PROFIS: Design of Target-Focused Libraries by Probing Continuous Fingerprint Space with Recurrent Neural Networks

Hubert Rybka, Tomasz Danel, Sabina PodlewskaJournal of Chemical Information and Modeling 2025 · DOI · OpenReview

TL;DR

This paper introduces PROFIS, a novel generative model that designs new drug-like molecules by exploring a continuous "fingerprint space" using recurrent neural networks. PROFIS can generate diverse and novel molecular libraries tailored to specific biological targets, offering a powerful new tool for drug discovery.

What Problem Does This Paper Address?

The drug discovery process is notoriously expensive and time-consuming. Computational methods, particularly machine learning (ML), are increasingly used to accelerate this process. A key challenge is generating novel molecular structures that are not only chemically valid but also possess desired biological activity for a specific target. Traditional methods often rely on screening existing libraries or using generative models that struggle to explore entirely new chemical spaces or focus on specific targets. Existing molecular representations like SMILES can be ambiguous for machine learning models, while molecular fingerprints, though compact, lose explicit structural information, making direct reconstruction difficult. This paper aims to bridge this gap by developing a generative model that can effectively navigate a continuous representation of molecular structures (fingerprint space) to design novel, target-specific compounds.

Methods

PROFIS is built upon a Variational Autoencoder (VAE) architecture. The VAE's encoder transforms molecular fingerprints (specifically, KRFP and ECFP4) into a low-dimensional, continuous latent space. This latent space is then used in two ways:

  1. Activity Prediction: A Quantitative Structure-Activity Relationship (QSAR) classifier is trained on this latent space to identify "active" subspaces that are likely to correspond to potent ligands for a given biological target.
  2. Molecule Generation: A recurrent neural network (RNN), specifically a three-layer Gated Recurrent Unit (GRU), acts as the decoder. It takes points from the identified active latent subspaces and translates them back into molecular structures represented as SMILES strings.

The model is trained using a two-factor loss function: one for reconstruction error (ensuring the decoded molecule matches the input fingerprint) and another for KL regularization (to maintain a well-behaved latent space). To address the common "KL-vanishing" problem in VAEs, the authors employ KLD annealing during training. For searching the latent space, a Bayesian optimization algorithm is used in conjunction with the trained QSAR classifier.

Datasets

The study utilized two primary datasets:

  • Training Dataset: Molecules were obtained from the ChEMBL and ZINC databases. For ChEMBL, molecules targeting the human D₂R receptor were filtered for specific molecular weight ranges and adherence to Lipinski's rule of five. The ZINC-250k subset was used without filters. All molecules were processed to neutralize charged atoms and remove counterions. This resulted in a dataset of 1,126,085 compounds, for which KRFP and ECFP4 fingerprints were generated.
  • D₂R Ligand Dataset: A specific dataset of D₂R ligands was extracted from the ChEMBL database, filtered for human D₂R ligands with Kᵢ values less than 10 µM. This dataset was used to train the latent space QSAR classifier. Molecules with Kᵢ ≤ 100 nM were labeled "active," and others "inactive."

Key Insights and Ideas

  • Probing Continuous Fingerprint Space: The core innovation is using a continuous latent space derived from molecular fingerprints as a navigable landscape for designing new molecules. This allows for exploration beyond discrete chemical structures.
  • Target-Focused Generation: By integrating a QSAR classifier trained on the latent space, PROFIS can specifically search for regions that predict high activity against a particular biological target, leading to more relevant compound libraries.
  • Scaffold-Hopping Capabilities: PROFIS demonstrates excellent scaffold-hopping potential, meaning it can generate molecules with novel core structures that are structurally different from known ligands but still retain desired activity. This is crucial for discovering truly novel chemical entities.
  • SMILES vs. DeepSMILES vs. SELFIES: The study highlights the trade-offs between different molecular string representations. While DeepSMILES offers higher validity (fewer invalid molecules generated), SMILES-based generation produced molecules with better structural complexity and more favorable physicochemical properties. SELFIES, despite its robustness, led to many undesirable artifacts.

Key Results

  • Latent Space Classification Performance: The trained latent space QSAR classifiers (MLP, SVM, XGB) achieved high performance (ROC AUC > 0.88) in distinguishing active from inactive D₂R ligands, comparable to or even better than classifiers trained on uncompressed fingerprints.
  • Novelty of Generated Libraries: Approximately 90% of the generated molecules by PROFIS were based on novel scaffolds, meaning they had no similar scaffolds (Tanimoto distance < 0.25) in the training dataset.
  • Target-Specific Binding: Molecular docking studies showed that PROFIS-generated candidate D₂R ligands exhibited stronger predicted binding affinities to the D₂R receptor compared to a random sample of drug-like molecules from ChEMBL. This affinity was not observed when docked to an unrelated protein (bacterial acetylcholine esterase), confirming the target-focused nature of the generated libraries.
  • Analog Generation: PROFIS can generate diverse libraries of structural analogs for a given drug molecule by introducing small perturbations (flipping bits) in its fingerprint.

Limitations and Future Work

  • Output Validity: The authors acknowledge that the validity of generated SMILES strings is moderately low (around 3-39% depending on the model and sampling strategy). This necessitates more generation trials to obtain a sufficient number of valid molecules.
  • SELFIES Artifacts: While SELFIES offers 100% validity, the generated molecules often contained undesirable artifacts (strained rings, unrealistic connectivities) and lacked structural complexity.
  • DeepSMILES Trade-offs: DeepSMILES improved validity but resulted in molecules with less structural complexity compared to SMILES.
  • Applicability Domain: The study defines an applicability domain for the QSAR models, suggesting that MLP predictors are reliable for compounds with a distance to model (DM) less than 0.15.
  • Future Work: The authors suggest that the protocol is universal and can be applied to any biological target with available ligand data. The ready-to-use scripts shared on GitHub aim to facilitate widespread adoption and further research. The potential for lead optimization and designing new drugs around existing patents is also highlighted.

Summary updated

BibTeX
@article{rybka2025profis,
  author = {Hubert Rybka and Tomasz Danel and Sabina Podlewska},
  title = {{PROFIS: Design of Target-Focused Libraries by Probing Continuous Fingerprint Space with Recurrent Neural Networks}},
  journal = {Journal of Chemical Information and Modeling},
  year = {2025},
  doi = {10.1021/acs.jcim.5c00698},
}
Figure