Technical Summary: SABLE Framework for Hit-to-Lead Optimization

Problem Statement

Hit-to-lead optimization in medicinal chemistry is an inherently multi-objective challenge requiring the iterative design of analogs that balance competing requirements: potency, selectivity, physicochemical properties, pharmacokinetics, safety, and synthetic tractability. Traditional Design-Make-Test-Analyze (DMTA) cycles are constrained by finite experimental capacity, necessitating the prioritization of a small fraction of the accessible chemical space.

Current computational approaches face significant limitations:

1. Fragmentation: Necessary capabilities (reaction-aware enumeration, property prediction, structure-based scoring, and Bayesian optimization) are distributed across standalone tools and services, requiring nontrivial expertise to integrate.
2. LLM Limitations: While Large Language Models (LLMs) offer natural language interfaces, relying on them for quantitative scientific reasoning (e.g., judging chemical validity, scoring affinity, or selecting candidates) leads to hallucinations and unreliable assessments. Existing agentic frameworks often place the LLM at the center of scientific decision-making, which is problematic for rigorous optimization.
3. Synthetic Feasibility: Generative models often propose chemically valid but synthetically inaccessible structures, creating a gap between computational proposals and experimental reality.

Methodology: The SABLE Framework

The authors introduce SABLE (Synthetically-accessible Agentic Bayesian Ligand Exploration), an open-source, modular framework that employs natural-language orchestration to guide chemical structure optimization without relying on the LLM for quantitative decision-making.

Core Architecture

SABLE separates the orchestration layer (LLM) from the quantitative execution layer (specialized tools).

• Orchestration (LLM): The LLM interprets user-defined goals in natural language, extracts parameters (seed molecule, target, objectives, iteration limits), and routes tasks. It does not propose molecules, score affinity, or reason about ADMET liabilities.
• Execution (Specialized Tools):
  • Enumeration: Uses HEALER (Hit Expansion to Advanced Leads using Enumerated Reactions) to construct a reaction-constrained analog space from commercially available building blocks and predefined reaction templates. This ensures all generated candidates are constructed from available building blocks, though the authors note that synthetic accessibility is a necessary but insufficient condition for true experimental tractability and does not constitute a full experimental protocol.
  • Characterization: Integrates modular backends including RDKit (physicochemical descriptors), STOPLIGHT (ADMET), and Boltz-2 (structure-based affinity scoring via diffusion-transformer architecture).
  • Optimization: Employs Bayesian Optimization (BO) using the BayBE library. It utilizes Gaussian process surrogate models and specific acquisition functions:
    • Single-objective: qLogExpectedImprovement (qLogEI).
    • Multi-objective: qLogNoisyExpectedHypervolumeImprovement (qLogNEHVI) to navigate the Pareto frontier.

Workflow Execution

Implemented as a stateful workflow using LangGraph, SABLE operates in iterative cycles:

1. Initialization: The user provides a prompt; the LLM extracts arguments and normalizes identifiers (e.g., drug names to SMILES, protein names to UniProt IDs).
2. Enumeration: HEALER generates a library of analogs based on the seed molecule.
3. Selection & Evaluation: The BO node selects a batch of unevaluated candidates. These are routed to the appropriate characterization tools.
4. Update: Results are appended to the campaign history, updating the surrogate model.
5. Termination: The loop continues until iteration limits, convergence criteria, or search space exhaustion is met.

The system supports a "Bring Your Own Tools" (BYOT) approach, allowing users to swap characterization or enumeration backends via a simple configuration file without altering the orchestration logic.

Key Contributions

1. Modular Agentic Design: SABLE establishes a design pattern where LLMs serve as accessible interfaces for task routing, while domain-specific models and explicit optimization algorithms retain responsibility for quantitative scientific decisions. This mitigates the risk of LLM hallucination in critical scoring tasks.
2. Synthetically Constrained Search Space: By leveraging HEALER, SABLE restricts the search space to molecules constructible from available building blocks. While this ensures proposals are based on known reaction templates and building blocks, the framework acknowledges that this constraint is necessary but insufficient to guarantee experimental tractability or provide a complete experimental protocol.
3. Unified Multi-Objective Optimization: The framework integrates single and multi-objective optimization (e.g., balancing affinity, QED, and CNS activity) within a chemically constrained environment, using BO to efficiently navigate large search spaces without exhaustive screening.
4. Extensibility: The architecture allows for the replacement of components (e.g., swapping Boltz-2 for a new affinity predictor) as the field advances, positioning SABLE as a durable substrate for future optimization pipelines.

Results

The paper evaluates SABLE across several scenarios:

• Single-Objective Optimization: Using CAMKK2 as a target, SABLE improved the predicted logIC50 by 1.03 log units relative to the seed compound, identifying high-affinity candidates while evaluating only a subset of the enumerated library.
• Multi-Objective Optimization: In campaigns optimizing for affinity, QED, and CNS activity, SABLE successfully generated Pareto frontiers, offering trade-off solutions rather than a single optimum.
• Efficiency: On a one-million-compound library, SABLE recovered the global maximum for QED in 3–7 iterations, demonstrating significant efficiency gains over exhaustive screening, particularly as the number of objectives increases.
• Retrospective Case Studies:
  • METTL3: SABLE identified an analog with a predicted concentration of ~102 nM (a ~95-fold improvement over the seed), though it did not rediscover the exact published lead due to limitations in the specific reaction templates available in the HEALER library.
  • Boltz-2 Validation Targets: On four targets (BACE-1, Carbonic Anhydrase XII, ABL1, S1PR1), SABLE enriched the local analog space, identifying molecules with predicted affinities 1–2 orders of magnitude better than the starting weak compounds, converging toward the range of known experimental hits.

Significance and Claims

The paper claims that SABLE provides a computational twin of the analytical and prioritization stages of the DMTA cycle. Its primary significance lies in:

• Bridging the Gap: It connects natural language intent with rigorous, quantitative chemical optimization, making advanced multi-objective optimization accessible to medicinal chemists without dedicated informatics support.
• Actionability: Unlike generative models that produce "hallucinated" or synthetically infeasible structures, SABLE's outputs are constructed from commercially available building blocks via known reaction templates, making them directly actionable for medicinal chemistry teams while acknowledging that this does not guarantee a successful experimental protocol without further review.
• Scalability and Adaptability: By decoupling orchestration from execution, SABLE avoids the "snapshot" limitation of current tools, allowing the community to integrate superior models (e.g., new affinity predictors or enumerators) as they emerge.
• Path to Automation: The framework is designed to eventually integrate with laboratory automation (via REST APIs or Model Context Protocol), enabling a transition from simulated to experimental closed-loop optimization.

The authors modestly note that SABLE's predictions are downstream of imperfect models (e.g., Boltz-2) and that the framework is intended to support, not replace, experimental validation and medicinal chemistry review. The system's performance is ultimately bounded by the coverage of its enumeration engine and the calibration of its characterization models.