Skip to content

intersection-generative-molecules.md

Latest commit

 

History

History
118 lines (82 loc) · 5.71 KB

intersection-generative-molecules.md

File metadata and controls

118 lines (82 loc) · 5.71 KB
Raw

Generative Models × Molecular Discovery

← Back to Wiki Index

How VAEs, diffusion models, and GFlowNets are designing the molecules of the future

The Convergence

Drug discovery, protein engineering, and materials science share a common challenge: navigating vast combinatorial spaces to find molecules with desired properties. Generative models — originally developed for images and text — have become the engine of molecular design.

VAEs: The First Generation of Molecular Generators

Gómez-Bombarelli et al. (2018) launched the field by encoding molecules (as SMILES strings) into a continuous latent space using a VAE, then optimizing in that latent space for desired properties (drug-likeness, solubility, synthetic accessibility).

Why this matters: Chemistry operates in a discrete space (atoms and bonds). VAEs transform this into a continuous manifold where gradient-based optimization becomes possible. Instead of testing molecules one by one, you can smoothly navigate chemical space.

Connection to brain science: The same VAE architecture used for molecular design (Gómez-Bombarelli) is used for normative brain modeling (Mai Ho 2025, Walter 2020). In both cases, the VAE learns "what's normal" — normal molecules that are drug-like, or normal brain structure at a given age. Abnormality is detected as departure from the learned distribution.

GFlowNets: Diverse Candidate Generation

The key limitation of standard optimization: it finds a single optimum. In drug discovery, you want many diverse candidates because:

  1. The top-scoring molecule might fail clinical trials
  2. Patent landscapes require alternatives
  3. Multi-objective design has a Pareto frontier, not a single best

GFlowNets (Bengio et al., 2021, Jain et al., 2022) solve this by sampling proportionally to reward — generating diverse high-quality candidates rather than collapsing to a single mode.

Application to molecules (Jain et al., 2023): GFlowNets discovered diverse small molecules for biological targets, outperforming RL and Bayesian optimization in both quality and diversity of candidates.

Ant Colony GFlowNets (Rektoris et al., 2024): Combined swarm intelligence with GFlowNets for combinatorial optimization — populations of agents explore solution space collaboratively, discovering diverse solutions.

AlphaFold: Structure Prediction Meets Generative Design

AlphaFold 2 (Jumper et al., 2021) solved protein structure prediction. AlphaFold 3 (Abramson et al., 2024) extended to protein-ligand complexes.

The design loop is now:

Design molecule → Predict binding pose (AlphaFold 3)
  ↑                        │
  │                        ▼
  │                  Evaluate binding affinity
  │                        │
  └── Generate new ────────┘
      candidates
      (GFlowNet/VAE/Diffusion)

Fpocket provides the binding site detection step — identifying where on a protein surface a drug could bind.

Large-scale DNA synthesis closes the loop from computational design to physical molecules — synthesized genes can express designed proteins for experimental validation.

Diffusion Models for Molecular Generation

Diffusion models' success in images has inspired molecular diffusion:

  • Structure generation: Diffusing atom coordinates and recovering 3D molecular structures
  • Sequence generation: Diffusing protein sequences to design new proteins
  • Guided generation (analogous to classifier-guided diffusion in images): Steering generation toward molecules with desired properties

AlphaFold 3's key innovation was switching from the coordinate regression of AlphaFold 2 to a diffusion-based structure generator — treating structure prediction as conditional generation.

Topological Data Analysis as a Bridge

Amézquita et al. (2020) applied persistent homology to biological structures, revealing that:

  • Protein shapes have topological features (tunnels, cavities, voids) that correlate with function
  • Molecular surfaces can be characterized by their Betti numbers — a more robust descriptor than geometric features

TDA provides a representation for molecular structure that is invariant to rotation, translation, and continuous deformation — exactly the properties needed for machine learning on molecules.

The Emerging Drug Discovery Pipeline

Target Identification
        │
        ▼
Binding Site Detection ← Fpocket
        │
        ▼
Candidate Generation ← VAE / GFlowNet / Diffusion
        │
        ▼
Binding Pose Prediction ← AlphaFold 3
        │
        ▼
Property Prediction ← GNNs on molecular graphs
        │
        ▼
Diversity Optimization ← GFlowNet
        │
        ▼
Synthesis Planning ← Large-scale DNA synthesis
        │
        ▼
Experimental Validation

The key insight across this collection: Every major generative architecture — VAEs, diffusion models, normalizing flows, GFlowNets — has found a natural application in molecular design. The mathematical structures that generate realistic images also generate realistic molecules.

Relevant Topic Pages