab/yiwan01/Spec-RAG/out/retrieval/library \
--specbridge-ckpt /cluster/tufts/liulab/yiwan01/SpecBridge/runs/specbridge_align_chemberta_pub_v3g_msgym_mapper_spec/ckpt_001200.pt \
--dreams-ckpt /cluster/tufts/liulab/yiwan01/SpecBridge/data/ssl_model.ckpt \
--variant B \
--K 100 \
--formula-filter \
--candidate-json /cluster/tufts/liulab/yiwan01/SpecBridge/data/MassSpecGym_retrieval_candidates_formula.json \
--candidate-key-field formula \
--out-jsonl /cluster/tufts/liulab/yiwan01/Spec-RAG/out/retrieval/candidates_B_benchmark.jsonl
```
`--candidate-key-field smiles_gt` is still supported, but it is oracle-only benchmarking because it selects the candidate pool using the ground-truth molecule.
**Variant C:**
```bash
python scripts/retrieve_generate.py \
--mgf-path /cluster/tufts/liulab/yiwan01/SpecBridge/data/MassSpecGym_test.mgf \
--library-dir /cluster/tufts/liulab/yiwan01/Spec-RAG/out/retrieval/library \
--specbridge-ckpt /cluster/tufts/liulab/yiwan01/SpecBridge/runs/specbridge_align_chemberta_pub_v3g_msgym_mapper_spec/ckpt_001200.pt \
--dreams-ckpt /cluster/tufts/liulab/yiwan01/SpecBridge/data/ssl_model.ckpt \
--variant C \
--K 100 \
--out-jsonl /cluster/tufts/liulab/yiwan01/Spec-RAG/out/retrieval/candidates_C.jsonl
```
---
## 5) Evaluate spectrum-based retrieval (Recall + Tanimoto)
```bash
python scripts/evaluate_massspecgym.py \
--pred-jsonl /cluster/tufts/liulab/yiwan01/Spec-RAG/out/retrieval/candidates_A.jsonl \
--report /cluster/tufts/liulab/yiwan01/Spec-RAG/out/retrieval/metrics_A.json \
--tanimoto
python scripts/evaluate_massspecgym.py \
--pred-jsonl /cluster/tufts/liulab/yiwan01/Spec-RAG/out/retrieval/candidates_B.jsonl \
--report /cluster/tufts/liulab/yiwan01/Spec-RAG/out/retrieval/metrics_B.json \
--tanimoto
python scripts/evaluate_massspecgym.py \
--pred-jsonl /cluster/tufts/liulab/yiwan01/Spec-RAG/out/retrieval/candidates_C.jsonl \
--report /cluster/tufts/liulab/yiwan01/Spec-RAG/out/retrieval/metrics_C.json \
--tanimoto
```
---
## 6) Self-retrieval sanity check (Priority 2) — true mol embeddings
Index = test molecules only; query = same true molecule embeddings. **Expected Recall@1 ≈ 1.0** for both.
- If either fails → that embedding/index pipeline is broken.
- If ChemBERTa passes and SMI-TED fails → SMI-TED embedding issue.
- If both pass → issue is likely library coverage / eval mismatch.
```bash
python scripts/self_retrieval_sanity_check.py \
--pred-jsonl /cluster/tufts/liulab/yiwan01/Spec-RAG/out/retrieval/candidates_B.jsonl \
--despecbridge-path /cluster/tufts/liulab/yiwan01/De-SpecBridge \
--report /cluster/tufts/liulab/yiwan01/Spec-RAG/out/retrieval/self_retrieval_report.json
```
(You can use any of `candidates_A.jsonl`, `candidates_B.jsonl`, or `candidates_C.jsonl`; they share the same `smiles_gt` per row.)
---
## 6b) Mapped-embedding self-retrieval sanity check
Index = **spectrum→mapper** embeddings of test set (same MGF); query = same mapped embeddings. **Expected Recall@1 ≈ 1.0.**
- Tests spectrum→ChemBERTa-mapped (SpecBridge) and spectrum→SMI-TED-mapped (DreamsToSmiTed or M_smi).
- If either fails → spectrum→mapped-embedding pipeline is broken.
```bash
# With De-SpecBridge SMI-TED mapper (recommended)
python scripts/self_retrieval_mapped_sanity_check.py \
--mgf-path /cluster/tufts/liulab/yiwan01/SpecBridge/data/MassSpecGym_test.mgf \
--specbridge-ckpt /cluster/tufts/liulab/yiwan01/SpecBridge/runs/specbridge_align_chemberta_pub_v3g_msgym_mapper_spec/ckpt_001200.pt \
--dreams-ckpt /cluster/tufts/liulab/yiwan01/SpecBridge/data/ssl_model.ckpt \
--smited-mapper-ckpt /cluster/tufts/liulab/yiwan01/De-SpecBridge/runs/smited_mapper_final/mapper_best.pt \
--despecbridge-path /cluster/tufts/liulab/yiwan01/De-SpecBridge \
--report /cluster/tufts/liulab/yiwan01/Spec-RAG/out/retrieval/self_retrieval_mapped_report.json
```
If you use Spec-RAG mappers instead of the pretrained SMI-TED mapper:
```bash
python scripts/self_retrieval_mapped_sanity_check.py \
--mgf-path /cluster/tufts/liulab/yiwan01/SpecBridge/data/MassSpecGym_test.mgf \
--specbridge-ckpt /cluster/tufts/liulab/yiwan01/SpecBridge/runs/specbridge_align_chemberta_pub_v3g_msgym_mapper_spec/ckpt_001200.pt \
--dreams-ckpt /cluster/tufts/liulab/yiwan01/SpecBridge/data/ssl_model.ckpt \
--mapper-dir /cluster/tufts/liulab/yiwan01/Spec-RAG/out/retrieval/mappers \
--report /cluster/tufts/liulab/yiwan01/Spec-RAG/out/retrieval/self_retrieval_mapped_report.json
```
---
## 6c) True-index + Mapped-query sanity check
**Index** = true molecule embeddings (ChemBERTa/SMI-TED of test SMILES). **Query** = mapped embeddings (spectrum → mapper). So we build a tiny index from the test set’s true mol embeddings, then query with the spectrum→mapper embeddings. This tests how well the mapper aligns spectrum to molecule space (Recall@1/10/50).
```bash
python scripts/sanity_check_true_index_mapped_query.py \
--mgf-path /cluster/tufts/liulab/yiwan01/SpecBridge/data/MassSpecGym_test.mgf \
--specbridge-ckpt /cluster/tufts/liulab/yiwan01/SpecBridge/runs/specbridge_align_chemberta_pub_v3g_msgym_mapper_spec/ckpt_001200.pt \
--dreams-ckpt /cluster/tufts/liulab/yiwan01/SpecBridge/data/ssl_model.ckpt \
--smited-mapper-ckpt /cluster/tufts/liulab/yiwan01/De-SpecBridge/runs/smited_mapper_final/mapper_best.pt \
--despecbridge-path /cluster/tufts/liulab/yiwan01/De-SpecBridge \
--report /cluster/tufts/liulab/yiwan01/Spec-RAG/out/retrieval/true_index_mapped_query_report.json
```
---
## 7) Oracle retrieval (true mol embedding vs library)
```bash
python scripts/evaluate_oracle_retrieval.py \
--pred-jsonl /cluster/tufts/liulab/yiwan01/Spec-RAG/out/retrieval/candidates_B.jsonl \
--library-dir /cluster/tufts/liulab/yiwan01/Spec-RAG/out/retrieval/library \
--despecbridge-path /cluster/tufts/liulab/yiwan01/De-SpecBridge \
--K 100 \
--report /cluster/tufts/liulab/yiwan01/Spec-RAG/out/retrieval/metrics_oracle.json
```
---
## Minimal “all commands” copy-paste (after library + FAISS exist)
```bash
cd /cluster/tufts/liulab/yiwan01/Spec-RAG
# Self-retrieval sanity check — true mol embeddings (expect Recall@1 ≈ 1.0)
python scripts/self_retrieval_sanity_check.py \
--pred-jsonl /cluster/tufts/liulab/yiwan01/Spec-RAG/out/retrieval/candidates_B.jsonl \
--despecbridge-path /cluster/tufts/liulab/yiwan01/De-SpecBridge \
--report /cluster/tufts/liulab/yiwan01/Spec-RAG/out/retrieval/self_retrieval_report.json
# Mapped-embedding self-retrieval (spectrum→mapper; expect Recall@1 ≈ 1.0)
python scripts/self_retrieval_mapped_sanity_check.py \
--mgf-path /cluster/tufts/liulab/yiwan01/SpecBridge/data/MassSpecGym_test.mgf \
--specbridge-ckpt /cluster/tufts/liulab/yiwan01/SpecBridge/runs/specbridge_align_chemberta_pub_v3g_msgym_mapper_spec/ckpt_001200.pt \
--dreams-ckpt /cluster/tufts/liulab/yiwan01/SpecBridge/data/ssl_model.ckpt \
--smited-mapper-ckpt /cluster/tufts/liulab/yiwan01/De-SpecBridge/runs/smited_mapper_final/mapper_best.pt \
--despecbridge-path /cluster/tufts/liulab/yiwan01/De-SpecBridge \
--report /cluster/tufts/liulab/yiwan01/Spec-RAG/out/retrieval/self_retrieval_mapped_report.json
# True-index + Mapped-query (index = true mol emb, query = spectrum→mapper)
python scripts/sanity_check_true_index_mapped_query.py \
--mgf-path /cluster/tufts/liulab/yiwan01/SpecBridge/data/MassSpecGym_test.mgf \
--specbridge-ckpt /cluster/tufts/liulab/yiwan01/SpecBridge/runs/specbridge_align_chemberta_pub_v3g_msgym_mapper_spec/ckpt_001200.pt \
--dreams-ckpt /cluster/tufts/liulab/yiwan01/SpecBridge/data/ssl_model.ckpt \
--smited-mapper-ckpt /cluster/tufts/liulab/yiwan01/De-SpecBridge/runs/smited_mapper_final/mapper_best.pt \
--despecbridge-path /cluster/tufts/liulab/yiwan01/De-SpecBridge \
--report /cluster/tufts/liulab/yiwan01/Spec-RAG/out/retrieval/true_index_mapped_query_report.json
# Oracle retrieval (true mol embedding vs library)
python scripts/evaluate_oracle_retrieval.py \
--pred-jsonl /cluster/tufts/liulab/yiwan01/Spec-RAG/out/retrieval/candidates_B.jsonl \
--library-dir /cluster/tufts/liulab/yiwan01/Spec-RAG/out/retrieval/library \
--despecbridge-path /cluster/tufts/liulab/yiwan01/De-SpecBridge \
--K 100 \
--report /cluster/tufts/liulab/yiwan01/Spec-RAG/out/retrieval/metrics_oracle.json
# Spectrum-based eval (if you have candidates_*.jsonl)
python scripts/evaluate_massspecgym.py --pred-jsonl /cluster/tufts/liulab/yiwan01/Spec-RAG/out/retrieval/candidates_A.jsonl --report /cluster/tufts/liulab/yiwan01/Spec-RAG/out/retrieval/metrics_A.json --tanimoto
python scripts/evaluate_massspecgym.py --pred-jsonl /cluster/tufts/liulab/yiwan01/Spec-RAG/out/retrieval/candidates_B.jsonl --report /cluster/tufts/liulab/yiwan01/Spec-RAG/out/retrieval/metrics_B.json --tanimoto
python scripts/evaluate_massspecgym.py --pred-jsonl /cluster/tufts/liulab/yiwan01/Spec-RAG/out/retrieval/candidates_C.jsonl --report /cluster/tufts/liulab/yiwan01/Spec-RAG/out/retrieval/metrics_C.json --tanimoto
```
|