QSAR-Guided Generative Framework for the Discovery of Synthetically Viable Odorants
Authors: Tim C. Pearce, Ahmed Ibrahim
Organizations: Biomedical Engineering Research Group, School of Engineering, University of Leicester, University Road, Leicester LE1 7RH, United Kingdom · School of Clinical Medicine, University of Cambridge, Addenbrooke’s Hospital, Hills Road, Cambridge CB2 0SP, United Kingdom
The discovery of novel odorant molecules is key for the fragrance and flavor industries, yet efficiently navigating the vast chemical space to identify structures with desirable olfactory properties remains a significant challenge. Generative artificial intelligence offers a promising approach for de novo molecular design but typically requires large sets of molecules to learn from. To address this problem, we present a framework combining a variational autoencoder (VAE) with a quantitative structure-activity relationship (QSAR) model to generate novel odorants from limited training sets of odor molecules. The self-supervised learning capabilities of the VAE allow it to learn SMILES grammar from ChEMBL database, while its training objective is augmented with a loss term derived from an external QSAR model to structure the latent representation according to odor probability. While the VAE demonstrated high internal consistency in learning the QSAR supervision signal, validation against an external, unseen ground truth dataset (Unique Good Scents) confirms the model generates syntactically valid structures (100% validity achieved via rejection sampling) and 93.9% unique structures. The latent space is effectively structured by odor likelihood, evidenced by a Frechet ChemNet Distance (FCD) of ~6.96 between generated molecules and known odorants, compared to ~21.6 for the ChEMBL baseline. An ablation of the property head shows that its contribution lies in the density of retrievable structures near an odorant seed -- a 4.3-fold increase in distinct novel compounds per seed -- rather than in the odor-likeness of any individual compound. Bemis-Murcko analysis shows 2,499 of 6,765 ring-containing candidates, spanning 1,429 distinct core frameworks, absent from all training sets, while a further 294 molecules...
Figures & tables
Figure 1
Figure 1: Variational Autoencoder framework concept. Abbreviations: CNN, convolutional neural network; GRU, gated recurrent unit; FC, fully connected (dense); KL, Kullback-Leibler divergence; Recon Loss, reconstruction loss (categorical cross-entropy); Prop Loss, property prediction loss (mean squared error, as described in text). The diagram illustrates the flow: input SMILES from ChEMBL are processed by an Encoder (CNN based) mapping to a continuous latent space. Latent vectors z are input to both a Decoder (GRU based) for reconstructing SMILES sequences and an Odor Property Prediction head (FC based) predicting odor probability P(odor) jointly with two auxiliary regression targets (LogP and boiling point), of which only P(odor) enters the property loss. The combined training objective minimizes the Recon Loss, the Prop Loss, and the KL divergence regularization term.
Hyperparameter
Values Evaluated
Latent Space Dimensions ( D )
64
128
196
256
512
GRU Hidden Units
128
488
512
768
1024
Convolutional Growth Factor
1.00
1.10
1.12
1.16
1.25
Learning Rate ( η )
1.0×10−3
5.0×10−4
3.12 × 10 -4
1.0×10−4
1.0×10−5
Adam Optimizer Decay ( β1 )
0.900
0.925
0.937
0.950
0.990
Batch Size ( B )
32
64
100
128
256
Table 1: Hyperparameter Optimization Search Space. The optimal values selected for the final VAE-QSAR model are highlighted in bold .
Figure 2: QSAR Odor molecule prediction results. a) Schematic showing the QSAR model pipeline where logistic regression was trained on 1,924 experimentally validated and/or literature reported molecules and subsequently used to label ChEMBL molecules with a target odor probability, b) UMAP visualization of training set in space of molecular descriptors showing the odor probability, c) classifier performance for logistic model, and d) Distribution of predicted probabilities across the training set, and e) residual plots for the logistic regression model.
Figure 3: a) UMAP dimensionality reduction and visualization of the molecular descriptor space showing ~5 x 105 VAE training set ChEMBL molecules (green) alongside the original set of odor molecules (red) and non-odor molecules (blue) and the same UMAP representation showing the predicted probabilities across all ChEMBL molecules used for VAE training as predicted by the QSAR model (right). Inset histogram for the frequency of odor probabilities in the ChEMBL set used to train the VAE, b) t-SNE dimensionality reduction of 1 x 104 randomly sampled ChEMBL molecules encoded into the latent space (green) together with validated odor (red) and non-odor molecules (blue), c) Validated odor molecule probabilities as predicted by the QSAR model, P(qsar) , compared against VAE odor property prediction, P(odor) . Shaded region shows where both models agree on odor molecule prediction ( P(odor)>0.5 and P(qsar)>0.5 ).
Dep. Variable:
odor.class
No. Observations
3230
Model
Logit
Df Residuals
2858
Method
MLE
Df Model
371
Pseudo R 2
0.894
Table 2: Logistic Regression Model Diagnostics
Figure 4: Generated molecules from two example seed odor molecules. a & b) Two example seed molecules (seed is distance 0.0, top left) with sampling of the latent space over increasing distance from the seed in latent space (units of σϕ(z∣X) ). Where available Pubchem catalogued compounds show their corresponding IUPAC name and CID reference, whereas potentially novel compounds show the corresponding SMILES and CID: N/A. SA: synthesis accessibility score. c) traversing the latent space between two odor seed molecules t=0 is start seed and t=1 is the end seed in 11 discrete steps. Attempts are the number of probabilistic decode cycles required to obtain a valid structure at that point of the latent space.
Figure 5: Generated molecules’ uniqueness. a) Jaccard similarity, and b) Fréchet ChemNet distance across data-sets.
Novelty Category
Count
% of ring-containing
MW
LogP
log 10 (VP)
In AD (%)
Interpretation
1. Exact Memorization
423
6.3
157.64
2.44
1.58
100.0
Model memorized training data (Overfitting).
2. Odorant Derivatization
3291
48.6
181.18
2.47
1.18
81.2
Optimizing side-chains of known
(QSAR Scaffold)
odorant frameworks.
3. Repurposing
258
3.8
158.31
2.23
1.70
50.8
Converting generic bioactive
(ChEMBL Scaffold)
frameworks into odorants.
4. Validated Scaffold Hop
294
4.3
153.27
2.01
2.61
33.0
True Generalization: Found known odorant
Table 3: Bemis-Murcko Scaffold Classification of Generated Molecules. The analysis distinguishes between memorization, derivatization of known structures, and the discovery of novel core frameworks. Average properties (Molecular Weight, LogP, and log10 Vapor Pressure) indicate that even novel scaffolds maintain physicochemical profiles consistent with volatility. Averages are arithmetic means over the molecules of each category; log10 (VP) is the mean of per-molecule log10(Psat/Pa) values, averaged over those molecules carrying a vapor pressure prediction ( n=325 of 423 for category 1; complete coverage for categories 2 to 5). Classification was performed on the ring-containing subset of the canonically de-duplicated generated molecules with P(qsar)>0.99 ( n=6,765 of 18,992 ; the remaining 12,227 are acyclic and have no defined scaffold, see Supplementary Table S1). Category 1 is a molecule-level match; categories 2 to 4 are framework-based and are therefore defined only for ring-containing molecules. Rows 2 to 4 are unaffected by the ring restriction, having no acyclic members. “In AD” is the percentage of each category’s molecules lying inside the applicability domain, that is, within a nearest-neighbor Tanimoto similarity of 0.35 on ECFP4 fingerprints to the QSAR training set. The applicability-domain column is reported, not applied: Table 3 classifies the full ring-containing library, while the 0.35 threshold is applied only to the proposed candidate set (Methods, “Applicability domain”; Results, “Applicability domain of the QSAR model”).
Figure 6: Generated molecules odor probability. a) UMAP representation of seed and generated molecules encoded in the latent space color-coded by odor (inset shows the P(qsar) distribution across all generated molecules). b) probability of generated molecule being odorous with distance from the seed (in units of σϕ(z∣X) ) averaged over all generated molecules (thickness shows 95% confidence interval). c) QSAR model predicted odor probabilities of random 1k molecules sampled from ChEMBL, generated molecules set and unique Good Scents (ground truth compounds).
Figure 7: Ablation of the odor prediction head. Both models share an identical architecture, training corpus and training schedule, differing only in the presence of the QSAR property loss ( γLP(odor) , Equation 12 ); both were seeded with the same 984 odorant compounds ( P(qsar)>0.9 ) and given 100 decode attempts per seed. Distance is normalized by each model’s own mean pairwise inter-seed distance ( 2.59 with the head, 1.57 ablated), so that 1.0 (dotted vertical line) corresponds to the typical separation between two seed odorants; without this normalization the 1.65× contraction of the ablated latent space would itself produce an apparent difference. (a) Distinct novel molecules per seed as a function of distance; shading gives the 95% confidence interval bootstrapped over seeds. (b) Mean QSAR odor probability of the same compounds, both models scored with the same external logistic-regression model. (c) Zero-perturbation decoding control, each seed’s own latent mean decoded 100 times: valid strings emitted, decodes returning the seed itself, and distinct molecules obtained. (d) Bemis-Murcko classification of the novel compounds ( P(qsar)>0.99 ) through the pipeline of Table 3 , restricted to ring-containing molecules. Values are given in the text.
Full model
Ablated
Novelty category
Count
% Gen.
Count
% Gen.
1. Exact Memorization
127
4.2
58
8.5
2. Odorant Derivatization (QSAR scaffold)
1,441
47.7
393
57.9
3. Repurposing (ChEMBL scaffold)
130
4.3
36
5.3
4. Validated Scaffold Hop (rediscovery)
122
4.0
23
3.4
5. Uncharted Scaffold Hop (new IP)
1,199
39.7
169
24.9
Table 4: Bemis-Murcko scaffold classification of the novel compounds generated by the full model and by the ablated model (no odor prediction head), from an identical seed set ( 984 seeds, P(qsar)>0.9 ) and an identical budget of 100 decode attempts per seed. Both sets were classified against the same three reference sets (QSAR training set, ChEMBL, Unique Good Scents) and filtered at P(qsar)>0.99 , exactly as in Table 3 and restricted to ring-containing molecules, as defined in Methods (“Scaffold classification”).
Chemical Validity Checks
Geometric
Physical
Structural Alerts (PAINS)
Novelty / Overlap
Dataset
RDKit
OpenBabel
Cross
Bredt
Energy Strain
Clean /
PAINS
Unique
Overlap
Valid (%)
Valid (%)
Valid (%)
Compliant (%)
Acceptable (%)
Free (%)
Alerts (%)
(%)
(%)
Generated molecules
100.00
100.00
100.00
99.90
99.38
97.74
2.26
93.94
6.06
QSAR Training Set
100.00
100.00
100.00
99.27
97.00
97.51
2.49
95.74
4.26
Unique Good Scents
100.00
100.00
100.00
96.85
99.60
99.24
0.76
99.08
0.92
All Good Scents
-
-
-
97.13
99.40
99.22
0.78
97.45
2.55
Table 5: Combined Validity, Energy, Structural, and Novelty Metric Checks
Figure 8: Physicochemical and ADMET property validation on generated molecules compared against ChEMBL training molecules and unique Good Scents (ground truth compounds) . a-h) Predicted physicochemical properties, i-l) Predicted ADMET properties, NR-AR: Nuclear Receptor – Androgen Receptor; NR-ER: Nuclear Receptor – Estrogen Receptor; AMES - Ames Mutagenicity Test; Skin Reaction. Panels f–h show integer-valued counts as discrete bars giving the relative frequency of each value; the remaining panels show kernel-density violins.
Data-set
% Solved
Avg state
Min state
Max state
State score
Avg synth
Min synth
Max synth
Synth Routes
Avg
Avg
Avg precurs.
score
score
Score
st. dev.
routes
routes
routes
st. dev
reactions
precursors
in stock
Generated molecules
100.00
0.76
0.05
1.00
0.27
5.30
1.00
25.00
2.71
2.89
2.54
2.11
Unique Good Scents
100.00
0.92
0.05
1.00
0.17
4.10
1.00
16.00
2.31
1.85
1.99
1.71
ChEMBL
100.00
0.90
0.05
1.00
0.15
4.85
1.00
21.00
2.50
2.65
2.67
1.71
All Good Scents
100.00
0.93
0.05
1.00
0.17
3.82
1.00
15.00
2.15
1.61
1.83
1.60
Table 6: Retrosynthesis analysis of data-sets
Figure 9: Synthesis accessibility and quantum chemistry checks. a) UMAP representation of seed and generated molecules encoded in the latent space color-coded by synthesis accessibility (SA – lower score relates to simpler synthesis complexity), b) comparison of total molecule and per atom energy calculations between generated molecules and Good Scents (ground truth molecules).
Molecular generation develops computational methods to create chemical structures and conformations for discovery, optimization, and structure elucidation.
We introduce Fraglingo, an autoregressive molecular generator that constructs molecules step by step from chemically meaningful fragments connected through predefined attachment sites. At each generation step, Fraglingo jointly predicts which fragment to add and how it should attach by producing an attachment-aware fragment embedding and retrieving the nearest fragment through latent-space search. A wildcard-anchored readout represents both the growing molecule and candidate fragments relative to their attachment sites, enabling a single latent prediction to determine both fragment identity and attachment configuration. Because prediction operates in a continuous embedding space rather than over fixed fragment identifiers, larger fragment libraries can be introduced at inference time without retraining. This retrieval-based formulation provides a unified generation primitive for molecule generation, scaffold generation, scaffold decoration, and molecule optimization. Fraglingo also supports property-conditional generation, allowing desired molecular properties to guide the generation process. On controlled property-conditional benchmarks, Fraglingo achieves stronger joint property control than comparably trained baselines while maintaining competitive validity, uniqueness, and novelty. It also generalizes to fragment libraries up to four times larger than those used during training without retraining. Code is available at: https://anonymous.4open.science/r/FragLingo-3551.
Thao Nguyen, Jeonghwan Kim, Zhenhailong Wang +1
Siebel School of Computing and Data Science University of Illinois Urbana-Champaign Urbana, IL 61801, USA
Many diffusion based molecule generation methods ignore the symbolic information of molecules and represent the atom and bond type as one hot representation. Methods based on Morgan fingerprints produce hash collisions and are hard to embed into a continuous space without information loss and random fingerprints correspond to no valid molecule. To circumvent this issue we use another paradigm and consider atom and bond codes as latent variables of VQ-VAE. We introduce VQ-SAD which first trains a VQ-VAE and uses the frozen pretrained VQ-VAE model and considers the codebooks for both atom and bond types as tokenizers for the downstream diffusion process. VQ-SAD is a neuro-symbolic model that utilizes both symbolic and neural structural information for a diffusion based model with learnable forward process. The large discrete code space provides a more balanced atom and bond types which enhances the denoising process. VQ-VAE slightly outperforms SOTA models for diffusion based molecule generation on QM9 and ZINC250k datasets.
Mosquito-borne infectious diseases cause more than 700000 deaths worldwide each year. The long-term use of conventional chemical insecticides has induced serious resistance problems, creating an urgent need to develop novel, highly effective, and ecologically sustainable alternatives. While existing artificial intelligence approaches in this domain have focused primarily on activity prediction and classification, they leave a critical gap in the de~novo generation of novel molecular scaffolds. In this study, we propose Mos-Gen, a motif-aware generative collaborative framework that couples the pretrained molecular representation model Uni-Mol with a variational autoencoder (VAE), specifically tailored for the design of disulfide-containing allicin derivatives as mosquito insecticides. Among the generated candidates, fourteen compounds -- comprising nine predicted positives and five predicted negatives -- were selected for chemical synthesis and experimental validation. The hit rate among the predicted positives reached 78%, whereas none of the predicted negatives exhibited mosquitocidal activity. These experimental results fully validated the high-precision screening capability of the Mos-Gen framework.
Lina Wang, Yaning Cui
Shanghai Institute of Organic Chemistry, Chinese Academy of Sciences, 345 Lingling Road, Xuhui District, Shanghai, China · DP Technology, Shanghai, China