code/docs/RETRIEVAL_PIPELINE.md

Retrieval pipeline (Retrival_plan.md)

This implements the retrieval plan from De-SpecBridge/Retrival_plan.md in Spec-RAG: molecule library with precomputed embeddings, spectrum→embedding mappers, and three retrieval/generation variants (A: SMI-TED only, B: ChemBERTa retrieval + SMI-TED generation, C: ChemBERTa only).

Scripts

Script	Purpose
scripts/build_library.py	Build library: compute v_smi (E_smi), v_chem (E_chem), meta.parquet (SMILES, formula, mass). SMI-TED optional via --despecbridge-path or DESPECBRIDGE_PATH.
scripts/build_faiss.py	Build FAISS indices from vectors_smi.npy and vectors_chem.npy → index_smi.faiss, index_chem.faiss.
scripts/train_mapper.py	Train M_smi and M_chem on MassSpecGym train MGF: E_mist(spec) → t_smi, t_chem. Saves mappers.pt.
scripts/retrieve_generate.py	Run Variant A/B/C: query MGF → candidates JSONL.
scripts/evaluate_massspecgym.py	Compute Recall@1/10/50 (and optional Tanimoto@1) from candidates JSONL.
scripts/evaluate_oracle_retrieval.py	Oracle: encode smiles_gt with ChemBERTa/SMI-TED, search library; report Recall@1/10/50, Tanimoto@1.
scripts/self_retrieval_sanity_check.py	Self-retrieval: index test molecules only, query with same embeddings; expect Recall@1 ≈ 1.0 (ChemBERTa and SMI-TED separately).
scripts/self_retrieval_mapped_sanity_check.py	Mapped self-retrieval: index = spectrum→mapper embeddings of test set, query = same; expect Recall@1 ≈ 1.0 (ChemBERTa-mapped and SMI-TED-mapped).
scripts/sanity_check_true_index_mapped_query.py	True-index + Mapped-query: index = true mol embeddings (test SMILES), query = spectrum→mapper embeddings; report Recall@1/10/50. Tests mapper alignment of spectrum to molecule space.

Artifacts

• Library dir: vectors_smi.npy, vectors_chem.npy, meta.parquet, index_smi.faiss, index_chem.faiss
• Mapper dir: mappers.pt (M_smi, M_chem state_dicts + d_spec, d_smi, d_chem)

Example

# 1) Build library (SMILES list, e.g. from PubChem or MassSpecGym unique SMILES)
python scripts/build_library.py --smiles-path /path/to/smiles.txt --out-dir /path/to/library --despecbridge-path /path/to/De-SpecBridge

# 2) Build FAISS indices
python scripts/build_faiss.py --library-dir /path/to/library

# 3) Train mappers (MassSpecGym train MGF)
python scripts/train_mapper.py --mgf-path /path/to/MassSpecGym_train.mgf --specbridge-ckpt /path/to/specbridge.pt --out-dir /path/to/mappers --despecbridge-path /path/to/De-SpecBridge

# 4) Retrieve (e.g. Variant B, K=100)
python scripts/retrieve_generate.py --mgf-path /path/to/MassSpecGym_test.mgf --library-dir /path/to/library --mapper-dir /path/to/mappers --specbridge-ckpt /path/to/specbridge.pt --variant B --K 100 --out-jsonl /path/to/candidates_b.jsonl

# 5) Evaluate
python scripts/evaluate_massspecgym.py --pred-jsonl /path/to/candidates_b.jsonl --report /path/to/metrics.json --tanimoto

Self-retrieval sanity check (Priority 2)

Build a tiny FAISS index from test molecules only; query with the same true molecule embeddings. Expected Recall@1 ≈ 1.0 for both ChemBERTa and SMI-TED.

• If either fails → that embedding/index pipeline is broken.
• If ChemBERTa passes and SMI-TED fails → SMI-TED embedding construction is the issue.
• If both pass → main issue is likely library coverage / eval mismatch.

Mapped-embedding self-retrieval (self_retrieval_mapped_sanity_check.py): index = spectrum→mapper embeddings (test MGF), query = same. Tests spectrum→ChemBERTa-mapped and spectrum→SMI-TED-mapped. If either fails, the spectrum→mapped-embedding pipeline is broken.

True-index + Mapped-query (sanity_check_true_index_mapped_query.py): index = true molecule embeddings (ChemBERTa/SMI-TED of test SMILES), query = mapped embeddings (spectrum→mapper). Measures how well the mapper aligns spectrum to molecule space (Recall@1/10/50). Low recall here with high oracle recall suggests the mapper or spectrum representation is the bottleneck.

Dependencies

• SMI-TED (E_smi): set DESPECBRIDGE_PATH or --despecbridge-path to De-SpecBridge repo so despecbridge.models.smited_decoder.load_smited can be imported.
• E_mist: SpecBridge checkpoint (DreaMS adapter) via --specbridge-ckpt (and optional --dreams-ckpt).