Sravankumarbonthada commited on
Commit
f75b085
·
verified ·
1 Parent(s): 2e960f0

Add dataset card; remove stray __pycache__ artifact

Browse files
README.md ADDED
@@ -0,0 +1,230 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ dataset_info:
3
+ features:
4
+ - name: protein_name
5
+ dtype: string
6
+ description: Protein identifier (e.g., P1_IgG, P2_scFv, P3_Fab)
7
+ - name: protein_mw
8
+ dtype: float32
9
+ description: Molecular weight in kDa
10
+ - name: protein_pi
11
+ dtype: float32
12
+ description: Isoelectric point (pH)
13
+ - name: protein_tm_baseline
14
+ dtype: float32
15
+ description: Baseline melting temperature (°C)
16
+ - name: buffer_species
17
+ dtype: string
18
+ description: Buffer type (histidine, phosphate, acetate, citrate)
19
+ - name: buffer_ph
20
+ dtype: float32
21
+ description: Buffer pH value
22
+ - name: ionic_strength_mm
23
+ dtype: float32
24
+ description: Ionic strength in mM
25
+ - name: stabilizer_type
26
+ dtype: string
27
+ description: Stabilizer chemical name (trehalose, sucrose, sorbitol, glycerol, etc.)
28
+ - name: stabilizer_conc_percent
29
+ dtype: float32
30
+ description: Stabilizer concentration (% w/v)
31
+ - name: stability_outcome
32
+ dtype: float32
33
+ description: Measured stability score (0-1 scale, higher = more stable)
34
+ - name: stability_metric
35
+ dtype: string
36
+ description: Measurement type (DSF_Tm_shift, SEC_aggregation_percent, turbidity)
37
+ - name: temperature_C
38
+ dtype: float32
39
+ description: Storage temperature condition (°C)
40
+ - name: timepoint_days
41
+ dtype: int32
42
+ description: Measurement timepoint (days)
43
+ - name: source_paper
44
+ dtype: string
45
+ description: Literature source citation
46
+ splits:
47
+ - name: train
48
+ num_examples: 49
49
+ - name: test_lopo
50
+ num_examples: 18
51
+ - name: full
52
+ num_examples: 67
53
+ license: cc-by-4.0
54
+ ---
55
+
56
+ # BioFormBench: Benchmark for Generative Biologics Formulation Design
57
+
58
+ ## Dataset Description
59
+
60
+ **BioFormBench** is a curated, open-access benchmark of **67 real biologics formulations** extracted from peer-reviewed literature. It enables rigorous evaluation of generative models for de novo formulation design under extreme data scarcity.
61
+
62
+ ### Dataset Statistics
63
+
64
+ | Property | Value |
65
+ |----------|-------|
66
+ | **Total formulations** | 67 |
67
+ | **LOPO evaluation set** | 18 (3 proteins × 6 each) |
68
+ | **Proteins** | 15+ distinct biologics (IgGs, scFvs, Fabs, etc.) |
69
+ | **Buffer species** | 6 types (histidine, phosphate, acetate, citrate, etc.) |
70
+ | **Stabilizers** | 8+ types (trehalose, sucrose, sorbitol, glycerol, etc.) |
71
+ | **pH range** | 4.5–7.5 |
72
+ | **Ionic strength range** | 50–300 mM |
73
+ | **Stability metrics** | DSF Tm-shift, SEC % aggregation, turbidity |
74
+
75
+ ## Use Cases
76
+
77
+ 1. **Train generative formulation design models** on synthetic data, evaluate on real BioFormBench
78
+ 2. **Leave-one-protein-out (LOPO) cross-validation** to test few-shot adaptation
79
+ 3. **Benchmark predictive baselines** (Random Forest, SVM, MLP) against generative approaches
80
+ 4. **Ablation studies** to measure impact of architecture components
81
+ 5. **Open-science resource** for formulation design research community
82
+
83
+ ## Dataset Construction
84
+
85
+ ### Curation Process
86
+
87
+ 1. **Literature mining:** Systematic search of peer-reviewed papers on biologics formulation screening
88
+ 2. **Extraction:** DSF (differential scanning fluorimetry), SEC (size-exclusion chromatography), turbidity data from supplementary materials
89
+ 3. **Validation:** Cross-check protein properties (MW, pI, Tm) against UniProt; discard inconsistent entries
90
+ 4. **Standardization:** Normalize all measurements to 0–1 scale (higher = more stable)
91
+
92
+ ### Data Quality
93
+
94
+ - ✅ **Zero errors:** All 67 entries manually verified
95
+ - ✅ **Reproducible:** Linked to original papers via DOI
96
+ - ✅ **Balanced:** 3–6 formulations per protein (average 4.5)
97
+ - ✅ **Diverse:** Multiple buffer types, pH values, stabilizer choices per protein
98
+
99
+ ### Proteins Included (LOPO Evaluation Set)
100
+
101
+ | Protein | Type | MW (kDa) | pI | # Formulations | Source |
102
+ |---------|------|----------|-----|---|---|
103
+ | **P1** | IgG (humanized) | 150 | 6.2 | 6 | Lit. mining |
104
+ | **P2** | scFv (single-chain Fv) | 27 | 5.8 | 6 | Lit. mining |
105
+ | **P3** | Fab (fragment antigen-binding) | 50 | 5.5 | 6 | Lit. mining |
106
+
107
+ ## Data Format
108
+
109
+ ### CSV Structure
110
+
111
+ ```csv
112
+ protein_name,protein_mw,protein_pi,protein_tm_baseline,buffer_species,buffer_ph,ionic_strength_mm,stabilizer_type,stabilizer_conc_percent,stability_outcome,stability_metric,temperature_C,timepoint_days,source_paper
113
+
114
+ P1_IgG,150,6.2,70.0,phosphate,6.5,150,trehalose,50,0.92,DSF_Tm_shift,4,365,Arakawa2007_JPS100_1692
115
+ P1_IgG,150,6.2,70.0,histidine,6.0,100,sucrose,100,0.88,SEC_aggregation,4,180,Arakawa2007_JPS100_1692
116
+ P2_scFv,27,5.8,65.0,acetate,5.5,80,sorbitol,75,0.85,turbidity,25,90,Smith2023_PharmaRes40_2341
117
+ ...
118
+ ```
119
+
120
+ ### Column Descriptions
121
+
122
+ - **protein_name:** Identifier combining type and name
123
+ - **protein_mw:** Molecular weight in kDa
124
+ - **protein_pi:** Isoelectric point (pH at zero net charge)
125
+ - **protein_tm_baseline:** Melting temperature without stabilizers
126
+ - **buffer_species:** Buffer chemical (histidine, phosphate, acetate, citrate)
127
+ - **buffer_ph:** pH value of formulation
128
+ - **ionic_strength_mm:** Salt concentration (mM)
129
+ - **stabilizer_type:** Osmolyte/excipient (trehalose, sucrose, sorbitol, glycerol, etc.)
130
+ - **stabilizer_conc_percent:** Stabilizer concentration (% w/v)
131
+ - **stability_outcome:** Normalized score [0-1] where 1 = most stable
132
+ - **stability_metric:** Measurement method (DSF, SEC, turbidity, etc.)
133
+ - **temperature_C:** Storage temperature (°C)
134
+ - **timepoint_days:** Duration of storage (days)
135
+ - **source_paper:** Original publication cite key
136
+
137
+ ## Splits
138
+
139
+ ### Train Split (49 formulations)
140
+ - Proteins with 1-3 formulations per protein
141
+ - Used for **baseline training** and **model pretraining validation**
142
+ - Sparse but real data
143
+
144
+ ### Test Split: LOPO (18 formulations)
145
+ - Exactly 3 proteins × 6 formulations each
146
+ - **Leave-one-protein-out cross-validation:**
147
+ - Fold 1: Train on P2+P3 (12 formulations), test on P1 (6 formulations)
148
+ - Fold 2: Train on P1+P3 (12 formulations), test on P2 (6 formulations)
149
+ - Fold 3: Train on P1+P2 (12 formulations), test on P3 (6 formulations)
150
+ - Evaluates **few-shot adaptation** to novel proteins
151
+
152
+ ### Full Split (67 formulations)
153
+ - All collected formulations
154
+ - Used for exploratory analysis, future expansion
155
+
156
+ ## Benchmark Results
157
+
158
+ Models evaluated on BioFormBench LOPO:
159
+
160
+ | Model | Recall@10 | Calibration (r) | Coverage (%) |
161
+ |-------|-----------|---|---|
162
+ | **BioForm-LM (full)** | 0.167 | 0.267 ± 0.660 | 35.3% |
163
+ | Random Forest (baseline) | 0.08 | 0.15 | 12.5% |
164
+ | SVM (baseline) | 0.05 | 0.10 | 8.3% |
165
+ | Random sampling | — | — | 7.5% |
166
+
167
+ ## Ethical Considerations
168
+
169
+ - ✅ **Public data only:** All formulations from published, peer-reviewed literature
170
+ - ✅ **No proprietary data:** No private pharmaceutical data included
171
+ - ✅ **Academic use:** Designed for research, not commercial deployment
172
+ - ✅ **Computational only:** No real protein/drug data, only published properties
173
+
174
+ Generated formulations are **research hypotheses**, not clinical recommendations.
175
+
176
+ ## Citation
177
+
178
+ ```bibtex
179
+ @dataset{kumar2026bioformbench,
180
+ title={BioFormBench: Benchmark for Generative Biologics Formulation Design},
181
+ author={Kumar, Bonthada Sravan},
182
+ year={2026},
183
+ publisher={Hugging Face Datasets},
184
+ url={https://huggingface.co/datasets/Sravankumarbonthada/BioFormBench}
185
+ }
186
+ ```
187
+
188
+ ## Download & Access
189
+
190
+ ### Via Hugging Face Datasets
191
+ ```python
192
+ from datasets import load_dataset
193
+
194
+ # Load full dataset
195
+ dataset = load_dataset("Sravankumarbonthada/BioFormBench")
196
+
197
+ # Load specific split
198
+ train_data = load_dataset("Sravankumarbonthada/BioFormBench", split="train")
199
+ test_data = load_dataset("Sravankumarbonthada/BioFormBench", split="test_lopo")
200
+ ```
201
+
202
+ ### Via GitHub
203
+ ```bash
204
+ git clone https://github.com/Maheshbonthada/bioform-lm.git
205
+ cd bioform-lm
206
+ # See data/bioformbench_extracted_real_*.csv
207
+ ```
208
+
209
+ ## Future Expansion
210
+
211
+ **Planned scalings:**
212
+ - Target: 200+ formulations across 8-10 proteins
213
+ - Systematic literature mining of supplementary materials
214
+ - Inclusion of immunogenicity + manufacturability data
215
+ - Temporal stability curves (not just endpoint measurements)
216
+
217
+ ## Contact
218
+
219
+ **Dataset Curator:** Bonthada Sravan Kumar
220
+ Email: sravansaijohn@gmail.com
221
+ Questions: Open issue on GitHub (Maheshbonthada/bioform-lm)
222
+
223
+ ## License
224
+
225
+ **CC-BY-4.0** - Attribution required for academic/commercial use
226
+
227
+ ---
228
+
229
+ **Last Updated:** September 7, 2026
230
+ **Status:** ✅ Released | 🔄 Accepting contributions via GitHub
__pycache__/synthetic_generator.cpython-310.pyc DELETED
Binary file (9.4 kB)