TL;DR
This paper introduces SmilesGEN, a novel generative model that designs drug-like molecules by learning from both molecular structures and their resulting cellular phenotypic changes. By jointly modeling these aspects in a shared latent space, SmilesGEN can generate molecules with a higher likelihood of inducing desired therapeutic effects, outperforming existing methods.
What Problem Does This Paper Address?
The traditional approach to drug discovery, target-based drug design (TDD), focuses on identifying and interacting with specific molecular targets. While effective, TDD can be resource-intensive, narrow in scope, and prone to off-target effects. Phenotypic drug discovery (PDD) offers an alternative by focusing on the observable effects of compounds on biological systems, even when the underlying molecular mechanisms are unknown. However, PDD faces challenges in optimizing lead compounds for efficacy and safety without a defined target. This paper addresses this gap by proposing a method that leverages gene expression profiles (phenotypic profiles) as a guide for de novo molecule generation, aiming to create molecules that elicit specific, desirable phenotypic changes.
Methods
SmilesGEN is built upon a dual-channel variational autoencoder (VAE) architecture, comprising two interacting VAEs: SmilesNet for molecules and ProfileNet for expression profiles.
- SmilesNet: This VAE is pre-trained on a large dataset of SMILES strings (a chemical notation) to learn the "grammar" of molecular structures. It encodes molecules into a latent representation and can decode this representation back into valid molecular structures.
- ProfileNet: This VAE is designed to encode gene expression profiles. A key innovation here is its ability to explicitly model the impact of drug perturbations. It's constrained to reconstruct the pre-treatment cellular state when the drug's influence is "removed" in the latent space. This allows it to capture the net effect of a drug's perturbation.
The core idea is to map both molecular structures and expression profiles into a common latent space. This shared representation allows the model to learn the interplay between a molecule's structure and the phenotypic changes it induces. Specifically, ProfileNet learns to reconstruct pre-treatment profiles, while SmilesNet is conditioned on desired phenotypic profiles to generate molecules that are likely to cause those changes.
Datasets
The paper utilizes several datasets for training and evaluation:
- Gene Expression Data: Expression profiles induced by genetic perturbations of 10 common cancer-associated genes (8 knockdown, 2 overexpression) from a benchmark dataset.
- Drug-Target Information: The Drug Target Commons (DTC) database was used to find known ligands for the target genes.
- Cancer Patient Data: Expression profiles from patients with triple-negative breast cancer (TNBC), colorectal cancer (CRC), and lung adenocarcinoma (LUAD) from The Cancer Genome Atlas (TCGA).
- Approved Drug Data: The DrugBank database was used to retrieve approved drugs for the selected cancer types.
Key Insights and Ideas
- Bridging Phenotype and Structure: The most significant insight is the successful integration of phenotypic information (gene expression changes) directly into the molecule generation process. This moves beyond simply generating "valid" molecules to generating molecules with a purpose – to induce specific cellular responses.
- Modeling Perturbation: Explicitly modeling the change in gene expression caused by a drug, rather than just the post-treatment state, is crucial. This allows the model to understand the drug's specific impact and reverse it or guide it towards a desired outcome.
- Latent Space Synergy: By mapping both molecules and expression profiles into a shared latent space, the model learns a richer representation that captures the complex relationship between chemical structure and biological effect.
- Therapeutic Relevance: The model's ability to generate molecules that are not only structurally sound but also highly similar to known ligands and approved drugs demonstrates its potential for real-world drug discovery.
Key Results
- Superior Performance: SmilesGEN significantly outperforms several state-of-the-art methods in generating molecules with higher validity, uniqueness, and novelty.
- High Similarity to Known Ligands: The generated molecules achieved significantly higher Tanimoto coefficients (a measure of structural similarity) compared to competing methods when aiming to target specific proteins. For example, for the AKT1 target, SmilesGEN achieved a Tanimoto coefficient of 0.98, compared to 0.66 for the next best method.
- Drug-Likeness: Generated molecules showed a higher QED (quantitative estimate of drug-likeness) score, with a peak around 0.7, compared to known ligands which peaked below 0.55.
- Scaffold Optimization: The model successfully optimized benzene ring scaffolds and short-chain ligands, incorporating functional groups and improving physicochemical properties while maintaining or enhancing drug-likeness.
- Therapeutic Molecule Generation: For specific cancer types (TNBC, CRC, LUAD), SmilesGEN generated molecules that showed high similarity to approved drugs, with one instance achieving a Tanimoto similarity of 0.98 for lung cancer.
Limitations and Future Work
- Linearity Assumption: The model assumes a linear additivity for modeling drug perturbations, which might not capture all complex biological interactions.
- Data Availability: The performance is contingent on the availability of high-quality gene expression data linked to drug perturbations.
- Future Directions: The authors suggest that SmilesGEN provides a robust framework for bridging chemistry and biology. Future work could explore more complex biological interactions, expand the range of phenotypic profiles considered, and further validate the generated molecules through experimental testing.
Summary updated