AI-Powered Ecosystem for Drug Discovery

Integrated drug discovery ecosystem for small molecules and peptides that blends together human and artificial intelligence. Backed up by state-of-the-art AI technologies, powerful data augmentation, and experience in 40+ diverse commercial projects.

Multilevel AI-driven Drug Discovery Infrastructure

Our drug discovery infrastructure consists of maintained, four hierarchical levels that facilitate seamless AI-driven preclinical drug discovery pipeline:

  • Level 1: Data Infrastructure
    Advanced data acquisition, generation and augmentation.
  • Level 2: AI Model Suite
    Predictive and generative AI models for all major drug discovery tasks.
  • Level 3: Drug Discovery Workflows
    Dedicated workflows for different drug modalities and target classes.
  • Level 4: Orchestration with LLM Agents
    Blending human expertise with AI for project planning, decision-making, and dynamic adaptation.

LAYER 01

Data Infrastructure

The data layer, which is responsible for automated data acquisition, generation, aggregation, analysis and augmentation, is the foundation of the system. It is continuously updated from a variety of biological, chemical, clinical and computational data sources:

  • Public and proprietary structured databases
  • Unstructured data such as scientific literature and patents
  • In house computational and experimental results
  • Iterative experimental feedback in the course of project execution
  • AI-driven data augmentation for enriching datasets
  • Human-in-the-loop feedback to refine AI predictions

This layer provides both generalized and context-aware project-specific insights, which is used for both training and validating AI models and for guiding human comprehension of the project strategy and results.

01

LAYER 02

Core AI Models

Our AI-driven platform is powered by dozens of specialized AI models, each fine-tuned for different aspects of drug discovery. Building on the continuously evolving data layer, these models offer broad applicability and project-specific precision. Of these, four flagship models power predictive and generative tasks:

  • ArtiDock: AI docking for high-throughput and accurate prediction of the binding poses for small molecules and peptides.
  • DeepTAG: PPI prediction for protein complexes even in the absence of homology templates and design of ternary complexes and proximity engagers.
  • Multitask ADMET profiling using 80+ endpoints for MPO.
  • 3DProtDTA: Generative AI for small molecules and peptides design directly in the binding pocket.

These models are continuously trained, fine-tuned, tested and deployed in the project-specific workflows, providing unmatched precision, scalability and customization.

02

LAYER 03

Modality and Target-Specific Workflows

This layer constructs optimized workflows for specific targets and modalities by providing concrete protocols, decision trees, and customizations — each one built upon an ensemble of specialized AI models. Current capabilities include:

  • Small molecules workflow
  • Linear and cyclic peptide workflow
  • PPI targeting workflow
  • Covalent inhibitors workflow
  • Molecular glues and proximity engagers workflow
  • Specific workflows for target classes: GPCRs, kinases, ionic channels, etc.

Each workflow is controlled by the LLM layer and is configured with optimal parameters, generation strategies, benchmarks, and execution pipelines, is fully customizable to the specific project demands, and has been validated through experimental testing in real drug discovery projects.

03

LAYER 04

Orchestration and supervision with LLM Agents

This layer is a mission control center of our infrastructure. Our LLM Agents supervise the project's planning, execution, decision-making and result assessment. They provides:

  • Workflow orchestration — ensuring seamless operation across all platform layers, by setting key parameters, filters and run-time options for each step of the drug discovery process.
  • Insights and recommendations — analyzing project data to guide decisions making at every stage of the project.
  • Research assistance — RAG-based literature analysis and hypothesis generation.
  • Drug discovery strategy planning — building and adapting strategies and approaches according to evolving project needs.

The LLM Agents are aware of biologically, chemical and clinical context, offering real-time support for both computational, experimental and management teams.

04

LAYER 01

Data Infrastructure

The data layer, which is responsible for automated data acquisition, generation, aggregation, analysis and augmentation, is the foundation of the system. It is continuously updated from a variety of biological, chemical, clinical and computational data sources:

  • Public and proprietary structured databases
  • Unstructured data such as scientific literature and patents
  • In house computational and experimental results
  • Iterative experimental feedback in the course of project execution
  • AI-driven data augmentation for enriching datasets
  • Human-in-the-loop feedback to refine AI predictions

This layer provides both generalized and context-aware project-specific insights, which is used for both training and validating AI models and for guiding human comprehension of the project strategy and results.

LAYER 02

Core AI Models

Our AI-driven platform is powered by dozens of specialized AI models, each fine-tuned for different aspects of drug discovery. Building on the continuously evolving data layer, these models offer broad applicability and project-specific precision. Of these, four flagship models power predictive and generative tasks:

  • ArtiDock: AI docking for high-throughput and accurate prediction of the binding poses for small molecules and peptides.
  • DeepTAG: PPI prediction for protein complexes even in the absence of homology templates and design of ternary complexes and proximity engagers.
  • Multitask ADMET profiling using 80+ endpoints for MPO.
  • 3DProtDTA: Generative AI for small molecules and peptides design directly in the binding pocket.

These models are continuously trained, fine-tuned, tested and deployed in the project-specific workflows, providing unmatched precision, scalability and customization.

LAYER 03

Modality and Target-Specific Workflows

This layer constructs optimized workflows for specific targets and modalities by providing concrete protocols, decision trees, and customizations — each one built upon an ensemble of specialized AI models. Current capabilities include:

  • Small molecules workflow
  • Linear and cyclic peptide workflow
  • PPI targeting workflow
  • Covalent inhibitors workflow
  • Molecular glues and proximity engagers workflow
  • Specific workflows for target classes: GPCRs, kinases, ionic channels, etc.

Each workflow is controlled by the LLM layer and is configured with optimal parameters, generation strategies, benchmarks, and execution pipelines, is fully customizable to the specific project demands, and has been validated through experimental testing in real drug discovery projects.

LAYER 04

Orchestration and supervision with LLM Agents

This layer is a mission control center of our infrastructure. Our LLM Agents supervise the project's planning, execution, decision-making and result assessment. They provides:

  • Workflow orchestration — ensuring seamless operation across all platform layers, by setting key parameters, filters and run-time options for each step of the drug discovery process.
  • Insights and recommendations — analyzing project data to guide decisions making at every stage of the project.
  • Research assistance — RAG-based literature analysis and hypothesis generation.
  • Drug discovery strategy planning — building and adapting strategies and approaches according to evolving project needs.

The LLM Agents are aware of biologically, chemical and clinical context, offering real-time support for both computational, experimental and management teams.

PharmaSphere Drug Discovery Ecosystem

Our integrated drug discovery ecosystem for small molecules and peptides blends together human and artificial intelligence and covers all major stages of the preclinical drug design pipeline. It is backed up by state-of-the-art AI technologies, powerful data augmentation techniques, and experience in 40+ diverse commercial projects.

Knowledge Engine

Multimodal knowledge graph, which is maintained, searched and updated by LLM RAGs and provides information about chemical, biological and clinical contextual data on all preclinical drug discovery stages:

  • Data from structured databases.
  • Data from scientific papers, patents, and clinical studies.
  • Multiple kinds of contextual relations with confidence scores.
  • Capable of advanced target identification and assessment.

Small Molecules Platform

De novo AI-driven design of small molecules by leveraging key interactions related to biological activity with optimization of over 80 drug properties:

  • x3 faster than the traditional iterative approach due to self-learning AI models.
  • Dedicated AI workflows for kinases, GPCRs, ion channels, and enzymes.
  • Identification and targeting of allosteric, hidden, transient, and cryptic pockets.
  • Reliable discrimination between agonists, antagonists, and neutral binders.
  • Selectivity against highly similar isoforms and mutants.
  • Optimization of the largest set of ADMET endpoints in the market.

Peptides Platform

AI-guided de novo design and optimization of linear and cyclic peptides against challenging targets, including “undruggable” protein-protein interactions:

  • Unlimited diversity is only constrained by synthetic capabilities.
  • Any non-natural amino acids, non-peptide building blocks, and linkage chemistry.
  • Seamless integration with phage and RNA display techniques.
  • Optimization of peptide activity, permeability, and oral availability.
  • Peptide-to-small-molecule workflow to transform peptide binders to drug-like molecules with superior ADMET properties.

Proximity Inducers Platform

Engineering ternary complexes to transform structurally unresolved native and induced PPIs into druggable targets:

  • Predicting and targeting PPIs without known structures and homology templates.
  • Designing warheads and linkers for degraders and proximity engagers.
  • Designing molecular glues and PPI disruptors.
  • Designing modulators of the lipid-mediated interaction of membrane proteins.

Internal Programs

co-dev Program
Start date
Target ID &
validation
Early discovery
Lead optimization
IND-enabling
Phase 1
Small molecules
RAI-001 (Nephrology)
Apr 2023
RAI-002 (Oncology)
Oct 2023
RAI-004 (Inflammation)
May 2024
Peptides
RAI-003 (Cardiology)
Jan 2024