PaperlayerView exact revision on arXiv

Source-linked AI summary

MADDIE: Multi-Agent Drug Discovery Engine

Ye Liu, Reza Averly, Avery Meyer, Frazier N. Baker, Ruoxi Gao, Daniel Adu-Ampratwum, Xia Ning

arXiv:2610.16975v1cs.AI

TL;DR

Drug discovery is costly and slow, while existing agentic systems have difficulty scaling to realistic tasks or adapting beyond predefined workflows. MADDIE coordinates four specialized workers through a central Orchestrator that adapts workflows across campaigns. Across 50 targets, it outperforms state-of-the-art baselines while generating satisfying molecules and synthesis routes.

  • Problem

    Drug discovery is costly and slow, while single-agent systems scale poorly to realistic tasks and multi-agent systems commonly rely on predefined workflows.

    Source: §1 · p. 2

  • Method

    MADDIE uses a central Orchestrator to adapt workflows among four specialized workers, drawing on memory and an evolving repository of candidate molecules.

    Source: §1 · pp. 2–3

  • Results

    Across 50 targets, MADDIE outperforms state-of-the-art baselines and generates 100 satisfying molecules per target with synthesis routes.

    Source: §6 · p. 20

  • Takeaways & Limitations

    The findings support dynamic orchestration of specialized workers as an effective design for autonomous drug discovery.

    Source: §1 · p. 4, §6 · p. 20

    MADDIE’s computational tools remain imperfect proxies for laboratory synthesis and in-vitro experiments, so generated molecules require those validations.

    Source: In vitro validation for MADDIE’s… · p. 18

Abstract

from arXiv · show

Drug discovery remains a resource-intensive and time-consuming process, involving billions of dollars and a decade of development to advance a drug to market. While many in silico tools have accelerated this process, operating them still demands substantial time, effort, and expertise from medicinal chemists. Autonomous systems built on Large Language Models (LLMs) have emerged as a promising solution, yet current designs limit their performance. Existing single-agent systems scale poorly to long-horizon, multi-step, and realistic tasks, whereas multi-agent systems typically execute predefined workflows that cannot capture the iterative, non-linear nature of a real drug discovery campaign. Here, we present MADDIE, a multi-agent drug discovery engine. It comprises a central agent, Orchestrator, and four specialized worker agents: HitIdentifier, LeadOptimizer, Synthesizer, and Evaluator, each representing a critical step in a drug discovery campaign. Under Orchestrator's orchestration and coordination, these worker agents are composed into flexible workflows, enabling MADDIE to adapt its strategy and navigate the drug discovery pipeline end-to-end. Across 50 protein targets, MADDIE produces molecules satisfying key property requirements and their synthesis routes, outperforming existing agentic systems, generative models, and general-purpose LLMs by a wide margin. Our analysis reveals how these agents collaborate and that MADDIE can employ different strategies across targets to achieve the campaign's goal, while still being efficient in time and tokens. Together, they demonstrate MADDIE as a low-cost, effective, and efficient autonomous engine for accelerating drug discovery. Code and dataset is available via https://github.com/ninglab/MADDIE.

1 Introduction

Drug discovery is costly and slow, and existing agentic systems either scale poorly to realistic tasks or rely on predefined workflows. MADDIE combines specialized workers under a central Orchestrator that adapts workflows across campaigns, and the paper reports strong results across 50 targets.

1 Introduction

  • Single-agent systems scale poorly to realistic, complex tasks, while multi-agent systems typically rely on predefined workflows that do not reflect drug discovery’s non-linear process.Finding molecules that bind well while maintaining favorable pharmacokinetic and physicochemical profiles requires many iterations across chemical space.Source: §1 · p. 2
  • MADDIE places a central Orchestrator over four specialized workers: HitIdentifier, LeadOptimizer, Synthesizer, and Evaluator.The Orchestrator dynamically determines workflows, maintains memory and an evolving candidate repository, and monitors campaign progress.Source: §1 · p. 2, Figure 1 · p. 3
  • Flexible workflows let MADDIE adapt its strategy during a campaign, while worker tools—including de novo generative models—support exploration beyond known chemical space.The workers identify hits, optimize candidates, propose synthesis routes, and evaluate molecule quality.Source: §1 · p. 3
  • Across 50 targets, MADDIE outperforms state-of-the-art baselines and generates 100 satisfying molecules per target with synthesis routes.The baselines span drug-design generative models, drug-discovery agents, and generalist LLMs.Source: §1 · pp. 3–4
  • The paper attributes performance to agent collaboration, with distinct worker contributions and the Orchestrator directing effort toward promising candidates; campaigns use different strategies across targets.The authors also report that MADDIE is low-cost and efficient in time and tokens.Source: §1 · pp. 3–4

2 Related Works

Prior work includes LLM-based agents for drug discovery and generative models for molecular design. MADDIE addresses downstream filtering and selection by coordinating generation, assessment, and synthesis tasks.

2.1 Agents for Drug Discovery

  • Drug-discovery agents range from task-specific assistants to end-to-end systems, but their performance is typically limited and does not scale well to more complex, realistic tasks.Examples include CACTUS and TxGemma assistants, AgentD’s early-stage pipeline, and LIDDIA’s autonomous molecule-search workflow.Source: §2.1 · p. 4
  • Multi-agent drug-discovery systems distribute work across specialized agents, but their coordination generally centers on predefined workflows and evaluations often focus on individual case studies.Systems described include DrugAgent, Robin, F.A.D.E., Prompt-to-Pill, PharmAgents, Mozi, and FROGENT.Source: §2.1 · pp. 4–5

2.2 Generative Models for Drug Design

  • Structure-based drug-design models generate ligands conditioned on 3D binding pockets, using approaches such as autoregressive sampling, diffusion, and LLM-based refinement.The passage highlights Pocket2Mol, conDitar-dev, and ELILLM.Source: §2.2 · p. 5
  • Ligand-based drug-design models condition generation on the shape of a ligand that binds the target pocket, with fragment-based, image-to-sequence, and diffusion approaches represented.Examples include SQUID, DESERT, and DiffSMol.Source: §2.2 · p. 5
  • Molecular generation is only an initial step because generated molecules often require substantial manual filtering and selection; MADDIE integrates these tasks under Orchestrator’s coordination.HitIdentifier and LeadOptimizer propose molecules, while Evaluator and Synthesizer assess them.Source: §2.2 · p. 5

3 Experimental Settings

The experiments use a 50-target dataset and assess molecular properties, synthesizability, and joint satisfaction of requirements. Comparisons include drug-discovery agents, generative models, and generalist LLMs under matched generation opportunities.

Dataset

  • The curated dataset contains 50 protein targets with binding-pocket structures and FDA-approved drugs associated with targets to define molecular-property requirements.Targets span major therapeutic areas; 42 come from existing datasets and eight from OpenTargets, and the targets are excluded from the CrossDocked2020 training set.Source: Dataset · pp. 5–6

Evaluation Metrics

  • Evaluation covers binding affinity, drug-likeness, synthesizability estimates, toxicity indicators, medicinal-chemistry filters, and whether a synthesis route can be generated.A molecule counts as satisfying only when it jointly meets every requirement; thresholds are based on curated FDA-approved drugs.Source: Evaluation Metrics · p. 6

Baselines

  • Baselines include drug-discovery agents, structure- and ligand-based generative models, and open-source generalist LLMs; generated molecules are filtered against requirements not assessed during generation.The comparisons include LIDDIA, AgentD, Prompt-to-Pill, five generative models, Qwen3-30B, and Nemotron-3-Ultra.Source: Baselines · pp. 7–8
  • Baselines receive matched generation opportunities: methods unable to produce a fixed output count are instructed to generate as many intermediate molecules as MADDIE produces per target.Intermediate molecules include any produced by HitIdentifier, LeadOptimizer, or Synthesizer, whether or not they become outputs.Source: Baselines · p. 8

Implementation Details

  • Qwen3-30B is the backbone LLM for the Orchestrator and all four worker agents, deployed locally using vLLM with a 64K-token context window.Decoding uses temperature 0.7 and top-p 0.8.Source: Implementation Details · p. 8

4 Experimental Results

MADDIE delivers the requested satisfying molecules across all 50 targets, while its agents contribute through adaptive evaluation, optimization, and synthesis. Campaign analyses show varied workflows, and the PSEN1 case study and efficiency results illustrate the system’s performance and operating cost.

4.1 Overall Performance

  • Only MADDIE delivered the requested 100 satisfying molecules on all 50 targets; closest baselines yielded 56.7 (conDitar, 49 targets), 33.1 (Pocket2Mol, 50), and 44.2 (AgentD, 35).Agentic systems still miss measured requirements, while filtering can enforce pass rates but may retain few or no molecules; generalist LLMs also produce fewer satisfying molecules or cover fewer targets.Source: Table 2 · p. 9, §4.1 · pp. 9–10

4.2 Ablation Study

  • Removing Synthesizer reduced satisfying molecules to 40.2; removing LeadOptimizer or Orchestrator raised intermediate-molecule exploration to 1,327.7 or 1,743.8, respectively.Synthesizer supplies synthesizable analogs, LeadOptimizer improves candidates, and Orchestrator uses feedback to focus exploration; without LeadOptimizer, the variant covers 49 rather than 50 targets.Source: §4.2 · p. 10, LeadOptimizer · pp. 10–11, Table 3 · p. 11

4.3.1 Overall Transition Probabilities

  • Evaluator acts as a quality checkpoint, assessing generated or optimized molecules before further processing; after evaluation, campaigns distribute next steps across worker agents rather than following a fixed path.From Evaluator, the probabilities of moving to HitIdentifier, LeadOptimizer, and Synthesizer are 0.22, 0.20, and 0.32; campaigns finish only after Evaluator or Synthesizer is called.Source: §4.3.1 · pp. 11–12, Figure 2 · p. 12

4.3.2 Molecule Flow Across Agents

  • HitIdentifier sends 541.4 molecules to Evaluator but yields 22.5 satisfying molecules; LeadOptimizer sends 241.5 and yields 33.4, while Synthesizer sends 260.9 and yields 72.8.HitIdentifier explores broadly, while LeadOptimizer and Synthesizer exploit candidates; their contributions are linked because Synthesizer can generate analogs from previously assessed molecules.Source: §4.3.2 · pp. 12–13, Figure 3 · p. 13

4.3.3 Agent Execution Across Campaigns

  • Across campaigns, MADDIE changes delegation strategies; on ABCD1_ALD it shifted to multigenerate-then-optimize, accumulating 86 property-satisfying molecules before Synthesizer helped raise the total from 44 to 136.The five other representative targets also use differing sequences of generation, optimization, and synthesis rather than a fixed workflow.Source: §4.3.3 · pp. 13–15, Figure 4 · p. 14

4.4 Case Study on PSEN1

  • For PSEN1, MADDIE proposed 16 molecules with in-silico properties better than either the reference ligand or Nirogacestat; PSEN1-01 exceeded both across all six properties.The reported binding rationale includes hydrogen bonds to the PAL and GxGD motifs and nine hydrophobic contacts, which the authors link to improved binding while retaining other key properties.Source: §4.4 · pp. 15–17, Figure 5 · p. 16

4.5 Computational Efficiency and Cost

  • >50% of targets required fewer than 1,000 intermediate molecules; campaigns averaged 1,043.8 molecules, 3.3 hours, and 21,125 output tokens.For the hardest 10% of targets, exploration exceeded 1,500 molecules and reached 2,601; averaged over satisfying molecules, generation cost roughly 2 minutes and 211 output tokens each.Source: Table 4 · p. 17, §4.5.1 · p. 17, Table 5 · p. 18

4.5.2 Computational Cost

  • Evaluator received 5.10 of the 10.0 average delegations per target, about as many as the other three agents combined; delegation count alone did not determine campaign duration.ADAMTS5 and HDAC8 took 1.2 and 11.6 hours despite differing by only two delegations, indicating that duration depended on work within delegations.Source: §4.5.2 · p. 18

5 Discussions

The discussion identifies experimental validation and expert input as routes toward assessing and steering MADDIE’s in-silico campaigns, while laboratory integration raises additional safety requirements.

In vitro validation for MADDIE’s generated molecules

  • In-silico tools remain imperfect proxies for laboratory synthesis and in-vitro experiments, which are indispensable next steps for validating generated molecules’ real-world potential.The authors identify scaling and incorporating in-vitro validation into MADDIE as future research directions.Source: In vitro validation for MADDIE’s… · p. 18

Human-in-the-loop design for MADDIE

  • Domain experts currently verify generated-molecule quality rather than drive generation; future feedback could steer Orchestrator toward preferred candidates and away from undesirable structures.The proposed human-in-the-loop framework could use occasional expert input to shape campaigns.Source: Human-in-the-loop design for MADDIE · p. 19

MADDIE as a fully autonomous and self-driving system with laboratory integration

  • Integrating automated synthesis and in-vitro platforms could close the loop, but physical-world deployment requires safeguards for handling synthesized molecules and preventing hazardous outputs.Candidate selection would also become a resource-allocation decision under synthesis budgets, with Orchestrator drawing on feasibility and predicted-property assessments.Source: MADDIE as a fully autonomous and… · pp. 19–20

6 Conclusions

MADDIE combines adaptive multi-agent campaigns with molecular generation, optimization, evaluation, and synthesis planning. Across 50 targets, it meets requested property requirements and provides synthesis routes, while reporting target-specific results and low campaign cost and time.

6 Conclusions

  • Across 50 targets, MADDIE generates 100 satisfying molecules per target and synthesis routes, outperforming state-of-the-art generative models, drug-discovery agents, and generalist LLMs.The conclusion attributes performance to distinct worker-agent contributions and Orchestrator’s focus on promising candidates, reducing the chemical space explored.Source: §6 · p. 20
  • MADDIE adapts its strategies across campaigns, including on a challenging target with no existing marketed FDA-approved drugs.The conclusion reports intelligent strategy adaptations for that target.Source: §6 · p. 20
  • On PSEN1, MADDIE’s PSEN1-01 has in-silico properties outperforming the reference ligand and Nirogacestat, with favorable binding interaction at the challenging target site.The conclusion identifies PSEN1 as a critical target associated with Alzheimer’s disease.Source: §6 · p. 20
  • One campaign produces roughly 21K output tokens and takes an average of 3.3 hours, which the authors report as strong performance at low cost and time.The authors envision MADDIE as a foundation for autonomous systems that could accelerate the timeline and cost of bringing new drugs to patients.Source: §6 · p. 20

7 Methods

MADDIE decomposes drug discovery into specialized worker tasks coordinated by Orchestrator, which uses campaign information to choose actions and candidates. Its workers generate and refine molecules, assess synthesis routes, and evaluate molecular properties.

7 Methods

  • Orchestrator coordinates four specialized workers: HitIdentifier supplies candidates, LeadOptimizer refines them, Synthesizer proposes routes and synthesizable analogs, and Evaluator assesses drug-relevant properties.The workers operate in separate workspaces, while Orchestrator alone accesses each workspace for a campaign-wide view.Source: §7 · p. 21

7.1 Orchestrator

  • Orchestrator chooses its next worker and candidates using campaign progress and worker profiles, making decisions that adapt to evolving campaign needs.It delegates subtasks and processes resulting molecules, property values, or synthesis routes to update which candidates meet requirements.Source: §7.1 · pp. 21–22

Memory module

  • A dedicated memory module preserves campaign artifacts, worker reports, and prior decisions, while summarizing and pruning other records when context limits are reached.This provides Orchestrator with compact, decision-relevant campaign history.Source: Memory module · p. 22

Satisfying candidate repositories

  • Repositories P and S track, respectively, property-satisfying candidates and those also verified synthesizable; Orchestrator uses them to choose actions and stops when S reaches its target or the action budget expires.Near the budget limit, it routes candidates in P to Synthesizer so they can still enter S.Source: Satisfying candidate repositories · pp. 22–23

7.2.1 HitIdentifier

7.2.2 LeadOptimizer

7.2.3 Synthesizer

7.2.4 Evaluator

  • Evaluator supplies property assessments—including binding affinity, drug-likeness, synthesizability, ADMET, and structural liabilities—that ground Orchestrator’s candidate decisions.Its rapid heuristic synthesizability score complements Synthesizer’s route-based assessment by providing a quantitative indication of ease of synthesis.Source: §7.2.4 · pp. 26–27

Ethics Statement

MADDIE operates purely in silico and cannot execute synthesis or produce physical molecules, but the authors acknowledge that it may generate unsafe or harmful molecules.

Ethics Statement

  • MADDIE generates molecules meeting specified property requirements from natural-language prompts, but not all generated molecules are safe and the system could be used to generate harmful molecules.The authors recommend responsible domain-expert supervision and compliance with applicable safety guidance and professional practices.Source: Ethics Statement · p. 27
  • MADDIE is designed without the capability to execute synthesis or produce physical molecules, limiting its operation to in-silico drug discovery.The authors present this as a step to minimize potential negative impact, not as evidence that generated molecules are safe.Source: Ethics Statement · p. 27

A Dataset and Annotations

The dataset pairs curated protein targets with disease annotations and current FDA-approved small-molecule drug associations. Target structures and binding pockets are selected using experimentally validated ligand-bound structures.

A Dataset and Annotations

  • The authors provide protocols for dataset curation.The protocols are available at the referenced location.Source: A Dataset and Annotations · p. 36
  • Targets were curated from LIDDIA, CrossDocked2020, and OpenTargets, excluding targets in the CrossDocked2020 training set, then annotated with disease associations from OpenTargets and GeneCards.Ligand-bound experimentally validated structures from RCSB were used to extract binding pockets; target annotations are listed in Table A1.Source: A Dataset and Annotations · p. 36
  • Current marketed FDA-approved small-molecule drug associations were identified for 22 dataset targets by mapping ChEMBL and DrugCentral records to Drugs@FDA active ingredients.Products were filtered as of July 31, 2026, and collapsed to one active-family representative per target; Table A2 lists the associations.Source: A Dataset and Annotations · p. 36

B Property Requirements

Property thresholds were calibrated using curated FDA-approved drugs to balance individual-property pass rates with the joint requirements. The final requirements and reference-drug pass rates are reported in Table A3.

B Property Requirements

  • At least 80% of reference drugs were required to pass each property, while at least 40% had to pass all requirements simultaneously.Initial per-property thresholds left fewer than 30% passing jointly, so the authors manually relaxed thresholds while retaining around 80% per-property passage.Source: B Property Requirements · p. 36
  • Table A1 catalogs the protein targets in the MADDIE dataset.Source: Table A1 · p. 37

C Baseline Selection Discussion

The authors selected publicly implemented agentic drug-discovery systems as baselines because incomplete or unavailable code made evaluation difficult. LIDDIA, AgentD, and Prompt-to-Pill were chosen as representatives.

C Baseline Selection Discussion

  • Incomplete or unavailable code for several drug-discovery agent systems made them difficult to evaluate, so the authors restricted agentic baselines to systems with public implementations.Source: C Baseline Selection Discussion · p. 37
  • Table A2 lists FDA-approved drug associations for targets in the MADDIE dataset, and Table A3 reports property requirements and reference-drug pass rates.Source: Table A2 · p. 38, Table A3 · p. 39
  • LIDDIA, AgentD, and Prompt-to-Pill were selected as representative agents for drug discovery.Source: C Baseline Selection Discussion · p. 38

D Baseline Implementation

The baseline implementations use public code and are configured to align model or evaluation settings with MADDIE where described. The appendix details the agent, generative-model, and generalist-LLM setups.

D Baseline Implementation

D.1 Agents for Drug Discovery

  • Agent baselines use Qwen3-30B and aligned computational evaluation tools; LIDDIA and Prompt-to-Pill replace docking or ADMET components, while AgentD uses adjusted pool, refinement, filtering, and recovery settings.AgentD runs may retry with an alternate disease association, and its paid web search and Boltz structure prediction are disabled.Source: §D.1 · pp. 39–40

D.2 Generative Models for Drug Design

  • Pocket2Mol’s beam size, generation steps, and requested samples were increased to obtain enough valid molecules for evaluation.Beam size rose from 300 to 1,000, maximum steps from 50 to 100, and requested samples from 100 to 2,000.Source: §D.2 · p. 40
  • TargetDiff and conDitar increased requested samples from 100 to 2,000 to retain enough chemically valid molecules after reconstruction.Source: §D.2 · p. 40
  • DiffSMol was supplied dataset ligands and run with pocket guidance at threshold 0.5, with samples increased from 100 to 2,000.Remaining settings followed the released configuration.Source: §D.2 · p. 40
  • ELILLM-rand was used to measure exploration without bias toward a particular structure-based generation model, with up to 2,000 iterations and a 7,200-second docking timeout.The setup allowed up to 10,000 generation opportunities per pocket.Source: §D.2 · pp. 40–41

D.3 Generalist LLMs

  • Generalist LLMs received a prompt requesting diverse, novel target binders meeting specified property thresholds, stopping at 100 qualifying molecules and returning only SMILES strings.The same prompt format was used for Nemotron-3-Ultra and Qwen3-30B.Source: §D.3 · p. 41

E MADDIE’s Implementation

This section describes MADDIE’s implementation, including how the Orchestrator coordinates budgeted specialist workflows and how molecule acceptance and optimizer selection are guided by evaluation and performance records.

E MADDIE’s Implementation

  • Campaigns receive 100 actions; HitIdentifier gets 3–10 iterations of 100 molecules, LeadOptimizer generates 10 derivatives per candidate with at most two executions, and Synthesizer searches 30 seconds per molecule up to 500 iterations.Orchestrator compacts working memory above 32K tokens and delegates molecules for synthesis once fewer than eight actions remain.Source: E MADDIE’s Implementation · pp. 41–42

E.2.4 Evaluator Profile

  • Evaluator critically reviews molecules and computes seven property values, excluding Syn; MADDIE uses AutoDock for HitIdentifier affinity prediction and UniDock for Evaluator docking.UniDock supports fast docking computation, including GPU use, while HitIdentifier needs affinity prediction for a given pose rather than docking.Source: E MADDIE’s Implementation · p. 42, §E.2.4 · p. 52

E.1.1 Orchestrator System Prompt

  • The Orchestrator analyzes the molecule pool, delegates exploration or optimization, and verifies each new batch; after an EXPLOIT cycle, it requires EXPLORE before optimizing again.It records qualifying EXPLORE results before choosing the next action and directs docked candidates that fail other constraints to optimization.Source: §E.1.1 · p. 44

E.2.1 HitIdentifier Profile

  • HitIdentifier generates candidate molecules against a target and can use SBDD, LBDD, or both, with repeated or alternating strategies to sample chemical space differently.Its delegation specifies a target-structure file, a hit count, a round count, and the selected search strategy.Source: §E.2.1 · p. 50

E.2.2 LeadOptimizer Profile

  • LeadOptimizer improves existing hits or leads, including ADMET toxicity scores, and supports multi-property optimization of up to three properties at once.It requires a CSV input file with a SMILES column; its output molecules must be reevaluated.Source: §E.2.2 · p. 51

E.2.3 Synthesizer Profile

  • Synthesizer searches for routes and synthesizable analogs, using MEEA and PrexSyn outputs to populate a synthesis-confirmed pool or send analogs for evaluation.It is triggered when the candidate pool can fill the remaining target count, reaches 50 molecules, or fewer than eight iterations remain; the pool is cleared after synthesis processing.Source: §E.2.3 · pp. 51–52

E.2.4 Evaluator Profile

  • The Evaluator is the sole source of docking and ADMET results, and molecules count toward the goal only after passing every required constraint.It evaluates files rather than raw SMILES and can assess physicochemical properties, molecular filters, docking, and selected ADMET properties. Molecule and protein files must be accessible in its own subdirectory, using absolute paths.Source: §E.2.4 · p. 53

E.3 LeadOptimizer’s Records

  • LeadOptimizer uses records from 11 MolOpt optimizers, tested on ZINC molecules for tasks such as increasing QED and lowering toxicity-related scores.Each structured JSON record stores the objective, optimizer, number of molecules meeting the request, runtime, and related metadata for optimizer selection.Source: §E.3 · p. 53

F.1 Property Value Analysis

MADDIE’s output molecules have property values comparable to strong baselines, and its diversity analysis reports competitive ranks alongside structurally diverse outputs. Additional examples show varied trajectories, including recovery from a stalled campaign.

F.1 Property Value Analysis

  • Table A4 reports average property values across output molecules from all targets; Div is average diversity per target, measured using Tanimoto similarity between Morgan fingerprints.Source: §F.1 · pp. 53–54, Table A4 · p. 54
  • 6.3 average rank puts MADDIE near AgentD (6.1), Pocket2Mol (6.0), TargetDiff (6.4), and conDitar (6.7); no method consistently outperforms the others.LIDDIA has the best average rank on property values, although relatively few of its molecules have those properties.Source: §F.1 · p. 54
  • QED (0.646), Mut (0.217), CTx (0.224), and hERG (0.535) outperform reference ligands and drugs; Vina (-7.897 vs -7.828) and SAS (3.563 vs 2.961) are on par with approved drugs.Source: §F.1 · p. 54

F.2 Additional Representative Trajectories Discussion

  • ABO_BGAT generates no synthesizable analogs and relies on HitIdentifier and LeadOptimizer; HTR2A recovers after more than 10 Evaluator calls fail to locate HitIdentifier artifacts.Source: §F.2 · p. 55, Figure A1 · p. 55
  • On PSEN1, MADDIE-generated molecules exceed either the reference ligand or Nirogacestat on all in silico properties defined in Section 3.Source: Figure A2 · p. 56

G Additional Case Studies

The ACHE and BCHE case studies describe generated molecules with favorable comparisons against reference ligands and approved drugs, including a BCHE candidate that passes filters Rivastigmine fails.

G Additional Case Studies

Related papers

Loading 2610.16975v1…