Methodological details
Experimental design
Molesthetics is a web-based pairwise comparison experiment designed to quantify the collective aesthetic perception of molecular structures. Users are repeatedly presented with two molecular depictions and asked to select the structure they find more visually appealing. Each vote is treated as a binary outcome contributing to the estimation of a latent aesthetic score.
The experiment is blind by default: no identifiers, annotations, or scores are displayed during voting, ensuring that decisions rely primarily on visual perception.
Molecular dataset
Molecular structures were randomly sampled from the LOTUS natural products database1, an open resource providing curated chemical structures and associated biological metadata. This ensures broad coverage of naturally occurring chemical space.
Each molecule is represented by a SMILES string and associated with an Elo score and a number of comparisons.
Molecular rendering
Structures are rendered from SMILES using RDKit2 as standardized 2D SVG depictions. Rendering parameters are fixed (bond width, scale), atom indices are hidden, and optional random rotations are applied to reduce orientation bias.
Pair selection algorithm
Pair selection follows a hybrid exploration–exploitation strategy. One molecule is preferentially sampled among those with fewer comparisons, increasing coverage.
- Exploration (≈30%): the second molecule is randomly sampled.
- Exploitation: the second molecule is selected among molecules with similar scores.
This strategy balances global exploration of chemical space with local refinement of rankings.
Scoring model
Molecular scores are estimated using an Elo rating system3, which can be interpreted as an online approximation of a Bradley–Terry model4 for pairwise comparisons.
P(A > B) = 1 / (1 + 10(RB − RA) / 400)
Ratings are updated after each comparison:
R′ = R + K(S − E)
The update factor K decreases with the number of comparisons:
K = max(Kmin, K0 / √(n + 1))
This adaptive formulation allows rapid convergence for under-sampled molecules while stabilizing well-characterized rankings.
Quality control and bias mitigation
- Blind presentation (no metadata displayed)
- Random rotation of structures
- Anonymous session tracking
- Insertion of control comparisons ("traps")
Trap comparisons consist of identical molecules presented against themselves. These are used to quantify user consistency and detect systematic biases such as side preference.
Data acquisition and storage
Each vote is recorded with timestamp, anonymous session identifier, molecule identifiers, selected molecule, trap status, and rendering parameters. Data are stored in a SQLite database operating in write-ahead logging mode to support concurrent access.
Chemical space visualization
A two-dimensional projection of chemical space is computed using molecular fingerprints (e.g., Morgan fingerprints) and dimensionality reduction (UMAP5). The resulting coordinates are displayed in an interactive map, where each point represents a molecule colored by its Elo score.
Limitations
This approach relies on 2D representations of molecules and therefore does not capture stereochemistry or 3D conformational effects. Aesthetic judgments may also be influenced by rendering conventions or user-specific biases. Additionally, the Elo framework assumes transitivity and may not fully capture complex preference structures.
Research objective
The objective of this experiment is to determine whether collective human perception reveals consistent structural preferences and whether these preferences correlate with molecular descriptors or regions of chemical space.
References
1 Rutz, A. et al. eLife 2022, 11, e70780. LOTUS database.
2 RDKit: Open-source cheminformatics. https://www.rdkit.org
3 Elo, A. E. The Rating of Chessplayers; 1978.
4 Bradley, R. A.; Terry, M. E. Rank analysis of incomplete block designs. Biometrika 1952, 39, 324–345.
5 McInnes, L.; Healy, J.; Melville, J. UMAP: Uniform Manifold Approximation and Projection. arXiv 2018.