Molecular Fingerprint

The most widely used circular fingerprint in cheminformatics, particularly ECFP4. It encodes molecular neighborhoods by iteratively hashing circular atom environments up to a specified diameter into a fixed-length bit vector.

• Mechanism: Uses the Morgan algorithm to assign initial atom identifiers, then iteratively updates them by hashing concatenated neighbor identifiers
• Key property: Captures stereochemistry and pharmacophoric features without requiring pre-defined substructure dictionaries
• Use case: Primary input for QSAR models and virtual screening campaigns
• Limitation: The hashing process causes bit collisions, where different substructures map to the same bit position

A learned alternative to fixed fingerprints. Rather than hashing predefined substructures, MPNNs learn task-specific molecular representations by iteratively passing messages between bonded atoms.

• Core operation: Node hidden states are updated via message functions that aggregate information from neighboring atoms and update functions that combine aggregated messages with the current state
• Advantage over fingerprints: Representations are differentiable and optimized end-to-end for the target property, avoiding the information loss from rigid hashing
• Readout phase: A permutation-invariant function pools all node states into a fixed-length graph-level vector, analogous to a learned fingerprint
• Framework: Generalizes architectures like GGNN, Interaction Networks, and MPNN variants

A 100% robust molecular string representation designed to replace SMILES in generative models. Every syntactically valid SELFIES string corresponds to a valid molecular graph, eliminating the invalid output problem that plagues SMILES-based generation.

• Design principle: Uses a formal grammar with derivation rules that guarantee semantic validity by construction
• Contrast with fingerprints: SELFIES is a discrete, sequence-based representation suitable for language models, while fingerprints are continuous, fixed-length vectors for discriminative models
• Encoding: Atoms and bonds are represented as tokens in a constrained vocabulary, with ring closure handled through explicit state constraints
• Application: Preferred for de novo drug design using RNNs, transformers, or variational autoencoders

A theoretically maximally powerful GNN architecture under the Weisfeiler-Lehman (WL) test. GIN is designed to be as discriminative as the WL graph isomorphism heuristic, making it a strong learned fingerprint generator.

• Key insight: The aggregation function must be injective to distinguish different multisets of neighbor features; sum aggregation achieves this while mean and max do not
• Update rule: Uses a multi-layer perceptron to model the composition of aggregation and update, with a learnable epsilon parameter controlling the weight of the central node
• Relationship to fingerprints: GIN embeddings can provably distinguish molecular graphs that traditional hashed fingerprints might conflate due to bit collisions
• Readout: Concatenates node representations from all layers to capture both local and global structural information

The Smooth Overlap of Atomic Positions descriptor generates a continuous, invariant representation of local atomic environments. Unlike discrete fingerprints, SOAP produces a smooth similarity measure between molecular geometries.

• Construction: Expands a Gaussian-smeared atomic density in a basis of radial functions and spherical harmonics, then computes the power spectrum to achieve rotational invariance
• Key advantage: The resulting kernel captures geometric similarity between 3D conformations, not just 2D topological features
• Integration: SOAP vectors can be used as input to kernel ridge regression models or as features for neural network potentials
• Contrast: Standard fingerprints encode topology; SOAP encodes 3D geometry, making it essential for tasks requiring conformational sensitivity

A self-supervised pre-training framework that learns transferable molecular representations without labeled property data. GraphCL maximizes agreement between differently augmented views of the same molecular graph.

• Augmentations: Applies stochastic perturbations including node dropping, edge perturbation, attribute masking, and subgraph sampling
• Objective: Uses an NT-Xent loss to pull representations of augmented views of the same molecule together while pushing views of different molecules apart
• Advantage over fingerprints: Produces contextualized embeddings that adapt to downstream tasks after fine-tuning, unlike static fingerprints
• Impact: Enables effective transfer learning from large unlabeled chemical databases to low-data property prediction tasks