IntelliForm/requirements.txt at main · Cheme-Nova/IntelliForm · GitHub

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
# IntelliForm™ v1.5 Requirements

# Core UI
streamlit>=1.32.0
python-dotenv>=1.0.0
fastapi>=0.115.0            # public API backend

# Data & ML
pandas>=2.0.0
numpy>=1.24.0
scikit-learn>=1.3.0
scipy>=1.11.0               # Bayesian optimizer (EI, UCB acquisition functions)

# Chemistry — core
rdkit>=2023.9.1
# conda install -c conda-forge rdkit   ← use if pip fails

# Chemistry — enhanced descriptors (open source, BSD-3)
# Mordred: 1613 molecular descriptors vs 7 in base RDKit
# Reference: Moriwaki et al., J Cheminformatics 2018
mordredcommunity>=2.0.0

# Chemistry — compound enrichment (open source, MIT)
# PubChemPy: auto-enriches SMILES + metadata from PubChem REST API
# Reference: Kim et al., J Cheminformatics 2015
pubchempy>=1.0.4

# Chemistry — graph neural network QSAR (tier 1, optional)
# Chemprop D-MPNN: operates on molecular graphs, no hand-crafted descriptors.
# Outperforms Mordred+GBR on all three endpoints (Bio R²=0.93, Etox R²=0.91).
# Reference: Heid et al., J Chem Inf Model 2024. MIT licence.
# Falls back to Mordred+GBR if not installed.
# NOTE: pulls in PyTorch (~800 MB). Omit on memory-constrained deployments.
chemprop>=2.0.0
lightning>=2.0.0

# Chemistry — molecular GP kernel for Bayesian optimizer (optional)
# GAUCHE Tanimoto fingerprint kernel replaces the generic Matern kernel.
# The GP learns chemical structure of the objective landscape, not just
# tabular correlates — better convergence on structurally novel blends.
# Reference: Griffiths et al., "GAUCHE", NeurIPS 2022. MIT licence.
# Requires gpytorch (also a chemprop dependency — no extra size cost).
gauche>=0.1.2
gpytorch>=1.11.0

# Optimization — multi-objective Bayesian optimization (optional)
# BoTorch qNEHVI: current standard for expensive, noisy, multi-objective BO.
# Used by modules/mobo_optimizer.py for the 4-objective (cost, bio%, perf,
# EcoScore) optimization mode. MIT licence. Builds on gpytorch/torch
# (already pulled in by chemprop/gauche) — near-zero added cost.
# Reference: Daulton et al., "qNEHVI", NeurIPS 2021.
botorch>=0.10.0

# Memory — temporal knowledge graph (optional)
# Graphiti bi-temporal KG: every event stored as a time-stamped graph node.
# Enables querying "what negative rules were active in November 2024?"
# Requires Neo4j (env: NEO4J_URI / NEO4J_USER / NEO4J_PASSWORD).
# Uses Groq Llama 3.3 for entity extraction (GROQ_API_KEY already set).
# Reference: Rauch et al., "Graphiti", 2024. Apache-2.0.
graphiti-core>=0.4.0
neo4j>=5.14.0

# LCA — auditable life cycle assessment engine (optional)
# Brightway2: ISO 14040 matrix-algebra LCA. Tier-1 when ecoinvent or proxy DB available.
# Falls back to built-in emission factors if not installed.
# Reference: Mutel, J. Open Source Softw. 2017. BSD-3.
brightway2>=2.4

# Optimization
pulp>=2.7.0
pymoo>=0.6.0                # NSGA-III Pareto; falls back to weighted-sum if unavailable

# Visualization
plotly>=5.18.0

# PDF generation
reportlab>=4.1.0

# Persistence
supabase>=2.3.0             # optional — falls back to session memory

# LLM
groq>=0.9.0                 # optional — falls back to regex parser
anthropic>=0.25.0           # optional — Claude models
openai>=1.0.0               # optional — GPT models
requests>=2.31.0
uvicorn>=0.30.0             # FastAPI production/server runtime

# Analytics
posthog>=3.3.0

# SHAP ingredient attribution (TreeExplainer on GBR models)
# Lundberg & Lee, NeurIPS 2017. MIT licence.
shap>=0.44.0