ChemLog: Making MSOL Viable for Ontological Classification and Learning

ChemLog: Making MSOL Viable for Ontological Classification and Learning Simon Flügel (Osnabrück University) Based on: Simon Flügel, Martin Glauer, Till Mossakowski, Fabian Neuhaus: ChemLog: Making MSOL Viable for Ontological Classification and Learning (IJCLR, 2025) Simon Flügel, Till Mossakowski, Fabian Neuhaus, Erik Pfanenstiel, Martin Glauer, Edgar Haak, Adnan Malik, Noel M. O’Boyle: Defining Peptides in ChEBI (JCI, submitted) ChemLog / Flügel et al. / 2 Ontological Classification tripeptide zwitterion dipeptide zwitterion peptide zwitterion dipolar compound chemical entity organic molecular entity nitrogen ylide ylide ? ??? Asp-Lys zwitterion (CHEBI:229953) ChemLog / Flügel et al. / 3 Chemical Entities of Biological Interest (ChEBI) ChemLog / Flügel et al. / 4 Ontological Classification With Logic organooxygenCompound ⇔∃a,b :C ( a ) ∧O ( b ) ∧Bond ( a,b ) ChEBI definitions & expert knowledge ChemLog / Flügel et al. / 7 Why Monadic Second-order Logic (MSOL)? Usual logic choices for ontologies: Description logics (e.g., OWL 2 DL) First-order logic (FOL) MSOL: FOL with quantification over subsets of the domain Problem: Reflexive, transitive closure TC of symmetric relation R not possible with FOL ChemLog / Flügel et al. / 8 Is This an Amino Acid? Amino acid (def.): “A carboxylic acid containing one or more amino groups.” (CHEBI:33709) Implicitly: The carboxy group and amino group have to be connected. CHEBI:25017 - leucine ChemLog / Flügel et al. / 9 Is This an Amino Acid? Amino acid (def.): “A carboxylic acid containing one or more amino groups.” (CHEBI:33709) Implicitly: The carboxy group and amino group have to be connected. Carboxy group Amino group Connection CHEBI:25017 - leucine ChemLog / Flügel et al. / 10 Carboxy group Amino group Connection Is This an Amino Acid? CHEBI:72772 - β-leucine Amino group Carboxy group Connection Amino acid (def.): “A carboxylic acid containing one or more amino groups.” (CHEBI:33709) Implicitly: The carboxy group and amino group have to be connected. CHEBI:25017 - leucine ChemLog / Flügel et al. / 11 CHEBI:72772 - β-leucine Amino group Carboxy group Connection Carboxy group Amino group Connection Is This an Amino Acid? Amino acid (def.): “A carboxylic acid containing one or more amino groups.” (CHEBI:33709) Implicitly: The carboxy group and amino group have to be connected. → Any chain of carbon atoms can be a valid connection. This cannot be expressed in FOL. CHEBI:73044 - lumiracoxib Amino group Connection Carboxy group CHEBI:25017 - leucine ChemLog / Flügel et al. / 18 Peptides – What happens at the N-Terminus? Amino groups may carry acyl or alkyl substituents. CHEBI:191173 - N-acetyl-L-seryl-L-aspartic acid ChemLog / Flügel et al. / 19 Peptides – How are amino acids connected? The amino acid residues are delimited by heteroatoms. CHEBI:191173 - Temocapril ChemLog / Flügel et al. / 20 Peptides – How are amino acids connected? The amino acid residues are delimited by heteroatoms. CHEBI:191173 - Temocapril ChemLog / Flügel et al. / 21 Making MSOL Viable ChemLog / Flügel et al. / 22 FOL Structures CHEBI:15428 - glycine 1 molecule = 1 FOL structure ChemLog / Flügel et al. / 23 Classification as a Model Checking Problem CHEBI:15428 - glycine A molecule is an amino acid iff the corresponding predicate holds for its FOL structure. ChemLog / Flügel et al. / 24 Small Scale: MONA / QBF MONA [1]: MSOL model finder QBF: Quantified Boolean Formulas pred AminoAcid(var2 X) = ex1 A: ex1 B1: ex1 B2: ex1 B3: (AminoGroup(A) & CarboxylicAcid(B1, B2, B3) & IsConnected(A, B1)); Var 2 X; AminoAcid(X); [1] Henriksen, J.G., et al.: MONA: Monadic second-order logic in practice. In: TACAS’95. pp. 89–110. Springer LNCS 1019 (1995) Medium Scale: First-order Logic Translate FOL-able predicates automatically Implement MSOL predicates in Python pred AminoAcid(var2 X) = ex1 A: ex1 B1: ex1 B2: ex1 B3: (AminoGroup(A) & CarboxylicAcid(B1, B2, B3) & IsConnected(A, B1)); Var 2 X; AminoAcid(X); AminoAcid ( x ) ⇔∃a,b1,b2,b3: (AtomIn ( a,x ) ∧AtomIn ( b1,x ) ∧AtomIn ( b2,x ) ∧AtomIn ( b3,x ) ∧AminoGroup ( a ) ∧CarboxylicAcid ( b1,b2,b3 ) ∧IsConnected ( x ) ) Large Scale: Algorithmic def amino_acid(mol, x): for a in mol.GetAtoms(): for b_1 in mol.GetAtoms(): for b_2 in mol.GetAtoms(): for b_3 in mol.GetAtoms(): if (a in x and b_1 in x and b_2 in x and b_3 in x and amino_group(a) and carboxylic_acid(b_1, b_2, b_3) and is_connected(mol, a, b_1)): return True return False Compile FOL to Python ChemLog / Flügel et al. / 25 Performance Dataset (small scale): 1,000 smallest peptides from ChEBI Dataset (medium scale): all molecules from ChEBI’s 3-STAR subset (45,000) Dataset (large scale): PubChem molecules (119.3 million) Dataset Method Timeout rate Time/molecule (s) 1,000 smallest MONA 61.1% 10.8 QBF 0% 832 FOL 0.4% 0.37 3-STAR FOL 2.2% 1.03 Algorithmic 0% 0.0084 PubChem Algorithmic 5.8*10-7 0.0017 ChemLog / Flügel et al. / 26 Limitations Is this a peptide? Our classification: amino acid ChEBI says: dipeptide CHEBI:71596 - mycocyclosin ChemLog / Flügel et al. / 27 Limitations Is this a peptide? Our classification: dipeptide ChEBI says: not a peptide CHEBI:6827 - methicillin ChemLog / Flügel et al. / 34 Large Scale: Algorithmic combine implementation of MSOL predicates with FOL-to-Python compilation optimise for performance AminoAcid ( x ) ⇔∃a,b1,b2,b3: (AtomIn ( a,x ) ∧AtomIn ( b1,x ) ∧AtomIn ( b2,x ) ∧AtomIn ( b3,x ) ∧AminoGroup ( a ) ∧CarboxylicAcid ( b1,b2,b3 ) ∧IsConnected ( x ) ) Implement MSOL part manually Translate to FOL Run in Python Compile to algorithmic def amino_acid(mol, x): for a in mol.GetAtoms(): for b_1 in mol.GetAtoms(): for b_2 in mol.GetAtoms(): for b_3 in mol.GetAtoms(): if (a in x and b_1 in x and b_2 in x and b_3 in x and amino_group(a) and carboxylic_acid(b_1, b_2, b_3) and is_connected(mol, a, b_1)): return True return False Optimise manually def is_connected(mol, a, b): graph = nx.Graph() for bond in mol.GetBonds(): graph.add_edge(bond.GetBeginAtomIdx(), bond.GetEndAtomIdx()) for comp in nx.connected_components(graph): if a in comp and b in comp: return True return False ChemLog / Flügel et al. / 35 Large scale: Algorithmic Dataset (small scale): 1,000 smallest peptides from ChEBI Dataset (medium scale): all molecules from ChEBI’s 3-STAR subset (45,000) Dataset (large scale): PubChem molecules (119.3 million) Dataset Method Timeout rate Time/molecule (s) 1,000 smallest MONA 61.1% 10.8 QBF 0% 832 FOL 0.4% 0.37 3-STAR FOL 2.2% 1.03 Algorithmic 0% 0.0084 PubChem Algorithmic 5.8*10-7 0.0017 ChemLog / Flügel et al. / 36 Improving Deep Learning Models Goal: Ontology classification with Deep Learning models Problem: Performance limited by data quality Solution: Improve data quality with ChemLog ChemLog / Flügel et al. / 37 Improving Deep Learning Models #Chemlog/ #ChEBI Trained on ChemLog data 1.23 41.07 Trained on ChEBI data Blue: tested on ChEBI data Orange: tested on ChemLog data 41.07 1.23 ChemLog / Flügel et al. / 38 Improving Deep Learning Models Deep Learning model: ELECTRA, trained on 14 peptide classes Train / test with labels from either ChEBI or ChemLog ChemLog / Flügel et al. / 39 Two Amino Acids, but no Peptide? CHEBI:31176 - Adipiodone ChemLog / Flügel et al. / 40 Four Amino acids, two dipeptides? CHEBI:17257 - bis-γ-glutamylcystine ChemLog / Flügel et al. / 41 Amino Acid or Peptide? CHEBI:133888 - tabtoxinine-δ-lactam ChemLog / Flügel et al. / 42 How Many Amino Acids? CHEBI:28001 - vancomycin ChemLog / Flügel et al. / 43 New Carboxylic Acid Derivatives Venturamide A (CHEBI:66353)