Skip to content

REQUIREMENTS.md

Latest commit

 

History

History
144 lines (107 loc) · 4.6 KB

REQUIREMENTS.md

File metadata and controls

144 lines (107 loc) · 4.6 KB
Raw

ChemLogic Requirements

Formal requirements specification. For introduction, see README. For API details, see SPEC.

Functional Requirements

FR1: Molecular Property Prediction

ID Requirement Priority
FR1.1 Predict binary molecular properties Must
FR1.2 Predict continuous molecular properties (regression) Must
FR1.3 Predict multi-class molecular properties Must
FR1.4 Predict multiple outputs simultaneously (multi-regression) Must
FR1.5 Accept SMILES strings as input Must
FR1.6 Auto-detect task type from label structure Should
FR1.7 Support batch prediction Should

FR2: Model Configuration

ID Requirement Priority
FR2.1 Support multiple GNN architectures Must
FR2.2 Configurable message-passing layers Must
FR2.3 Configurable embedding dimensions Must
FR2.4 Support architecture-specific parameters Must

FR3: Knowledge Base Integration

ID Requirement Priority
FR3.1 Encode functional groups as learnable logical rules Must
FR3.2 Selective enabling of functional group categories Must
FR3.3 Encode subgraph patterns as learnable rules Must
FR3.4 Selective enabling of subgraph pattern types Must
FR3.5 Three integration modes (BARE, CCE, CCD) Must

FR4: Training

ID Requirement Priority
FR4.1 Configurable train/test split Must
FR4.2 Configurable learning rate and epochs Must
FR4.3 Early stopping Must
FR4.4 Report loss and evaluation metrics Must
FR4.5 AUROC for classification/multi-class, R² for regression/multi-regression Must

FR5: Datasets

ID Requirement Priority
FR5.1 Support TUD benchmark datasets Must
FR5.2 Support TDC ADMET datasets Must
FR5.3 Accept custom SMILES datasets Must
FR5.4 Automatic SMILES to relational conversion Must

FR6: Interpretability

ID Requirement Priority
FR6.1 Visualize learned templates with weights Must
FR6.2 Funnel mode for scalar weights Should
FR6.3 Weights traceable to chemical concepts Should

FR7: Inference

ID Requirement Priority
FR7.1 Inference on new SMILES after training Must
FR7.2 Return prediction scores Must

FR8: Extended Features

ID Requirement Priority
FR8.1 Atom-level RDKit features as node predicates Must
FR8.2 Bond-level RDKit features as edge predicates Must
FR8.3 Graph-level scalar features via synthetic node or broadcast Must

FR9: Checkpointing

ID Requirement Priority
FR9.1 Save trained model weights to disk Must
FR9.2 Load weights and resume inference without retraining Must
FR9.3 Optionally save checkpoint during training at fixed intervals Should

Non-Functional Requirements

NFR1: Performance

ID Requirement Target
NFR1.1 MUTAG training completes within 5 minutes Target
NFR1.2 Batched dataset building for large datasets Must
NFR1.3 Inference under 1 second per molecule Target

NFR2: Usability

ID Requirement Target
NFR2.1 Basic pipeline in <10 lines of code Must
NFR2.2 Sensible parameter defaults Must
NFR2.3 Auto-detect task and num_outputs from label structure Should
NFR2.4 Clear error messages Should

NFR3: Compatibility

ID Requirement Target
NFR3.1 Python >=3.11 Must
NFR3.2 Any OS with Java 1.8+ Must
NFR3.3 Installable via pip Must

NFR4: Extensibility

ID Requirement Target
NFR4.1 New models via Model extension Must
NFR4.2 New functional groups via KnowledgeBase extension Must
NFR4.3 New datasets via Dataset extension Must

NFR5: Reliability

ID Requirement Target
NFR5.1 Input validation with clear errors Must
NFR5.2 Graceful handling of invalid SMILES Should
NFR5.3 Reproducible training with fixed seed Must

Constraints

ID Constraint
C1 Built on PyNeuraLogic—bound by its capabilities
C2 Requires Java runtime
C3 Some patterns require explicit hydrogens