> For the complete documentation index, see [llms.txt](https://docs.pending.ai/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.pending.ai/capabilities/generative-ai.md). # Generative AI {% hint style="info" %} Pending AI’s various capabilities are powered by artificial intelligence which is an experimental technology and may occasionally be misleading or incorrect. {% endhint %} Pending AI's **Molecule Generator** is a state-of-the-art, transformer-based service engineered to cut down the time and cost of early-stage drug discovery. It delivers **ultra-high-throughput sampling** of novel, diverse, and pharmacologically-relevant compounds, empowering R\&D teams to rapidly explore vast chemical space and construct highly effective virtual libraries. Our advanced deep learning models perform unconditional molecule generation based on the distribution of massive, quality-focused input datasets. Models are built with drug discovery as a primary concern: * **Validity**: Molecules are chemically sound, meeting standards for synthetic accessibility and structural integrity required for progression. * **Novelty**: Minimal structural overlap with known or patented compounds, ensuring a maximum RoI by focusing on true novel chemical entities. * **Diversity**: Compounds exhibit significant structural dissimilarity, reducing repeated chemical screens and maximising the chemical space explored for a target. See the **PAI Generator** page for more information about the service. {% content-ref url="/pages/Cjr6ws19xxEBwzOesIdY" %} [PAI Generator](/api-reference/pai-generator.md) {% endcontent-ref %} *** ## Applications The Molecule Generator is built to address critical bottlenecks in the therapeutic development cycle - it flexibly addresses several applicable areas for integration into an existing drug discovery pipeline:


De Novo Drug Design	Rapid creation of entirely new molecular scaffolds to address novel or challenging therapeutic targets.
Ultra-High-Throughput Screening	Efficiently enumerating and screening massive virtual libraries (in the billions) to identify promising lead candidates faster than traditional methods.
Novel Hit-Identification	Pinpointing unique, high-potential compounds ready for experimental validation and subsequent lead optimisation.
Focused Library Construction	Targeted sampling to build custom libraries for specific drug targets, scaffolds, or physicochemical property profiles.

*** ## Features & Advantages ### 1. High-Performance Architecture for Scale Pending AI's proprietary transformer models are fine-tuned for the unique demands of chemical generation, prioritising speed without compromising quality. * **Industry-Leading Throughput**: Generate millions of high-quality molecules, drastically compressing the timeline for virtual library creation ([screening](/capabilities/retrosynthesis.md) is made readily available). * **Precision Customisation**: Leverage fine-tuning capabilities to train the generator on in-house datasets, directing sampling toward specific scaffolds or therapeutically relevant ADMET profiles. ### 2. Rigorous Quality Control & Benchmarking A comprehensive benchmarking suite is used that validates every output against industry standards for drug-like quality. * **Drug-Likeness Validation**: Automated checks for validity, uniqueness, and adherence to established drug-like filters (e.g., Lipinski's Rule of Five). * **Optimised Diversity**: Continuous measurement of novelty and average dissimilarity to confirm the effective exploration of new chemical space. * **Feature Consistency**: Sampled physicochemical feature distributions (e.g., molecular weight, $$LogP$$, $$TPSA$$) are compared against optimal ranges, ensuring molecules possess certain characteristics.

### 3. Extensive and Vetted Chemical Data Foundation The model is trained on a massive, diverse, and meticulously curated collection of molecular datasets drawn from both public and commercial sources. This foundation ensures the model has a vast embedded understanding of the complexity and rules of medicinal chemistry. * **Data Integrity**: A rigorous preparation process ensures underlying data is high-quality and standardised, resulting a robust model that generates reliable, chemically-sound outputs. *** ## **Model Overview** There are several models available for sampling molecules. Each model contains a semantically limited chemical space based on underlying training distributions sourced from large compound libraries. Architecture components are specially designed for the drug discovery domain; advanced token embedding models optimised for novelty and diversity have been improved to further extract chemically relevant features for early Design & Make stages. {% hint style="info" %} Model utility is uniquely validated against Pending AI's own drug discovery pipeline which has accelerated the early Hit Identification and refinement stages. {% endhint %}

Name	Description	Overall Rank
Diverse Small Transformer	This model excels at generating a broad range of general chemical structures, including those found in patents and reaction databases, natural products, and known drug molecules. It offers high diversity in its generated output.	2
Docking Tiny Transformer	This model specialises in generating molecules with high drug-likeness and is particularly well-suited for virtual screening applications. It offers a very high throughput for generating potentially relevant compounds for early-stage drug discovery.	4
Docking Small Transformer	Building on the strengths of its predecessor, this model also excels at virtual screening and generating drug-like molecules. It has the highest throughput and improved rates of generating valid and unique molecules compared to previous iterations.	3
Docking Medium Transformer	This model is the most robust for virtual screening molecule generation. It offers an excellent balance of generating valid, unique, and drug-like molecules, and its larger architecture provides a more comprehensive exploration of the chemical space relevant to drug discovery, leading to a high amount of novel, drug-like compounds.	4

{% @mermaid/diagram content="radar-beta title PAI Generator Model Performance Rankings axis t\["Throughput"], v\["Validity"], n\["Novelty"], d\["Diversity"] curve aa\["Diverse Sml"]{1, 1, 1, 4} curve ab\["Docking Tny"]{3, 3, 3, 3} curve ac\["Docking Sml"]{4, 4, 2, 1} curve ad\["Docking Med"]{2, 2, 4, 2} max 4 min 0" fullWidth="false" %}