Orchestrate Agent
Plans the task graph, delegates work, tracks state, and controls feedback loops.
Flagship platform / FROGENT
Coordinate the full discovery loop through shared scientific context
FROGENT is EasternDawn's full-process multi-agent drug design platform. Four collaborating agents plan, retrieve, generate, evaluate, and refine work through a common project state and a Model Context Protocol (MCP)-connected capability layer.
- Collaborating agents
- 4
- Discovery capabilities
- 11
- Benchmark families
- 8
- End-to-end case studies
- 3
01 / Abstract
A campaign is one inspectable scientific record
FROGENT turns a high-level goal into a task graph and keeps retrieved evidence, candidate structures, scores, artifacts, failures, and agent handoffs in shared context. Specialized engines can then contribute modality-specific work without losing the scientific question that motivated it.
To our knowledge, the first multi-agent drug design framework that integrates and automates end-to-end discovery workflows for both small-molecule and peptide therapeutics.
Evidence boundary. The priority statement is preserved exactly from the supplied manuscript. Benchmark and case-study results describe computational workflows and do not establish experimental efficacy, safety, affinity, or clinical performance.
02 / Architecture
Four agents, one Global Context, and a connected capability surface
The architecture separates responsibility while preserving a common evidence trail. Each agent can call models, databases, tools, and evaluators through a shared MCP interface.
Retrieve Agent
Gathers literature, database records, structures, and project evidence while preserving provenance.
Forge Agent
Generates and optimizes candidates, then connects design decisions to synthesis planning.
Gauge Agent
Scores candidates, exposes diagnostics, and returns evaluation evidence to the next decision.
03 / End-to-end workflow
Every stage returns artifacts and diagnostics to the next decision
- 01
Define
Record the scientific goal, modality, constraints, thresholds, and human decision points.
- 02
Retrieve
Collect task-relevant evidence and retain the source attached to each claim or artifact.
- 03
Design
Generate candidates through modality-specific engines under explicit task constraints.
- 04
Evaluate
Run quantitative evaluators, preserve component scores, and expose failure diagnostics.
- 05
Refine
Return evidence to the shared context so the next proposal responds to the current record.
- 06
Deliver
Assemble ranked candidates, synthesis plans, provenance, limitations, and a consolidated report.
04 / Connected engines and agents
FROGENT delegates scientific depth to product-specific systems
Each product retains its own method, experiments, evidence, limitations, and visual identity while contributing a focused capability to the broader workflow.
Small molecule
TrioMol2
All-target molecular generation and target-memory-guided search.
Targeted degradationTrioPROTAC
Fixed-anchor, structure-aware degrader design.
PeptideTrioPep
Structure-feedback peptide design demonstrated for xanthine oxidase (XOD).
RNATrioIRES
Structure-constrained RNA prioritization.
DNATrioDNA
Cell-type-aware regulatory DNA design.
Scientific agentPio
Evidence-aware peptide optimization and review.
05 / Evaluation and limits
Architecture, task benchmarks, and case studies remain separate evidence layers
The supplied manuscript reports comparisons with six progressively capable baselines across eight benchmark families and documents three end-to-end computational design case studies. Task definitions and scoring scales remain specific to each evaluation.
Inspect the FROGENT evidence index →
Limitations. Automated coordination can improve coverage and traceability while still inheriting model error, database gaps, evaluator approximation, and task-definition bias. Consequential decisions require expert review and prospective experimental validation.