proteintensor 0.2.0__tar.gz → 0.3.0__tar.gz

{proteintensor-0.2.0 → proteintensor-0.3.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: proteintensor
-Version: 0.2.0
+Version: 0.3.0
 Summary: AI-native biomolecular tensor format for structural biology ML
 Author-email: Clayton Moore <claytonwaynemoore@gmail.com>
 License-Expression: MIT
@@ -34,6 +34,8 @@ Provides-Extra: cloud
 Requires-Dist: fsspec>=2023.1; extra == "cloud"
 Requires-Dist: s3fs>=2023.1; extra == "cloud"
 Requires-Dist: gcsfs>=2023.1; extra == "cloud"
+Provides-Extra: ligands
+Requires-Dist: rdkit>=2023.3; extra == "ligands"
 Provides-Extra: dev
 Requires-Dist: pytest>=7; extra == "dev"
 Requires-Dist: pytest-benchmark; extra == "dev"
@@ -41,7 +43,9 @@ Requires-Dist: pytest-cov; extra == "dev"
 Requires-Dist: fsspec>=2023.1; extra == "dev"
 Dynamic: license-file
 
-# HelixDB / ProteinTensor
+![ProteinTensor - AI-native protein data format: convert structure or sequence into cached tensors](assets/banner.png)
+
+# ProteinTensor Introduction
 
 **ProteinTensor** is an AI-native biomolecular storage format designed to eliminate
 the preprocessing bottleneck in modern structural biology machine learning pipelines.
@@ -108,20 +112,21 @@ performance format that turns a recurring computational tax into a one-time cost
 
 ## Benchmark: Traditional Pipeline vs ProteinTensor
 
-All timings are median over 30 rounds on an NVIDIA RTX 5080, CUDA 12.8, Python 3.11.
-Proteins span the full range from a 76-residue domain to a 3,525-residue CRISPR enzyme.
-Run `python boltz_benchmark.py` to reproduce.
+All timings are median over 30 rounds on a Windows workstation (RTX 5080, Python
+3.11.9); mmCIF parsing and `.ptt` reads are CPU-bound, so these reflect CPU
+performance. Proteins span the full range from a 76-residue domain to a
+3,525-residue CRISPR enzyme. Run `python boltz_benchmark.py` to reproduce.
 
 ### Per-structure load times
 
 | Structure | Method | Res | MSA seqs | mmCIF parse | ptt: full | ptt: backbone | ptt: bonds | ptt: MSA | ptt: dist mx |
 |---|---|---|---|---|---|---|---|---|---|
-| 1UBQ - Ubiquitin | X-ray | 76 | 512 | 7.2 ms | 2.8 ms | 1.2 ms | 0.7 ms | 1.6 ms | 0.8 ms |
-| 6LU7 - SARS-CoV-2 Mpro | X-ray | 312 | 1,024 | 29.6 ms | 2.9 ms | 1.2 ms | 0.7 ms | 5.1 ms | 2.0 ms |
-| 4HHB - Hemoglobin | X-ray | 574 | 2,048 | 55.3 ms | 2.9 ms | 1.2 ms | 0.7 ms | 11.3 ms | 3.5 ms |
-| 6M0J - ACE2 + RBD | Cryo-EM | 791 | 2,048 | 74.7 ms | 2.9 ms | 1.2 ms | 0.7 ms | 14.7 ms | 6.4 ms |
-| 6VXX - Spike trimer | Cryo-EM | 2,916 | 8,192 | 283.4 ms | 3.3 ms | 1.3 ms | 0.9 ms | 208.3 ms | 71.1 ms |
-| 6OHW - Cas12a | Cryo-EM | 3,525 | 8,192 | 352.4 ms | 3.3 ms | 1.2 ms | 1.0 ms | 240.7 ms | 104.5 ms |
+| 1UBQ - Ubiquitin | X-ray | 76 | 512 | 7.4 ms | 3.2 ms | 1.3 ms | 0.8 ms | 1.8 ms | 0.8 ms |
+| 6LU7 - SARS-CoV-2 Mpro | X-ray | 312 | 1,024 | 28.7 ms | 3.3 ms | 1.3 ms | 0.8 ms | 5.2 ms | 1.9 ms |
+| 4HHB - Hemoglobin | X-ray | 574 | 2,048 | 54.1 ms | 3.3 ms | 1.3 ms | 0.8 ms | 11.5 ms | 3.6 ms |
+| 6M0J - ACE2 + RBD | Cryo-EM | 791 | 2,048 | 73.2 ms | 3.3 ms | 1.4 ms | 0.8 ms | 15.3 ms | 6.9 ms |
+| 6VXX - Spike trimer | Cryo-EM | 2,916 | 8,192 | 283.9 ms | 3.7 ms | 1.4 ms | 1.0 ms | 213.7 ms | 74.7 ms |
+| 6OHW - Cas12a | Cryo-EM | 3,525 | 8,192 | 346.5 ms | 3.7 ms | 1.3 ms | 1.0 ms | 243.9 ms | 107.3 ms |
 
 **Column definitions**
 - `ptt: full` - `read()` - all atoms, backbone, bonds, metadata
@@ -134,32 +139,42 @@ Run `python boltz_benchmark.py` to reproduce.
 
 | Structure | Res | full | backbone | bonds | MSA | dist mx |
 |---|---|---|---|---|---|---|
-| 1UBQ - Ubiquitin | 76 | 3x | 6x | 11x | 4x | 9x |
-| 6LU7 - SARS-CoV-2 Mpro | 312 | 10x | 24x | 43x | 6x | 15x |
-| 4HHB - Hemoglobin | 574 | 19x | 45x | 78x | 5x | 16x |
-| 6M0J - ACE2 + RBD | 791 | 26x | 61x | 102x | 5x | 12x |
-| 6VXX - Spike trimer | 2,916 | 87x | 223x | 308x | 1x* | 4x |
-| 6OHW - Cas12a | 3,525 | 108x | 284x | 370x | 1x* | 3x |
+| 1UBQ - Ubiquitin | 76 | 2x | 6x | 10x | 4x | 9x |
+| 6LU7 - SARS-CoV-2 Mpro | 312 | 9x | 21x | 38x | 5x | 15x |
+| 4HHB - Hemoglobin | 574 | 17x | 40x | 70x | 5x | 15x |
+| 6M0J - ACE2 + RBD | 791 | 22x | 54x | 92x | 5x | 11x |
+| 6VXX - Spike trimer | 2,916 | 76x | 201x | 285x | 1x* | 4x |
+| 6OHW - Cas12a | 3,525 | 95x | 257x | 343x | 1x* | 3x |
 
 *MSA speedup shown as 1x vs mmCIF parse because both are in the same time range for
 large proteins - the real MSA comparison is vs JackHMMER generation (see below).
 
 ### Feature assembly: time to prepare all tensors for model.forward()
 
-Traditional = mmCIF parse + read MSA from A3M file. ProteinTensor = single .ptt read
-with all features pre-cached (sequence, backbone, bonds, MSA, distance matrix,
-ESM2 embedding).
-
-| Structure | Res | Traditional | ProteinTensor | Speedup |
-|---|---|---|---|---|
-| 1UBQ - Ubiquitin | 76 | 22.7 ms | 5.2 ms | 4x |
-| 6LU7 - SARS-CoV-2 Mpro | 312 | 157.3 ms | 9.9 ms | 16x |
-| 4HHB - Hemoglobin | 574 | 525.5 ms | 17.7 ms | 30x |
-| 6M0J - ACE2 + RBD | 791 | 722.7 ms | 23.9 ms | 30x |
-| 6VXX - Spike trimer | 2,916 | 9,838.5 ms | 282.7 ms | 35x |
-| 6OHW - Cas12a | 3,525 | 11,903.1 ms | 348.4 ms | **34x** |
-
-Average speedup across all six structures: **34x** for full feature assembly.
+Traditional = mmCIF parse + A3M MSA parse + distance-matrix compute. ProteinTensor
+= read the structure, MSA, distance matrix, and ESM2 embedding from a single
+pre-cached `.ptt`. Reproduce with `python benchmarks/assembly_benchmark.py`
+(MSA depth and embedding shape are realistic; numeric content is synthetic, so
+timing reflects tensor dimensions, not values).
+
+| Structure | Res | MSA depth | Traditional | ProteinTensor | Speedup |
+|---|---|---|---|---|---|
+| 1UBQ - Ubiquitin | 76 | 512 | 14.1 ms | 7.1 ms | 2.0x |
+| 6LU7 - SARS-CoV-2 Mpro | 312 | 1,024 | 48.7 ms | 13.6 ms | 3.6x |
+| 4HHB - Hemoglobin | 574 | 2,048 | 118.0 ms | 22.7 ms | 5.2x |
+| 6M0J - ACE2 + RBD | 791 | 2,048 | 196.4 ms | 38.3 ms | 5.1x |
+| 6VXX - Spike trimer | 2,916 | 8,192 | 1,395 ms | 309 ms | 4.5x |
+| 6OHW - Cas12a | 3,525 | 8,192 | 1,462 ms | 381 ms | 3.8x |
+
+Average speedup across all six structures: **4x** for full feature assembly
+(measured on a Windows CPU box - see
+[`benchmarks/ASSEMBLY_RESULTS.md`](benchmarks/ASSEMBLY_RESULTS.md)).
+
+> **On an earlier 34x figure:** prior versions reported ~34x here. That number was
+> measured against ProteinTensor's original scalar A3M parser, which dominated the
+> traditional side (~11 s to parse an 8,192-deep MSA). Vectorizing that parser in
+> v0.2.0 cut the traditional baseline ~8x, so the *fair* feature-assembly speedup
+> is now ~4x. The `.ptt` read side was unchanged - only the baseline got faster.
 
 ### Drug target benchmark
 
@@ -169,21 +184,21 @@ IgG1 antibody. Numbers are consistent with the structural biology benchmark abov
 
 | Target | Res | mmCIF parse | ptt: full | ptt: backbone | ptt: bonds | ptt: MSA | ptt: dist mx |
 |---|---|---|---|---|---|---|---|
-| 6OIM - KRAS G12C + Sotorasib | 167 | 16.6 ms | 2.8 ms | 1.2 ms | 0.7 ms | 2.8 ms | 1.1 ms |
-| 3HTB - HIV-1 protease | 163 | 16.0 ms | 2.8 ms | 1.2 ms | 0.7 ms | 2.7 ms | 1.1 ms |
-| 5WT9 - PD-L1 checkpoint | 533 | 53.8 ms | 2.9 ms | 1.2 ms | 0.7 ms | 13.1 ms | 3.3 ms |
-| 1TUP - p53 tumor suppressor | 585 | 56.5 ms | 2.8 ms | 1.2 ms | 0.7 ms | 12.4 ms | 3.4 ms |
-| 2P4E - PCSK9 | 586 | 54.7 ms | 2.8 ms | 1.2 ms | 0.7 ms | 12.1 ms | 3.4 ms |
-| 1IGT - IgG1 antibody | 1,316 | 123.4 ms | 2.9 ms | 1.2 ms | 0.8 ms | 46.8 ms | 16.4 ms |
+| 6OIM - KRAS G12C + Sotorasib | 167 | 17.1 ms | 3.4 ms | 1.3 ms | 0.8 ms | 3.0 ms | 1.3 ms |
+| 3HTB - HIV-1 protease | 163 | 16.5 ms | 3.3 ms | 1.4 ms | 0.8 ms | 2.8 ms | 1.3 ms |
+| 5WT9 - PD-L1 checkpoint | 533 | 54.8 ms | 3.8 ms | 1.4 ms | 0.8 ms | 11.9 ms | 3.8 ms |
+| 1TUP - p53 tumor suppressor | 585 | 57.4 ms | 3.4 ms | 1.4 ms | 0.8 ms | 13.0 ms | 4.0 ms |
+| 2P4E - PCSK9 | 586 | 55.4 ms | 3.4 ms | 1.4 ms | 0.8 ms | 12.8 ms | 4.1 ms |
+| 1IGT - IgG1 antibody | 1,316 | 127.3 ms | 3.5 ms | 1.4 ms | 0.8 ms | 47.1 ms | 17.9 ms |
 
 | Target | Res | full | backbone | bonds | MSA | dist mx |
 |---|---|---|---|---|---|---|
-| 6OIM - KRAS G12C + Sotorasib | 167 | 6x | 14x | 24x | 6x | 15x |
-| 3HTB - HIV-1 protease | 163 | 6x | 14x | 23x | 6x | 14x |
-| 5WT9 - PD-L1 checkpoint | 533 | 19x | 44x | 77x | 4x | 16x |
-| 1TUP - p53 tumor suppressor | 585 | 20x | 47x | 80x | 5x | 17x |
-| 2P4E - PCSK9 | 586 | 19x | 46x | 77x | 5x | 16x |
-| 1IGT - IgG1 antibody | 1,316 | 42x | **100x** | **162x** | 3x | 8x |
+| 6OIM - KRAS G12C + Sotorasib | 167 | 5x | 13x | 22x | 6x | 13x |
+| 3HTB - HIV-1 protease | 163 | 5x | 12x | 21x | 6x | 13x |
+| 5WT9 - PD-L1 checkpoint | 533 | 15x | 40x | 69x | 5x | 14x |
+| 1TUP - p53 tumor suppressor | 585 | 17x | 42x | 71x | 4x | 14x |
+| 2P4E - PCSK9 | 586 | 16x | 41x | 70x | 4x | 14x |
+| 1IGT - IgG1 antibody | 1,316 | 37x | **92x** | **156x** | 3x | 7x |
 
 ### DataLoader batch throughput
 
@@ -192,26 +207,38 @@ padded batches ready for `model.forward()`. Single process, no prefetch workers.
 
 | Batch size | ms / batch | Structures / sec |
 |---|---|---|
-| 1 | 0.01 ms | 88,106 |
-| 4 | 0.04 ms | 108,696 |
-| 8 | 0.37 ms | 21,707 |
-| 16 | 0.95 ms | 16,783 |
-| 32 | 2.0 ms | **15,854** |
+| 1 | 0.01 ms | 97,088 |
+| 4 | 0.03 ms | 116,279 |
+| 8 | 0.42 ms | 19,242 |
+| 16 | 0.97 ms | 16,412 |
+| 32 | 2.1 ms | **15,033** |
 
 ### Scale projection: 100,000 structures, one training epoch
 
+These are **projections**, extrapolated from the measured per-structure timings
+above - not end-to-end measurements at 100k scale.
+
 | Operation | Traditional pipeline | ProteinTensor | Speedup |
 |---|---|---|---|
-| Structure load (parse mmCIF each epoch) | 3.7 hours | 5 min | **45x** |
-| Backbone-only load (template search) | 3.7 hours | 2 min | **109x** |
-| Full feature assembly (seq + MSA + pairs + emb) | 4.5 days | 3.2 hours | **34x** |
-| MSA generation (JackHMMER, 32-core CPU, once) | 4,000 hours | 2.2 hours | **1,794x** |
+| Structure load (parse mmCIF each epoch) | 3.8 hours | 6 min | **37x** |
+| Backbone-only load (template search) | 3.8 hours | 2 min | **95x** |
+| Full feature assembly (seq + MSA + pairs + emb) | 16 hours | 3.9 hours | **4x** |
+| MSA generation (JackHMMER, 32-core CPU, once) | 4,000 hours | 2.7 hours | **1,477x** |
 
 > MSA generation assumes 2.4 min/protein on a 32-core server (PDB90 database, standard
 > AlphaFold settings). ProteinTensor generates MSAs once and loads from the `.ptt` cache
 > on every subsequent run. The 4,000-hour figure is the real cost AlphaFold2 and Boltz
 > users pay to build training datasets from scratch.
 
+> **Measured vs projected - read this.** The **1,477x** above is MSA *generation*
+> (building the alignment once with JackHMMER) and is a **literature-based
+> projection**, not something benchmarked here. What *is* measured on hardware is
+> the recurring per-epoch MSA **load** - reading a cached MSA from `.ptt` vs
+> re-parsing A3M text each epoch (against a vectorized A3M parser baseline):
+> **3.4x-5.9x**, growing with MSA depth. See
+> [`benchmarks/MSA_RESULTS.md`](benchmarks/MSA_RESULTS.md). These are different
+> quantities; do not read the 1,477x as a measured load speedup.
+
 ### Disk tradeoff
 
 A full-featured `.ptt` (8,192-sequence MSA + distance matrix + ESM2-650M embedding at
@@ -267,6 +294,18 @@ pt.write(data, "ubq.ptt")
 data = pt.from_fasta("complex.fasta")
 ```
 
+### Batch-convert a directory
+
+Convert an entire directory of structures in parallel, with progress reporting.
+Files that fail to parse are skipped and listed in the summary; already-converted
+outputs are skipped by default.
+
+```bash
+proteintensor convert-dir ./pdb_files/ ./ptt_files/            # auto worker count
+proteintensor convert-dir ./pdb_files/ ./ptt_files/ --workers 16 --recursive
+proteintensor convert-dir ./pdb_files/ ./ptt_files/ --overwrite  # rebuild existing
+```
+
 ### Benchmark against mmCIF
 
 ```bash
@@ -351,6 +390,20 @@ pt.add_pair_feature("1abc.ptt", my_array, name="template_pair",
 emb = pt.read_embedding("1abc.ptt", "esm2_t33_650M_UR50D")
 emb.data.shape      # (N_res, 1280)  float32  (upcast from float16 on load)
 
+# ------ Ligands / small molecules ------
+# Capture drugs, cofactors, and ions from a structure (opt-in)
+data = pt.from_mmcif("6oim.cif", include_ligands=True)
+[l.name for l in data.ligands]        # ['MG', 'GDP', 'MOV']  (MOV = sotorasib)
+
+ligs = pt.read_ligands("6oim.ptt")
+ligs[0].elements                      # (N_atoms,)  S2  element symbols
+ligs[0].positions                     # (N_atoms, 3)  float32
+pt.list_ligands("6oim.ptt")           # ['MG', 'GDP', 'MOV']
+
+# Build a ligand from SMILES (needs `pip install "proteintensor[ligands]"`)
+aspirin = pt.from_smiles("CC(=O)Oc1ccccc1C(=O)O", name="AIN")
+pt.add_ligand("target.ptt", aspirin)  # attach to an existing .ptt
+
 # ------ Lazy / zero-copy access ------
 positions = pt.mmap_positions("1abc.ptt")       # zarr.Array - no full load
 backbone  = pt.mmap_backbone("1abc.ptt")        # [N_res, 4, 3]
@@ -392,6 +445,7 @@ data = pt.read(
 )
 
 # ------ Multi-structure dataset ------
+# Structure .ptt files and sequence-only .ptt files can be mixed in one dataset.
 pt.create_dataset("training.ptt")
 for ptt_file in Path("ptt_files").glob("*.ptt"):
     pt.add_to_dataset("training.ptt", ptt_file)
@@ -407,8 +461,13 @@ loader = DataLoader(ds, batch_size=8, collate_fn=pt.ProteinDataset.collate)
 for batch in loader:
     coords  = torch.from_numpy(batch["atom_positions"])   # (B, max_atoms, 3)
     pad     = torch.from_numpy(batch["padding_mask"])     # (B, max_res)  True=real
+    has_str = torch.from_numpy(batch["has_structure"])    # (B,)  False = sequence-only
 ```
 
+Sequence-only entries contribute zero atoms to the batch (`n_atoms == 0`,
+`has_structure == False`), so sequence-driven and structure-based samples can be
+loaded together in one `DataLoader`.
+
 ---
 
 ## .ptt file layout
@@ -445,10 +504,16 @@ structure.ptt/                      Zarr directory store (v0.7)
 │   └── <name>/                              one sub-group per named feature
 │       ├── .zattrs                          channels, symmetric, dtype, description
 │       └── data           [N_res, N_res, C] any dtype, chunked 128x128xC
-└── embeddings/
-    └── <model>/                             one sub-group per PLM model
-        ├── .zattrs                          model, layer, dim, dtype, seq SHA-256
-        └── data           [N_res, D]        float32 or float16, chunked 256xD
+├── embeddings/
+│   └── <model>/                             one sub-group per PLM model
+│       ├── .zattrs                          model, layer, dim, dtype, seq SHA-256
+│       └── data           [N_res, D]        float32 or float16, chunked 256xD
+└── ligands/
+    └── <index>/                            one sub-group per non-polymer ligand
+        ├── .zattrs                          name (CCD), chain_id, res_num, smiles
+        ├── elements       [N_atoms]         S2       element symbols
+        ├── positions      [N_atoms, 3]      float32  Angstrom coordinates
+        └── b_factors      [N_atoms]         float32
 ```
 
 ### Multi-structure dataset layout
@@ -486,9 +551,9 @@ Each sub-group under `structures/` is identical to a standalone `.ptt` root, so
 pytest tests/ -v
 ```
 
-106 tests across structure roundtrip, backbone/bonds/MSA/pairs/embeddings,
-A3M parsing, Boltz adapter, multi-structure dataset, and cloud streaming
-(memory:// fsspec - no real cloud account required).
+150 tests across structure roundtrip, backbone/bonds/MSA/pairs/embeddings/ligands,
+sequence conversion, A3M parsing, Boltz adapter, multi-structure dataset, and cloud
+streaming (memory:// fsspec - no real cloud account required).
 
 ---
 
@@ -509,11 +574,11 @@ A3M parsing, Boltz adapter, multi-structure dataset, and cloud streaming
 - [ ] Chai-1 adapter
 
 **Data pipeline**
-- [ ] Batch convert CLI - convert entire PDB directories in parallel with progress reporting
+- [x] Batch convert CLI - convert entire PDB directories in parallel with progress reporting
 - [ ] Sequence-identity dataset splitting - MMseqs2-based cluster splits to prevent data leakage between train / val / test
 
 **Format extensions**
-- [ ] Ligand / small-molecule support - SMILES, CCD-based atom graphs, binding site annotations for drug-protein interaction models
+- [x] Ligand / small-molecule support - CCD-based extraction from structures, SMILES input via RDKit, element/coordinate storage (bond graphs and binding-site annotations still to come)
 - [ ] MD trajectory storage - time axis `[N_frames, N_atoms, 3]` for conformational ensembles and AlphaFold 3 diffusion trajectories
 
 **Performance**

{proteintensor-0.2.0 → proteintensor-0.3.0}/README.md

@@ -1,4 +1,6 @@
-# HelixDB / ProteinTensor
+![ProteinTensor - AI-native protein data format: convert structure or sequence into cached tensors](assets/banner.png)
+
+# ProteinTensor Introduction
 
 **ProteinTensor** is an AI-native biomolecular storage format designed to eliminate
 the preprocessing bottleneck in modern structural biology machine learning pipelines.
@@ -65,20 +67,21 @@ performance format that turns a recurring computational tax into a one-time cost
 
 ## Benchmark: Traditional Pipeline vs ProteinTensor
 
-All timings are median over 30 rounds on an NVIDIA RTX 5080, CUDA 12.8, Python 3.11.
-Proteins span the full range from a 76-residue domain to a 3,525-residue CRISPR enzyme.
-Run `python boltz_benchmark.py` to reproduce.
+All timings are median over 30 rounds on a Windows workstation (RTX 5080, Python
+3.11.9); mmCIF parsing and `.ptt` reads are CPU-bound, so these reflect CPU
+performance. Proteins span the full range from a 76-residue domain to a
+3,525-residue CRISPR enzyme. Run `python boltz_benchmark.py` to reproduce.
 
 ### Per-structure load times
 
 | Structure | Method | Res | MSA seqs | mmCIF parse | ptt: full | ptt: backbone | ptt: bonds | ptt: MSA | ptt: dist mx |
 |---|---|---|---|---|---|---|---|---|---|
-| 1UBQ - Ubiquitin | X-ray | 76 | 512 | 7.2 ms | 2.8 ms | 1.2 ms | 0.7 ms | 1.6 ms | 0.8 ms |
-| 6LU7 - SARS-CoV-2 Mpro | X-ray | 312 | 1,024 | 29.6 ms | 2.9 ms | 1.2 ms | 0.7 ms | 5.1 ms | 2.0 ms |
-| 4HHB - Hemoglobin | X-ray | 574 | 2,048 | 55.3 ms | 2.9 ms | 1.2 ms | 0.7 ms | 11.3 ms | 3.5 ms |
-| 6M0J - ACE2 + RBD | Cryo-EM | 791 | 2,048 | 74.7 ms | 2.9 ms | 1.2 ms | 0.7 ms | 14.7 ms | 6.4 ms |
-| 6VXX - Spike trimer | Cryo-EM | 2,916 | 8,192 | 283.4 ms | 3.3 ms | 1.3 ms | 0.9 ms | 208.3 ms | 71.1 ms |
-| 6OHW - Cas12a | Cryo-EM | 3,525 | 8,192 | 352.4 ms | 3.3 ms | 1.2 ms | 1.0 ms | 240.7 ms | 104.5 ms |
+| 1UBQ - Ubiquitin | X-ray | 76 | 512 | 7.4 ms | 3.2 ms | 1.3 ms | 0.8 ms | 1.8 ms | 0.8 ms |
+| 6LU7 - SARS-CoV-2 Mpro | X-ray | 312 | 1,024 | 28.7 ms | 3.3 ms | 1.3 ms | 0.8 ms | 5.2 ms | 1.9 ms |
+| 4HHB - Hemoglobin | X-ray | 574 | 2,048 | 54.1 ms | 3.3 ms | 1.3 ms | 0.8 ms | 11.5 ms | 3.6 ms |
+| 6M0J - ACE2 + RBD | Cryo-EM | 791 | 2,048 | 73.2 ms | 3.3 ms | 1.4 ms | 0.8 ms | 15.3 ms | 6.9 ms |
+| 6VXX - Spike trimer | Cryo-EM | 2,916 | 8,192 | 283.9 ms | 3.7 ms | 1.4 ms | 1.0 ms | 213.7 ms | 74.7 ms |
+| 6OHW - Cas12a | Cryo-EM | 3,525 | 8,192 | 346.5 ms | 3.7 ms | 1.3 ms | 1.0 ms | 243.9 ms | 107.3 ms |
 
 **Column definitions**
 - `ptt: full` - `read()` - all atoms, backbone, bonds, metadata
@@ -91,32 +94,42 @@ Run `python boltz_benchmark.py` to reproduce.
 
 | Structure | Res | full | backbone | bonds | MSA | dist mx |
 |---|---|---|---|---|---|---|
-| 1UBQ - Ubiquitin | 76 | 3x | 6x | 11x | 4x | 9x |
-| 6LU7 - SARS-CoV-2 Mpro | 312 | 10x | 24x | 43x | 6x | 15x |
-| 4HHB - Hemoglobin | 574 | 19x | 45x | 78x | 5x | 16x |
-| 6M0J - ACE2 + RBD | 791 | 26x | 61x | 102x | 5x | 12x |
-| 6VXX - Spike trimer | 2,916 | 87x | 223x | 308x | 1x* | 4x |
-| 6OHW - Cas12a | 3,525 | 108x | 284x | 370x | 1x* | 3x |
+| 1UBQ - Ubiquitin | 76 | 2x | 6x | 10x | 4x | 9x |
+| 6LU7 - SARS-CoV-2 Mpro | 312 | 9x | 21x | 38x | 5x | 15x |
+| 4HHB - Hemoglobin | 574 | 17x | 40x | 70x | 5x | 15x |
+| 6M0J - ACE2 + RBD | 791 | 22x | 54x | 92x | 5x | 11x |
+| 6VXX - Spike trimer | 2,916 | 76x | 201x | 285x | 1x* | 4x |
+| 6OHW - Cas12a | 3,525 | 95x | 257x | 343x | 1x* | 3x |
 
 *MSA speedup shown as 1x vs mmCIF parse because both are in the same time range for
 large proteins - the real MSA comparison is vs JackHMMER generation (see below).
 
 ### Feature assembly: time to prepare all tensors for model.forward()
 
-Traditional = mmCIF parse + read MSA from A3M file. ProteinTensor = single .ptt read
-with all features pre-cached (sequence, backbone, bonds, MSA, distance matrix,
-ESM2 embedding).
-
-| Structure | Res | Traditional | ProteinTensor | Speedup |
-|---|---|---|---|---|
-| 1UBQ - Ubiquitin | 76 | 22.7 ms | 5.2 ms | 4x |
-| 6LU7 - SARS-CoV-2 Mpro | 312 | 157.3 ms | 9.9 ms | 16x |
-| 4HHB - Hemoglobin | 574 | 525.5 ms | 17.7 ms | 30x |
-| 6M0J - ACE2 + RBD | 791 | 722.7 ms | 23.9 ms | 30x |
-| 6VXX - Spike trimer | 2,916 | 9,838.5 ms | 282.7 ms | 35x |
-| 6OHW - Cas12a | 3,525 | 11,903.1 ms | 348.4 ms | **34x** |
-
-Average speedup across all six structures: **34x** for full feature assembly.
+Traditional = mmCIF parse + A3M MSA parse + distance-matrix compute. ProteinTensor
+= read the structure, MSA, distance matrix, and ESM2 embedding from a single
+pre-cached `.ptt`. Reproduce with `python benchmarks/assembly_benchmark.py`
+(MSA depth and embedding shape are realistic; numeric content is synthetic, so
+timing reflects tensor dimensions, not values).
+
+| Structure | Res | MSA depth | Traditional | ProteinTensor | Speedup |
+|---|---|---|---|---|---|
+| 1UBQ - Ubiquitin | 76 | 512 | 14.1 ms | 7.1 ms | 2.0x |
+| 6LU7 - SARS-CoV-2 Mpro | 312 | 1,024 | 48.7 ms | 13.6 ms | 3.6x |
+| 4HHB - Hemoglobin | 574 | 2,048 | 118.0 ms | 22.7 ms | 5.2x |
+| 6M0J - ACE2 + RBD | 791 | 2,048 | 196.4 ms | 38.3 ms | 5.1x |
+| 6VXX - Spike trimer | 2,916 | 8,192 | 1,395 ms | 309 ms | 4.5x |
+| 6OHW - Cas12a | 3,525 | 8,192 | 1,462 ms | 381 ms | 3.8x |
+
+Average speedup across all six structures: **4x** for full feature assembly
+(measured on a Windows CPU box - see
+[`benchmarks/ASSEMBLY_RESULTS.md`](benchmarks/ASSEMBLY_RESULTS.md)).
+
+> **On an earlier 34x figure:** prior versions reported ~34x here. That number was
+> measured against ProteinTensor's original scalar A3M parser, which dominated the
+> traditional side (~11 s to parse an 8,192-deep MSA). Vectorizing that parser in
+> v0.2.0 cut the traditional baseline ~8x, so the *fair* feature-assembly speedup
+> is now ~4x. The `.ptt` read side was unchanged - only the baseline got faster.
 
 ### Drug target benchmark
 
@@ -126,21 +139,21 @@ IgG1 antibody. Numbers are consistent with the structural biology benchmark abov
 
 | Target | Res | mmCIF parse | ptt: full | ptt: backbone | ptt: bonds | ptt: MSA | ptt: dist mx |
 |---|---|---|---|---|---|---|---|
-| 6OIM - KRAS G12C + Sotorasib | 167 | 16.6 ms | 2.8 ms | 1.2 ms | 0.7 ms | 2.8 ms | 1.1 ms |
-| 3HTB - HIV-1 protease | 163 | 16.0 ms | 2.8 ms | 1.2 ms | 0.7 ms | 2.7 ms | 1.1 ms |
-| 5WT9 - PD-L1 checkpoint | 533 | 53.8 ms | 2.9 ms | 1.2 ms | 0.7 ms | 13.1 ms | 3.3 ms |
-| 1TUP - p53 tumor suppressor | 585 | 56.5 ms | 2.8 ms | 1.2 ms | 0.7 ms | 12.4 ms | 3.4 ms |
-| 2P4E - PCSK9 | 586 | 54.7 ms | 2.8 ms | 1.2 ms | 0.7 ms | 12.1 ms | 3.4 ms |
-| 1IGT - IgG1 antibody | 1,316 | 123.4 ms | 2.9 ms | 1.2 ms | 0.8 ms | 46.8 ms | 16.4 ms |
+| 6OIM - KRAS G12C + Sotorasib | 167 | 17.1 ms | 3.4 ms | 1.3 ms | 0.8 ms | 3.0 ms | 1.3 ms |
+| 3HTB - HIV-1 protease | 163 | 16.5 ms | 3.3 ms | 1.4 ms | 0.8 ms | 2.8 ms | 1.3 ms |
+| 5WT9 - PD-L1 checkpoint | 533 | 54.8 ms | 3.8 ms | 1.4 ms | 0.8 ms | 11.9 ms | 3.8 ms |
+| 1TUP - p53 tumor suppressor | 585 | 57.4 ms | 3.4 ms | 1.4 ms | 0.8 ms | 13.0 ms | 4.0 ms |
+| 2P4E - PCSK9 | 586 | 55.4 ms | 3.4 ms | 1.4 ms | 0.8 ms | 12.8 ms | 4.1 ms |
+| 1IGT - IgG1 antibody | 1,316 | 127.3 ms | 3.5 ms | 1.4 ms | 0.8 ms | 47.1 ms | 17.9 ms |
 
 | Target | Res | full | backbone | bonds | MSA | dist mx |
 |---|---|---|---|---|---|---|
-| 6OIM - KRAS G12C + Sotorasib | 167 | 6x | 14x | 24x | 6x | 15x |
-| 3HTB - HIV-1 protease | 163 | 6x | 14x | 23x | 6x | 14x |
-| 5WT9 - PD-L1 checkpoint | 533 | 19x | 44x | 77x | 4x | 16x |
-| 1TUP - p53 tumor suppressor | 585 | 20x | 47x | 80x | 5x | 17x |
-| 2P4E - PCSK9 | 586 | 19x | 46x | 77x | 5x | 16x |
-| 1IGT - IgG1 antibody | 1,316 | 42x | **100x** | **162x** | 3x | 8x |
+| 6OIM - KRAS G12C + Sotorasib | 167 | 5x | 13x | 22x | 6x | 13x |
+| 3HTB - HIV-1 protease | 163 | 5x | 12x | 21x | 6x | 13x |
+| 5WT9 - PD-L1 checkpoint | 533 | 15x | 40x | 69x | 5x | 14x |
+| 1TUP - p53 tumor suppressor | 585 | 17x | 42x | 71x | 4x | 14x |
+| 2P4E - PCSK9 | 586 | 16x | 41x | 70x | 4x | 14x |
+| 1IGT - IgG1 antibody | 1,316 | 37x | **92x** | **156x** | 3x | 7x |
 
 ### DataLoader batch throughput
 
@@ -149,26 +162,38 @@ padded batches ready for `model.forward()`. Single process, no prefetch workers.
 
 | Batch size | ms / batch | Structures / sec |
 |---|---|---|
-| 1 | 0.01 ms | 88,106 |
-| 4 | 0.04 ms | 108,696 |
-| 8 | 0.37 ms | 21,707 |
-| 16 | 0.95 ms | 16,783 |
-| 32 | 2.0 ms | **15,854** |
+| 1 | 0.01 ms | 97,088 |
+| 4 | 0.03 ms | 116,279 |
+| 8 | 0.42 ms | 19,242 |
+| 16 | 0.97 ms | 16,412 |
+| 32 | 2.1 ms | **15,033** |
 
 ### Scale projection: 100,000 structures, one training epoch
 
+These are **projections**, extrapolated from the measured per-structure timings
+above - not end-to-end measurements at 100k scale.
+
 | Operation | Traditional pipeline | ProteinTensor | Speedup |
 |---|---|---|---|
-| Structure load (parse mmCIF each epoch) | 3.7 hours | 5 min | **45x** |
-| Backbone-only load (template search) | 3.7 hours | 2 min | **109x** |
-| Full feature assembly (seq + MSA + pairs + emb) | 4.5 days | 3.2 hours | **34x** |
-| MSA generation (JackHMMER, 32-core CPU, once) | 4,000 hours | 2.2 hours | **1,794x** |
+| Structure load (parse mmCIF each epoch) | 3.8 hours | 6 min | **37x** |
+| Backbone-only load (template search) | 3.8 hours | 2 min | **95x** |
+| Full feature assembly (seq + MSA + pairs + emb) | 16 hours | 3.9 hours | **4x** |
+| MSA generation (JackHMMER, 32-core CPU, once) | 4,000 hours | 2.7 hours | **1,477x** |
 
 > MSA generation assumes 2.4 min/protein on a 32-core server (PDB90 database, standard
 > AlphaFold settings). ProteinTensor generates MSAs once and loads from the `.ptt` cache
 > on every subsequent run. The 4,000-hour figure is the real cost AlphaFold2 and Boltz
 > users pay to build training datasets from scratch.
 
+> **Measured vs projected - read this.** The **1,477x** above is MSA *generation*
+> (building the alignment once with JackHMMER) and is a **literature-based
+> projection**, not something benchmarked here. What *is* measured on hardware is
+> the recurring per-epoch MSA **load** - reading a cached MSA from `.ptt` vs
+> re-parsing A3M text each epoch (against a vectorized A3M parser baseline):
+> **3.4x-5.9x**, growing with MSA depth. See
+> [`benchmarks/MSA_RESULTS.md`](benchmarks/MSA_RESULTS.md). These are different
+> quantities; do not read the 1,477x as a measured load speedup.
+
 ### Disk tradeoff
 
 A full-featured `.ptt` (8,192-sequence MSA + distance matrix + ESM2-650M embedding at
@@ -224,6 +249,18 @@ pt.write(data, "ubq.ptt")
 data = pt.from_fasta("complex.fasta")
 ```
 
+### Batch-convert a directory
+
+Convert an entire directory of structures in parallel, with progress reporting.
+Files that fail to parse are skipped and listed in the summary; already-converted
+outputs are skipped by default.
+
+```bash
+proteintensor convert-dir ./pdb_files/ ./ptt_files/            # auto worker count
+proteintensor convert-dir ./pdb_files/ ./ptt_files/ --workers 16 --recursive
+proteintensor convert-dir ./pdb_files/ ./ptt_files/ --overwrite  # rebuild existing
+```
+
 ### Benchmark against mmCIF
 
 ```bash
@@ -308,6 +345,20 @@ pt.add_pair_feature("1abc.ptt", my_array, name="template_pair",
 emb = pt.read_embedding("1abc.ptt", "esm2_t33_650M_UR50D")
 emb.data.shape      # (N_res, 1280)  float32  (upcast from float16 on load)
 
+# ------ Ligands / small molecules ------
+# Capture drugs, cofactors, and ions from a structure (opt-in)
+data = pt.from_mmcif("6oim.cif", include_ligands=True)
+[l.name for l in data.ligands]        # ['MG', 'GDP', 'MOV']  (MOV = sotorasib)
+
+ligs = pt.read_ligands("6oim.ptt")
+ligs[0].elements                      # (N_atoms,)  S2  element symbols
+ligs[0].positions                     # (N_atoms, 3)  float32
+pt.list_ligands("6oim.ptt")           # ['MG', 'GDP', 'MOV']
+
+# Build a ligand from SMILES (needs `pip install "proteintensor[ligands]"`)
+aspirin = pt.from_smiles("CC(=O)Oc1ccccc1C(=O)O", name="AIN")
+pt.add_ligand("target.ptt", aspirin)  # attach to an existing .ptt
+
 # ------ Lazy / zero-copy access ------
 positions = pt.mmap_positions("1abc.ptt")       # zarr.Array - no full load
 backbone  = pt.mmap_backbone("1abc.ptt")        # [N_res, 4, 3]
@@ -349,6 +400,7 @@ data = pt.read(
 )
 
 # ------ Multi-structure dataset ------
+# Structure .ptt files and sequence-only .ptt files can be mixed in one dataset.
 pt.create_dataset("training.ptt")
 for ptt_file in Path("ptt_files").glob("*.ptt"):
     pt.add_to_dataset("training.ptt", ptt_file)
@@ -364,8 +416,13 @@ loader = DataLoader(ds, batch_size=8, collate_fn=pt.ProteinDataset.collate)
 for batch in loader:
     coords  = torch.from_numpy(batch["atom_positions"])   # (B, max_atoms, 3)
     pad     = torch.from_numpy(batch["padding_mask"])     # (B, max_res)  True=real
+    has_str = torch.from_numpy(batch["has_structure"])    # (B,)  False = sequence-only
 ```
 
+Sequence-only entries contribute zero atoms to the batch (`n_atoms == 0`,
+`has_structure == False`), so sequence-driven and structure-based samples can be
+loaded together in one `DataLoader`.
+
 ---
 
 ## .ptt file layout
@@ -402,10 +459,16 @@ structure.ptt/                      Zarr directory store (v0.7)
 │   └── <name>/                              one sub-group per named feature
 │       ├── .zattrs                          channels, symmetric, dtype, description
 │       └── data           [N_res, N_res, C] any dtype, chunked 128x128xC
-└── embeddings/
-    └── <model>/                             one sub-group per PLM model
-        ├── .zattrs                          model, layer, dim, dtype, seq SHA-256
-        └── data           [N_res, D]        float32 or float16, chunked 256xD
+├── embeddings/
+│   └── <model>/                             one sub-group per PLM model
+│       ├── .zattrs                          model, layer, dim, dtype, seq SHA-256
+│       └── data           [N_res, D]        float32 or float16, chunked 256xD
+└── ligands/
+    └── <index>/                            one sub-group per non-polymer ligand
+        ├── .zattrs                          name (CCD), chain_id, res_num, smiles
+        ├── elements       [N_atoms]         S2       element symbols
+        ├── positions      [N_atoms, 3]      float32  Angstrom coordinates
+        └── b_factors      [N_atoms]         float32
 ```
 
 ### Multi-structure dataset layout
@@ -443,9 +506,9 @@ Each sub-group under `structures/` is identical to a standalone `.ptt` root, so
 pytest tests/ -v
 ```
 
-106 tests across structure roundtrip, backbone/bonds/MSA/pairs/embeddings,
-A3M parsing, Boltz adapter, multi-structure dataset, and cloud streaming
-(memory:// fsspec - no real cloud account required).
+150 tests across structure roundtrip, backbone/bonds/MSA/pairs/embeddings/ligands,
+sequence conversion, A3M parsing, Boltz adapter, multi-structure dataset, and cloud
+streaming (memory:// fsspec - no real cloud account required).
 
 ---
 
@@ -466,11 +529,11 @@ A3M parsing, Boltz adapter, multi-structure dataset, and cloud streaming
 - [ ] Chai-1 adapter
 
 **Data pipeline**
-- [ ] Batch convert CLI - convert entire PDB directories in parallel with progress reporting
+- [x] Batch convert CLI - convert entire PDB directories in parallel with progress reporting
 - [ ] Sequence-identity dataset splitting - MMseqs2-based cluster splits to prevent data leakage between train / val / test
 
 **Format extensions**
-- [ ] Ligand / small-molecule support - SMILES, CCD-based atom graphs, binding site annotations for drug-protein interaction models
+- [x] Ligand / small-molecule support - CCD-based extraction from structures, SMILES input via RDKit, element/coordinate storage (bond graphs and binding-site annotations still to come)
 - [ ] MD trajectory storage - time axis `[N_frames, N_atoms, 3]` for conformational ensembles and AlphaFold 3 diffusion trajectories
 
 **Performance**

{proteintensor-0.2.0 → proteintensor-0.3.0}/proteintensor/__init__.py

@@ -34,8 +34,10 @@ from .bonds import (
 from .dataset import ProteinDataset, create_dataset, add_to_dataset
 from .remote import consolidate
 from .converters import from_mmcif, from_sequence, from_fasta, parse_fasta
+from .ligands import read_ligands, list_ligands, add_ligand, from_smiles
+from .schema import LigandData
 
-__version__ = "0.2.0"
+__version__ = "0.3.0"
 
 __all__ = [
     # Converters - input
@@ -51,6 +53,8 @@ __all__ = [
     "compute_and_store_distances", "compute_and_store_contacts",
     # I/O - embeddings
     "read_embedding", "add_embedding", "list_embeddings", "mmap_embedding",
+    # Ligands / small molecules
+    "read_ligands", "list_ligands", "add_ligand", "from_smiles", "LigandData",
     # Data containers
     "ProteinTensorData", "BackboneData", "BondData", "MsaData", "PairFeature", "EmbeddingData",
     # MSA utilities