MSNet-Sirius-1.260626

For NEGATIVE mode
-Tune 1 to 4 and launch.

For POSITIVE mode
-Tune 1 to 4 and launch

For BOTH mode integration
-Tune 1 to 3 in both mode, launch separatly
-Tune 5 and 6 then launch

-Samples in pos and neg mode MUST have the same names or share the same index

-If fail, launch pos and neg mode separatly

MS-Net Input: MZmine/MSdial & SIRIUS

Launch Report PDF to write charts in the output result folder

SOME WARNINGS

• MS-Net use MS2 data. Features without MS/MS spectra are discarded

• A column named "Class" must be present in input height/area table defining sample category/class(Control, WT, QC...)

• MS-Net use pearson correlation among samples in many processing steps. As a consequence, MSnet do not handle one sample only.

• MS-Net need to seed the Mass Spectral Similarity network with confident annotation (level 1 or 2). Weencourage using mass spectral library matching during data processing (FragHub, GNPS2, MassBank, ...).

• The size of input chemical space (number of feature x topK hit) will drastically influence processing time

• If pos and neg mode must be merged, launch node 1 to 3 in one mode before the other to not overload your processor.

Before You Start

- [ ] KNIME 5.2+ installed with RDKit extension

- [ ] All software sites enabled in KNIME preferences

- [ ] Memory allocation configured (leave only 8GB for system)

- [ ] All other programs closed (MS-Net is memory-intensive)

- [ ] Directory structure organized (separate folders for each export)

- [ ] POS and NEG files have matching names/order (if merging modes)

🔄 Typical Workflow Pipeline

Raw Data (mzML/raw)

↓

MZmine/MS-DIAL (Feature Detection)

↓

Sirius-CSI (In Silico Annotation)

↓

MS-Net (Network-Based Prioritization)

↓

Results + Optional Cytoscape Visualization

📁 Directory Organization

Project_Folder/

├── raws_pos/ # LC-MS raw files (positive mode)

├── raws_neg/ # LC-MS raw files (negative mode)

├── MZmine_pos/ # ⚠️ ONE export per folder!

│ ├── [prefix]_full_feature_table.csv

│ ├── [prefix]_edges_ms2_cosine.csv

│ ├── [prefix]_annotations.csv

│ └── [prefix]_sirius.mgf

├── MZmine_neg/ # Same structure for negative mode

├── Sirius_pos/ # ⚠️ ONE export per folder!

│ └── structure_identifications_all.csv and all other sirius outputs

├── Sirius_neg/ # Same for negative mode

└── MSNet-POSNEG/ # Results folder (workflow output)

Module 1: Data Import

# DO NOT LEAVE EMPTY - use "xxxxx" if no value

target_genus: "Cannabis" # Seed elevation criterion

target_family: "Cannabaceae" # Seed elevation criterion

topK_insilicoDB: 50 # Balance: speed vs coverage

inhouse_DBmatch:10 #look for level 3a only in first top n ranked in silico matches

chromato_mode: "_C18pos" # For POS/NEG merging

IIN_filtering: TRUE # Keep most common adducts only

Module 2: Network Population

cosine_threshold: 0.70 # MSS network cutoff (0.6-0.8)

deltaRT_threshold: 8.0 # RT window (minutes) treshold between tow nodes

apply_RT_clustering: TRUE # ⏱️ slow! Disable for short gradient acquisition (<12 min)

RT_bucket_size: 0.01 # For UHPLC: 0.0015

Module 3: Annotation

fingerprint_type: "PubChem" # PubChem/Morgan/RDKit

topN_link_score: 5 # Neighbor links (5-20)

alpha: 0.3 # 0.2-0.3 for most datasets

deltaRT_xlogP: 4.0 # Only for C18 methods, avoid spurious annotation regarding XlogP vs RT

double_ID_filtering: TRUE # Remove ID duplicates

Module 4: Export

tanimoto_threshold: 0.8 # Structural network cutoff

look_for_DB_identifiers: TRUE # ⏱️ Slow, need internet!

🚨 Common Issues & Solutions

Issue: Workflow fails on first execution

Solution: Click Reset → Start again (normal behavior)

Issue: "Processing time very long"

Solutions:

- Close all other programs

- Lower `topK_insilicoDB` (try 20 instead of 50)

- Disable RT clustering if not critical

- Do NOT run POS and NEG simultaneously

Issue: "No Level 3a annotations"

Solutions:

- Check taxonomic spelling (case-sensitive!, based on NCBI taxonomy)

- Try broader filters (family instead of genus, in-house SMILES list used in annotation tools)

Issue: "POS/NEG merging fails"

Solution: Files must have:

- Identical names across modes

- Same number of samples

- Same acquisition order

→ If not possible, process modes separately and manual concatenate afterwards

📊 Output Files Explained

| File | Description | Use |

|------|-------------|-----|

| `AllAnnotation[mode].csv` | Raw annotations before processing | Reference check |

| `EdgestopK_MSMS_[mode].csv` | MSS network edges | Cytoscape |

| `EdgestopK_Tanimoto_[mode].csv` | Structural network edges | Cytoscape |

| `Nodes[mode].csv` | Node metadata | Cytoscape |

| `Results[mode].xlsx` | Final annotated feature table | Main output |

| `RTCluster[mode].csv` | RT clustering results | QC |

| `StatTable[mode].csv` | Intensity matrix (sum normalized) | Statistics |

| `Report[mode].pdf` | Figures and summaries | QC/Publication |

🎨 Fingerprint Selection Guide

| PubChem | 881 bits | General-purpose, diverse datasets | Higher score (less discriminative) |

| Morgan | 2048 bits | Structurally similar compounds/MSS cluster| Lower score value (highly discriminative) |

| RDKit | 2048 bits | Balanced, medium diversity/large MSS clusters | Medium |

📈 Alpha Parameter (α) Selection

| 0.2-0.3 | Spectral/structural similarity | High-quality seeds, well MSS clustered network |

| 0.4-0.5 | Balanced | Mixed quality, moderate network density |

| 0.6-0.8 | In silico predictions | Few seeds, high network density|

🔬 Confidence Level Summary

| 1 | Authentic standard | Highest confidence |

| 2a | High-confidence library match (>0.85) | Very High confidence |

| 2b | Moderate library match (0.70-0.85) | High confidence |

| 3a | Taxonomically/ user defined elevated in silico | 1 to 3a:Seed for propagation |

| 3b | Network-propagated annotation | Medium confidence |

| 4 | MS/MS Analogs | Low confidence |

| 5 | Unassigned | Lowest confidence |

⏱️ Expected Processing Times

Dataset Size Guidelines

- Small (<500 features): 5-15 min per mode

- Medium (500-2000 features): 15-30 min per mode

- Large (>2000 features): 30-60+ min per mode

Time-Consuming Steps

1. RT Clustering: Can take minutes for large datasets

2. DB Identifier Lookup: Depends on internet

3. Network Propagation: Scales with network density

Speed Optimization

- Process POS and NEG separately (never simultaneously)

- Lower topK values (20 instead of 50)

- Disable RT clustering if not essential

- Skip DB identifier lookup for initial runs

🎓 Tutorial Contact Information

Authors:

- Guillaume Marti (guillaume.marti@utoulouse.fr)

Institutions:

- Laboratoire de Recherche en Sciences Végétales (LRSV)

- Metatoul-AgromiX Platform, Toulouse, France

- MetaboHUB National Infrastructure

## 📚 Key Citation

MS-Net Manuscript (in preparation):

Pereira Francisco V, Duthen S, Crossay E, et al. (2025)

MS-Net: Multi-Similarity based network annotation for untargeted metabolomics