A feature vector is an ordered collection of features that represents one input instance. Its dimensionality is the number of features it contains. For a toxicity classifier, each molecule can have a vector describing properties such as molecular weight and counts of atoms or bond types; its label says whether it is toxic. section 1.2.2
In a molecular graph, features can describe individual atoms or bonds, as well as the whole graph. These attributes give a graph neural network information it can use to make predictions. section 11.2 section 11.2.2
An atom as a vector
An atom can be represented by a 78-dimensional vector combining its element, degree, hydrogen count, implicit valence, and aromaticity. The element is one-hot encoded: a vocabulary assigns a position to each element, and the atom’s element activates its corresponding position. RDKit extracts these properties when converting a SMILES string into a molecular graph. section 11.2.2
Listing 11.2 Generate 78-dimensional Feature Vector for a Single Atom
def atom_features(atom):
features = []
features += one_of_k_encoding_unk(
atom.GetSymbol(), ATOM_SYMBOLS
)
features += one_of_k_encoding_unk(
atom.GetDegree(), [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10]
)
features += one_of_k_encoding_unk(
atom.GetTotalNumHs(), [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10]
)
features += one_of_k_encoding_unk(
atom.GetImplicitValence(), [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10]
)
features += [int(atom.GetIsAromatic())]
return np.array(features, dtype=np.float32)Inspect how the atom properties are encoded and combined into a NumPy array. section 11.2.2
A protein residue as a vector
A protein graph can represent each amino acid residue with 54 features: 21 for amino acid identity, 21 from a position-specific scoring matrix, and 12 biochemical properties. The scoring matrix provides evolutionary context from aligned sequences. In this implementation it contains position probabilities rather than traditional log-odds scores. section 11.3.2
Listing 11.6 Generate 54-dimensional Feature Matrix for Protein
def compute_target_feature(aln_file: str, pro_seq: str) -> np.ndarray:
pssm = PSSM_calculation(aln_file, pro_seq)
pssm_transposed = np.transpose(pssm, (1, 0))
other_feature = seq_feature(pro_seq)
return np.concatenate(
(pssm_transposed, other_feature), axis=1
) #ASee how the position-specific matrix and other residue features are concatenated into a protein feature matrix. section 11.3.2
More features can provide useful information, but high-dimensional inputs can overfit when the training dataset is too small. section 11.3.2




