Use LibInvent Online

LibInvent: Reaction-Based Generative Scaffold Decoration for Focused Chemical Library Design

Lib-INVENT is a de novo drug design tool that rapidly proposes focused chemical libraries of compounds sharing the same core while maximizing desirable properties, with optional user-specified chemical reactions guiding how each scaffold is decorated. It was developed by Vendy Fialkova, Jiaxi Zhao, Kostas Papadopoulos, Ola Engkvist, Esben Jannik Bjerrum, Thierry Kogej and Atanas Patronov at Molecular AI, Discovery Sciences, AstraZeneca (Gothenburg, Sweden), with Uppsala University. The authors state the code is freely available in a public repository.

The desired activity of a compound is usually tied to its structural core, so lead optimization often means keeping a scaffold and varying only selected moieties. In practice this is done by screening very focused libraries that share the same core, ideally with moieties introduced through the same or similar reactions so they can be made under the same conditions.

According to the authors, no tool capable of designing such libraries had been proposed. Earlier decorator models sliced molecules with exhaustive RECAP rules, which the authors found does not teach a model to respect real reactions, so it cannot satisfy reaction requirements when designing libraries.

How LibInvent Designs Reaction-Aware Compound Libraries

1. Reaction-based slicing: compounds from ChEMBL are sliced according to chemical reactions into scaffold and decoration pairs, so decorations are tied to reactions the model can learn.

2. Pretrained prior: a SMILES-based decorator is pretrained by teacher forcing to produce valid decorations for each attachment point of a scaffold with a variable number of attachment points.

3. Reinforcement learning focus: the prior is adapted to a task using a scoring function with scores in the range 0 to 1. Compounds scoring above a user-defined threshold enter a scaffold memory.

4. Reaction filters: reactions can be specified for each attachment point of the scaffold and included in the score, steering decorations toward chosen chemistry.

Choosing the Reward Strategy

The authors compared four reinforcement learning strategies (DAP, SDAP, MAULI and MASCOF). DAP combines prior likelihood with the score to anchor the agent to learned chemical space and was selected as the method of choice.

Benchmark Results for Scaffold Decoration with LibInvent

Experiments targeted DRD2 actives using a QSAR scoring model, with results averaged over three runs and no reaction filters.

With a selective reaction filter (amide coupling at one attachment point and a Buchwald reaction at the other), more than 80% of DAP proposals still exceeded the 0.4 score threshold and nearly 90% of the scaffold memory satisfied both filters. In experiments with scaffolds of one and three attachment points, output validity stayed consistently above 90%.

What is Tamarind Bio?

Tamarind Bio is a no-code bioinformatics platform built to give life scientists and researchers access to powerful computational tools. Many cutting-edge machine learning models are hard to deploy and use. Tamarind provides an intuitive, web-based environment that removes the complexity of high-performance computing, software dependencies and command-line interfaces.

The platform is designed for biologists, chemists and other researchers who may not have a background in programming or cloud infrastructure but want to run models on their own data. Key features include:

• A user-friendly graphical interface for setting up and launching experiments

• A robust API for integration into existing research pipelines

• An automated system for managing and scaling computational resources

Tamarind treats information and data security as a top priority, as detailed in its Trust Center and Terms of Service.

Accelerating Discovery with LibInvent on Tamarind Bio

• Lead optimization: generate core-sharing virtual libraries around a hit or lead scaffold.

• Reaction-aware design: constrain decorations to reactions a chemistry team can run, such as amide coupling or Buchwald coupling.

• Property-guided libraries: steer generation with scoring components such as a QSAR model or ROCS shape similarity, both demonstrated by the authors.

• Varied scaffolds: decorate scaffolds with one, two or three attachment points.

How to Use LibInvent on Tamarind Bio

1. Log in and open the tool: sign in at tamarind.bio and select LibInvent.

2. Provide the scaffold: enter the core as SMILES with labeled attachment points, in the format the tool defines.

3. Choose reactions: if exposed, specify the reaction to apply at each attachment point.

4. Set scoring and size: if exposed, configure the scoring function, score threshold and number of compounds.

5. Run the job: submit and wait for generation to complete.

6. Download outputs: retrieve the decorated compounds and their scores.

7. Interpret results: check validity, scores and the fraction satisfying reaction filters.

8. Use downstream: filter, dock or prioritize compounds for synthesis.

Things to Keep in Mind

• The user is responsible for supplying feasible reactions for each attachment point. An infeasible reaction always scores 0 and can cause a failed run with small, irrelevant output.

• A reaction filter score plateauing below one is a sign of a wrong reaction requirement.

• A QSAR scoring model has a strong inductive bias and favors certain decorations, so it can struggle with reaction routes incompatible with those groups.

• Finding the best balance of guidance and exploration often needs some experimentation.

Source: Fialkova V., Zhao J., Papadopoulos K., Engkvist O., Bjerrum E. J., Kogej T., Patronov A. "Lib-INVENT: Reaction Based Generative Scaffold Decoration for in silico Library Design." Molecular AI, AstraZeneca. Code: https://github.com/MolecularAI/Lib-INVENT