AI-DRIVEN DRUG DESIGN (AIDD)

Accelerated by structure-first AI.

AI-DRIVEN DRUG DESIGN

Accelerated by structure-first AI.

AI-DRIVEN DRUG DESIGN

Accelerated by structure-first AI.

Structure-first AI

AI for small-molecule design has advanced from rule-based enumeration to generative models and reinforcement learning (RL). Many platforms still operate as black boxes trained primarily on public data, which can limit novelty and produce brittle results.

CrystalsFirst® takes a structure-first path. Our FragAI platform is a 3D-aware generative AI trained on experimental protein–ligand complexes produced in-house, then optimized in a closed loop with wet-lab validation. The result is faster convergence on optimized and IP-ready molecules because learning is grounded in real structural interactions.

In-house expertise and technology

3D-aware generative modeling — Unlike many AI tools that only work with chemical strings (SMILES), our FragAI platform works directly with 3D protein–ligand structures. This means the molecules it proposes are already consistent with how they should bind in the pocket. In practice, FragAI has delivered real, testable molecules. For example on our in-house target, over half of the suggested compounds were synthetically accessible, with multiple binders confirmed crystallographically and in biophysical assays.

Design rules and constraints

FragAI doesn’t just “imagine molecules.” It designs compounds under realistic medicinal chemistry rules:

It respects property windows (e.g., MW, lipophilicity), scaffolds, and synthetic feasibility.
It ensures smooth chemical changes to support SAR exploration.
Future upgrades include graph- and diffusion-based models to propose novel chemotypes at the scaffold level.

Reinforcement learning for chemistry

To make the AI better aligned with medicinal chemistry practice, we use reinforcement learning:

The system learns from human preferences and prior project outcomes.
It can incorporate historical data and simulated feedback to guide designs.
We control the optimization with “reward functions” that balance potency, selectivity, novelty, IP space, and synthetic tractability.
Our own optimization approach for Preference Optimization calibrates the model to stay productive and continues to generate high-quality, relevant molecules.

Robust operations and validation

Every model run is tracked.
Benchmarks are run against internal and external reference sets to validate performance.
Our in-house GPU/CPU infrastructure allows campaigns to run in parallel.

Closed-loop design cycle of the MAGNET Platform

The AI is not left on its own — every campaign runs in a closed loop:

The models suggest compounds.
Selected molecules are synthesized externally.
Molecules are tested experimentally (binding, co-crystallography, activity).
Results are fed back into FragAI to improve the next design cycle.

This make–test–learn loop ensures that proposals rapidly converge on potent, selective, and novel chemotypes that are supported by both AI predictions and experimental proof.

Project Execution

Data and objective definition
We assemble high-quality structural data (X-ray/Cryo-EM) and define multi-objective goals: potency, selectivity, ADME, novelty/IP, and synthetic accessibility.
Model setup and benchmarking
Baseline generative and scoring models are configured; internal benchmarks verify pose fidelity and property prediction before large-scale generation.
Constrained generative design
FragAI generates diverse chemotypes consistent with 3D binding requirements. Latent-space continuity supports controlled modifications; graph/diffusion components (where applicable) propose topology changes.
RL fine-tuning
Designs are optimized via actor–critic/off-policy RL, RLHF, and curriculum schedules. Rewards combine predicted affinity/free energy proxies, selectivity, liabilities, 3D alignment, novelty and synthetic routes.
Triage and synthesis hand-off
Shortlists include retrosynthetic suggestions and vendor availability; synthesis partners are selected programmatically to minimize cycle time.
Wet-lab validation and learning
Generated molecules are tested; co-structures and assay data feed back to continually improve the model (closed-loop of the MAGNET cycle). Demonstrated on internal and partner campaigns across protein classes.

Deliverables

Prioritized design set (top N molecules) with 3D poses, rationale, novelty/IP analysis, and synthetic accessibility.
Model artifacts & audit trail: configs, benchmarks, and versioned datasets for regulatory/partner diligence.
Validation package linking AI predictions to experimental outcomes (structures, biophysics, activity).
Scale & scheduling plan to run additional waves on our GPU/CPU infrastructure and expand to adjacent targets.

FAQ: AIDD at CrystalsFirst

How is AIDD different from CADD?

CADD prioritizes physics-based docking and simulations; AIDD generates novel molecules de novo, then optimizes them with multi-objective RL and real feedback. At CrystalsFirst, AIDD is structure-first and tightly coupled to experiments, improving reliability and IP differentiation.

What makes your generative AI “3D-aware”?

FragAI learns from experimental protein–ligand complexes, so it reasons about poses and interactions in 3D. This improves binding-mode plausibility and speeds up convergence to active, selective chemotypes.

How do you prevent “clever, unsynthesizable” designs?

We enforce constrained generation, retrosynthesis filters, and reward terms for synthetic accessibility. Designs are benchmarked, then triaged to synthesis partners; prior campaigns showed high synthesis success and multiple validated binders.

What scale can you support?

Our in-house GPU/CPU infrastructure supports parallel campaigns. We can deploy the models 100+ GPUs on cloud infrastructure.

How do you keep models from drifting or collapsing?

We use in-house preference optimization-based RLHF, actor–critic/off-policy training, and learning curricula; continuous benchmarking and wet-lab feedback stabilize training and keep top-tier quality high across iterations.