--- license: cc-by-4.0 --- # BRAIDBERTa BRAIDBERTa is the first pretrained language model for BRAID, a molecular notation in which every valid string decodes to a valid molecule. Compared against a matched SMILES baseline on BBBP, BRAID achieves statistically indistinguishable ROC-AUC while guaranteeing valid decoding. Check it out: https://huggingface.co/spaces/aakothari/Conformer --- ## The BRAID notation BRAID (**B**ranch-counted, **R**elative-ring, **A**ttachment-native, **I**nvalid-free, **D**ense) is built from four mechanisms, all of which are recombined prior art: | Mechanism | Syntax | Borrowed from | |---|---|---| | Organic-subset lexing, `[...]` only when needed | `C`, `Cl`, `[N+]` | SMILES | | Length-counted branches (no parentheses) | `>k` then *k* atoms | SELFIES branch counts | | Relative ring closure (no paired digits) | `^d` = bond *d* atoms back | DeepSMILES relative rings | | Fragment separator, global index continues across it | `.` | SAFE | Only **atom tokens** advance the global index; `>k`, `^d`, bond symbols and `.` do not. ``` aspirin SMILES CC(=O)Oc1ccccc1C(=O)O BRAID CC>1=OOcccccc^5C>1=OO (aromatic mode) BRAID CC>1=OOC=CC=CC=C^5C>1=OO (kekulé mode, default) ``` ### Novelty: none claimed BRAID is **not novel at the mechanism level.** Relative ring closure is DeepSMILES (O'Boyle & Dalke, 2018). Length-counted branches are the SELFIES branch-count idea (Krenn et al., 2020) with the brackets stripped. Guaranteed-valid decoding via a valence state machine is the entire point of SELFIES. The fragment separator is SAFE (Noutahi et al., 2023). Please treat BRAID as a recombination. ### The property BRAID does have Every string decodes. On 10,000 random and mutated strings, the decoder produced a sanitizable molecule with **0 crashes**. This is achieved by making the decoder *total*: unparseable tokens and over-valent bonds become silent no-ops. Validity is guaranteed. Under an order-*k* Markov generator on a 79-molecule corpus: | representation | valid% | unique% | novel% | vocab | tok/mol | |---|--:|--:|--:|--:|--:| | SMILES | 40.0 | 38.4 | 34.0 | 24 | 13.6 | | DeepSMILES | 35.4 | 46.9 | 42.4 | 27 | 13.3 | | SELFIES | 100.0 | 59.6 | 57.9 | 29 | 12.9 | | **BRAID** | **100.0** | 60.9 | 59.2 | 30 | 13.2 | A Markov model is far weaker than a real CLM, so treat validity% as the meaningful (model-independent) signal and uniqueness/novelty as illustrative. --- ## Model details - **Architecture:** RoBERTa (masked language modelling) - **Pretraining corpus:** ZINC 100k, encoded in BRAID - **Tokenizer:** BPE trained on the BRAID corpus - **Objective:** MLM --- ## Downstream evaluation: BBBP Finetuned for binary classification (blood–brain barrier penetration) with a fresh 2-logit head and cross-entropy loss. Hyperparameters (learning rate, epochs) were selected by Optuna on the validation split, then the best configuration was retrained across 5 seeds. Metric is test ROC-AUC. | Model | Notation | BBBP ROC-AUC | Seed 0 | Seed 1 | Seed 2 | Seed 3 | Seed 4 | |---|---|---|---|---|---|---|---| | **BRAIDBERTa-v9** | BRAID | **0.725 ± 0.017** | 0.727 | 0.710 | 0.751 | 0.710 | 0.727 | | *SMILES control* | SMILES | *0.720 ± 0.010* | 0.719 | 0.738 | 0.715 | 0.714 | 0.712 | (± is the sample standard deviation across seeds.) ### Interpretation **The two are statistically indistinguishable.** Welch's *t*-test on the five seeds per model gives *t* ≈ 0.61, *p* ≈ 0.56. The 0.005 gap in favour of BRAID sits well inside seed noise, and the per-seed ranges overlap almost completely (BRAID 0.710–0.751, control 0.712–0.738). The correct reading is > Under matched pretraining and finetuning, BRAID's guaranteed-validity property comes > **at no measurable cost to downstream predictive performance** on BBBP. Note: The seed-to-seed spread (±0.017) exceeds the between-model difference. Any comparison at this scale reporting a single seed is measuring noise. --- ## Usage The model consumes **BRAID strings, not SMILES**, so you need the codec: ```bash pip install rdkit transformers pip install git+https://github.com/AayushK-othari/braid.git ``` ```python from braids import smiles_to_braid from transformers import AutoTokenizer, AutoModelForSequenceClassification tok = AutoTokenizer.from_pretrained("aakothari/BRAIDBERTa-v9") model = AutoModelForSequenceClassification.from_pretrained( "aakothari/BRAIDBERTa-v9", num_labels=2 ) braid = smiles_to_braid("CC(=O)Oc1ccccc1C(=O)O") # -> CC>1=OOC=CC=CC=C^5C>1=OO inputs = tok(braid, return_tensors="pt", truncation=True, max_length=128) logits = model(**inputs).logits ``` > **Do not feed raw SMILES to this tokenizer.** It will not error — BPE always backs off to > characters — it will simply produce a meaningless tokenization and a plausible-looking, > wrong prediction. Always encode with `smiles_to_braid` first. **Encode in the same mode the model was pretrained in.** The codec has Kekulé (default) and aromatic modes, and they produce different strings for the same molecule (`C=CC=CC=C^5` vs `cccccc^5`). Mixing modes between pretraining and inference silently degrades performance. The reference implementation — encoder, decoder, valence state machine, stereo handling, tokenizer/`Vocab` builder, and the test suites — lives at **[github.com/aakothari/braids](https://github.com/aakothari/braids)**. ### Verified codec behaviour - **Round-trip:** 42/42 molecules recover their constitution in both Kekulé and aromatic modes — including fused/bridged systems (naphthalene, indole, adamantane, caffeine, purine, quinoline, nicotine) and ions (nitromethane, acetate, ammonium, Na⁺·Cl⁻). - **Stereo:** 26/26 chiral / E-Z molecules round-trip in both modes — L-alanine, (S)-ibuprofen, meso vs. L-tartaric acid, menthol, L-DOPA, adrenaline, carvone, fumaric/maleic acid, conjugated dienes, mixed chiral + E/Z. - **Validity:** 10,000/10,000 random and mutated strings decode to a sanitizable molecule with 0 crashes. --- ## Limitations 1. **Stereochemistry is partial.** Tetrahedral chirality and E/Z double bonds round-trip on 26/26 test cases. Allene/axial/planar chirality, atropisomers, and non-tetrahedral stereo centres fall back to unspecified. 2. **The valence state machine is approximate.** A hand-rolled charge→valence rule, correct for common ions but liable to mis-clamp hypervalent, organometallic, or unusual-charge atoms. It is not a substitute for RDKit's model. 3. **Insertion/deletion is non-local.** Inserting or deleting an atom shifts every `^d` that spans the edit point. Inherent to relative back-references; applies equally to DeepSMILES. 4. **Not a canonical hash.** One molecule has many valid BRAID strings. The encoder uses RDKit canonical-rank rooting for determinism, but uniqueness is not proven. --- ## Citation BRAID is a recombination of published mechanisms. Please cite the underlying work: - O'Boyle, N. & Dalke, A. *DeepSMILES: An Adaptation of SMILES for Use in Machine-Learning of Chemical Structures.* ChemRxiv, 2018. - Krenn, M. et al. *Self-Referencing Embedded Strings (SELFIES): A 100% robust molecular string representation.* Machine Learning: Science and Technology, 2020. - Noutahi, E. et al. *Gotta be SAFE: A New Framework for Molecular Design.* 2023. - Weininger, D. *SMILES, a chemical language and information system.* J. Chem. Inf. Comput. Sci., 1988.