TL;DR
This paper introduces a novel drug discovery strategy: starting with "null" compounds that already possess excellent drug-like properties (bioavailability, low toxicity) and then using generative AI to imbue them with potency against a specific disease target. This approach successfully generated a compound that inhibits alpha-synuclein aggregation, a key factor in Parkinson's disease, while retaining its favorable drug-like profile.
What Problem Does This Paper Address?
Traditional drug discovery often begins by screening large libraries for compounds that are potent against a disease target. However, these initial "hits" frequently suffer from poor drug metabolism and pharmacokinetics (DMPK) and other developability issues, making them difficult and time-consuming to optimize. This often leads to a trade-off where improving potency can worsen toxicity or bioavailability, and vice-versa. For diseases affecting the central nervous system (CNS), like Parkinson's, the blood-brain barrier presents an additional hurdle, as many potent compounds cannot reach the brain. The authors propose an alternative: starting with compounds that are already "drug-like" (good DMPK, low toxicity, minimal off-target effects) but lack specific potency, and then using AI to introduce that potency.
Methods
The core of the paper lies in the application of MolDQN, a generative machine learning model. MolDQN operates on a deep Q-learning framework, treating chemical modifications as "actions" to guide the generation of new compounds. The model is trained using a weighted compromise of several QSAR (Quantitative Structure-Activity Relationship) models, optimizing for:
- Potency: Inhibition of alpha-synuclein aggregation.
- DMPK/Developability: Specifically, the Central Nervous System Multiparameter Optimization (CNS MPO) score, which predicts a compound's suitability for CNS penetration.
- Toxicity: Assessed through experimental data and computational filters.
- Synthetic Accessibility: Ensuring generated compounds can be practically synthesized.
The authors explored three strategies:
- Repurposing existing drugs: Mining databases for drugs that failed clinical trials without safety concerns.
- In silico screening: Filtering large datasets for compounds predicted to be both non-toxic and potent.
- Generative modeling (MolDQN): The primary focus, used to modify an initial "null" compound.
Datasets
- Alpha-synuclein aggregation inhibition data: A small set of previously generated inhibitors used for training and testing QSAR models.
- Cayman dataset: A library of ~10,000 drug-like compounds used for initial screening and as a source for the "null" parent compound.
- Cell Painting (CP) data: Used for toxicity prediction models.
- ZINC15 database: Used for similarity searches to find commercially available analogs of generated compounds.
Key Insights and Ideas
- "Null Library" Concept: The idea of starting with compounds that are already "good citizens" in terms of DMPK and toxicity, rather than trying to fix these issues later, is a powerful paradigm shift.
- Generative AI for Targeted Optimization: MolDQN proved effective at introducing specific potency without sacrificing the desirable pre-existing properties of the parent compound.
- Balancing Conflicting Optimization Goals: The paper highlights the challenge of simultaneously optimizing multiple, often conflicting, properties (potency vs. DMPK vs. toxicity) and demonstrates how a weighted generative approach can navigate this.
- Limitations of In Silico Screening: The authors found that direct in silico screening for both potency and low toxicity was insufficient, underscoring the need for more sophisticated generative approaches.
- Importance of CNS MPO: For CNS-targeted drugs, the CNS MPO score is a critical metric, and maintaining a good score during optimization is paramount.
Key Results
- The authors successfully applied MolDQN to an initial "null" compound with a perfect CNS MPO score and low experimental toxicity.
- The generative process yielded a compound, G1, which demonstrated significant inhibition of alpha-synuclein aggregation.
- G1 showed potency approaching that of a known inhibitor (Anle-138b) and, crucially, retained the low toxicity of its parent compound.
- While G1's CNS MPO score dropped from a perfect 6 to 3.29, it was still considered viable for CNS penetration.
- The initial attempts at repurposing and direct in silico screening were less successful, with the latter yielding a compound with high cytotoxicity concerns.
Limitations and Future Work
- Commercial Availability vs. Custom Synthesis: The need to find commercially available analogs via similarity searches led to structural deviations from the ideal generated compounds, impacting CNS MPO scores. The authors suggest that custom synthesis of generated compounds would be a more direct and effective approach.
- CNS MPO Degradation: While G1 retained low toxicity, its CNS MPO score decreased. Future work could involve more stringent weighting of the CNS MPO score within the generative model to minimize this degradation.
- Generative Adversarial Networks (GANs): The authors suggest that future approaches could utilize GANs to bias inert compound populations towards regions of chemical space with higher potency while maintaining desirable DMPK properties.
- Broader Applicability: The presented approach is a proof-of-concept and could be extended to other disease targets and "null" compound populations.
Summary updated