AI-MedCraft: A Strategy-Driven AI Platform for Multi-Objective Molecular Design

Case Study 1 — GDC-0834 Solubility Rescue: Structure-Based Benchmarking Against REINVENT 4 and Ligand-Based Optimization

Case Study 1 includes two workflows: ligand-based and structure-based. Both use the “Improve solubility while maintaining target engagement” strategy.

A. Structure-Based Workflow

1. Project Setup

a) Select “Improve solubility while maintaining target engagement.”b) Choose “Pareto Front.”c) Select “2D fingerprint match” and upload IN.smi from the Case1-structure-based folder.d) Select the reference 2D structure and click Save selections.e) Keep default transformations.

2. Add Additional Scoring Functions

h) Add: • Drug-likeness score (QED) • AI-predicted synthesis ease • Solubility Level (LogS)i) Leave transformations at default.

3. Configure Structure-Based Docking

j) Under structure-based settings, choose “Pharmacophore-Based Docking.”k) Upload BTK.pdb and BTK.ph4 from the Case1-structure-based directory, then click Save & configure.l) Keep default energy and pharmacophore transformations.

4. Training Parameters

m) Click Next.n) Set AIMEDCRAFT Core Parameters as: • Batch size: 256 • Training steps: 1000

Simulations take approximately 2 days with these settings.

5. Run & Analyze

o) Click Next, then Submit.p) After completion, open the job via the Completed Jobs tab.q) View detailed results through Analysis, where all generated molecules can be filtered and visualized.r) Use Pareto Analysis to view multi-objective performance and select top analogues.

B. Ligand-Based Workflow

1. Project Setup

a) Select “Improve solubility while maintaining target engagement.”b) Choose “Pareto Front” as the optimization method.c) Select “2D fingerprint match” and upload IN.smi from the Case1-ligand-based folder.d) Select the reference 2D structure and click Save selections.e) Leave default transformations.

2. Add Pharmacophore Fit

f) Select “3D pharmacophore fit”, click Configure, upload BTK.ph4, then click Save & configure. 

g) Keep transformation settings at default.

3. Add Additional Scoring Functions

h) Add: • Drug-likeness score (QED) • AI-predicted synthesis ease • Solubility Level (LogS)i) Leave default transformations.

4. Training Parameters

j) Click Next.k) Under Show AIMEDCRAFT Core Parameters, set: • Batch size: 256 • Training steps: 2000

Simulations take approximately 12 hours with these settings.

5. Run & Analyze

l) Click Next, then Submit.m) After completion, open the job under Completed Jobs → View Analysis.n) Visualization and filtering options allow users to explore all generated molecules.o) Use Pareto Analysis to visualize the fronts and inspect top candidates.

Case Study 2 — Selectivity-Driven Redesign of Efavirenz to Preserve HIV-1 Reverse Transcriptase Activity While Reducing 5-HT2A Off-Target Engagement

To reproduce Case Study 2, follow these steps:

1. Project Setup

a) Select “Redesign a drug to reduce off-target effects”b) Choose “Pareto Front” as the optimization method.c) Select “2D fingerprint match” as the scoring function and upload the IN.smi file from the Case2 directory.d) Select the reference 2D structure shown in the manuscript table and click Save selections.e) Keep the default transformation settings for the 2D fingerprint function.

2. Add Additional Scoring Functions

f) Add the scoring functions: • Drug-likeness score (QED) • AI-predicted synthesis easeg) Leave their transformations at default values.

3. Configure On-Target Constraints

h) Under On Target, select “On-target Pharmacophore-Based Docking” and click Configure.i) Upload 1ikw.pdb and 1ikw.ph4, then click Save & configure.j) Leave transformation settings for energy and pharmacophore fit as default.

4. Configure Off-Target Constraints

k) Under Off Target, choose “Off-target Pharmacophore-Based Docking.”l) Upload 5ht2a.pdb and 5ht2a.ph4, then click Save & configure.m) Leave transformations at their default values.

5. Training Parameters

n) Click Next.o) Under Show AIMEDCRAFT Core Parameters, set: • Batch size: 128 • Training steps: 2000

Using these settings, simulations require approximately 3 days. Users may choose smaller values for fast, exploratory runs.

6. Run & Analyze

p) Click Next.q) Click Submit. The completed job will appear under Completed Jobs.r) Open the job via Analysis to explore generated compounds, visualizations, and filtering tools.s) To examine multi-objective performance, use Pareto Analysis to view the fronts and inspect top-ranked molecules.

The attached data package contains all input files used to generate the results presented in the manuscript “AI-MedCraft: A Strategy-Driven AI Platform for Multi-Objective Molecular Design.” These include the input PDB structures, pharmacophore definitions, and SMILES files that were supplied to the AI-MedCraft platform.

Following the steps below, any user should be able to reproduce all results and output files reported in the manuscript.