🔬 materials science

MolMiner: Toward Controllable, 3D-Aware, Fragment-Based Molecular Design

MolMiner is a novel fragment-based, 3D-aware autoregressive model that enables controllable molecular design by generating structures conditioned on twelve physicochemical properties and force-field-relaxed geometries, significantly improving hit rates for targeted property windows without requiring auxiliary loss functions.

Original authors: Raul Ortega-Ochoa, Tejs Vegge, Jes Frellsen

Published 2026-07-15

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Raul Ortega-Ochoa, Tejs Vegge, Jes Frellsen

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you're trying to build a custom LEGO castle, but instead of snapping together individual tiny bricks one by one, you have a magical box of pre-made, chemically perfect "super-bricks" (like whole towers, arches, or windows). This is exactly how MolMiner works. It's a new computer program designed to invent new molecules for medicine and materials, but it does so by snapping together these larger, meaningful chunks rather than atom-by-atom.

Here's the cool part: most computer programs that build molecules are like a rigid assembly line. They start at one specific corner and march forward in a straight line, never looking back or changing their path. If they make a mistake early on, the whole thing falls apart. MolMiner, however, is order-agnostic. Think of it like a chaotic but brilliant artist who can start building a molecule from any valid piece and grow it in any direction they want, as long as the pieces fit together. The paper shows that letting the model practice building the same molecule in many different ways (like a musician jamming on a song in different keys) makes it much better at generalizing and creating valid structures.

But here's the real magic trick: MolMiner doesn't just build; it sees in 3D. While building, it constantly relaxes the shape of the molecule using a "force field" (a physics simulation that makes atoms settle into comfortable positions). It's like if you were building a sandcastle and had to constantly check if the wet sand was holding its shape before adding the next bucket. This ensures the molecule isn't just a flat drawing but a physically plausible 3D object.

The paper's main finding is that MolMiner can be controlled with incredible precision. You can tell it, "I need a molecule that is this heavy, this soluble, and has this many rings," and it will listen. In fact, when the researchers asked it to build molecules fitting specific "drug-like" windows, it succeeded 5.25 times more often than if it had just guessed randomly, and 3.5 times more often than the molecules it was originally trained on. This is huge because it means the model can override its own habits to hit a specific target.

However, there are some rules and limits the paper is very clear about:

It's not a magic wand for everything: The model is great at 12 specific properties (like weight, solubility, and flexibility), but it struggles a bit with QED (a complex score for "drug-likeness"). The authors suggest this is because QED is just a mix of the other 11 properties, so the model handles it indirectly rather than directly.
It's not perfect yet: The model tends to build slightly smaller molecules than the training data because of how it decides when to stop adding pieces. The authors measured this and found a small bias toward smaller sizes, which they can tweak but haven't completely erased.
No "cheating" allowed: The paper explicitly rules out using "oracle" evaluations (where a super-computer checks the molecule's properties after it's built to decide if it's good). MolMiner learns to control properties during the building process, without needing a second opinion from a separate calculator. This makes it much faster and more practical for real-world use.

The authors are measured and confident in their results. They didn't just simulate this on a small scale; they trained the model on about 200,000 real drug-like molecules and tested it rigorously. They proved that by using these "super-bricks" and 3D awareness, they can generate molecules that are 100% chemically valid (no broken bonds or impossible shapes) and incredibly diverse.

In short, MolMiner is like a master LEGO architect who can build a 3D structure from any starting point, constantly checking the physics of the build, and following a detailed checklist of 12 different requirements. It doesn't just guess; it constructs. And while it's not a solved problem for every single chemical challenge, the paper shows it's a massive step forward in making molecular design controllable, efficient, and ready for the next generation of drug discovery.

Technical Summary: MolMiner

Problem Statement

Deep generative models are increasingly utilized in high-throughput screening (HTS) pipelines to generate candidate molecules tailored to specific properties. However, existing unconstrained models often explore chemically implausible regions of molecular space, particularly when training signals are limited or representational ambiguity is high. Current approaches face several specific limitations:

Lack of Geometric Awareness: Few autoregressive frameworks integrate dynamic three-dimensional (3D) geometry during generation, often relying on frozen or static structures.
Fixed Construction Order: Many models rely on a fixed rollout order (e.g., growing from a predefined atom), which does not reflect the intrinsic symmetry of molecular graphs and limits generalization.
Limited Conditional Control: Most models support only single-target conditioning, whereas practical drug discovery requires simultaneous control over multiple physicochemical and structural properties.
Validity and Bias: Unconstrained generation can lead to chemically invalid structures or biased exploration of chemical space.

Methodology

The authors introduce MolMiner, a fragment-based, geometry-aware, and order-agnostic autoregressive model designed to impose strong inductive biases on molecular generation. The methodology comprises four core components:

1. Fragment-Based Representation

Molecules are decomposed into chemically meaningful fragments corresponding to rings (identified via RDKit's Smallest Set of Smallest Rings) and isolated bonds. These fragments are treated as discrete tokens. To handle the loss of connection information during fragmentation, the model tracks the mapping between original atom indices and local fragment indices.

2. Symmetry-Aware Attachment

To ensure consistent representation of fragment attachments despite chemical symmetry (e.g., benzene rings where all carbons are equivalent), MolMiner employs a symmetry-aware standardization procedure. This involves enumerating symmetry-equivalent attachment configurations and constructing a canonical reference mapping. This ensures that attachment decisions are invariant to atom indexing and fragment automorphisms.

3. Order-Agnostic Rollout

Unlike models with fixed traversal orders (breadth-first or depth-first), MolMiner adopts an order-agnostic strategy. At each step, a focal attachment site is sampled uniformly from the set of currently open sites. The model predicts either a fragment to attach or a termination decision for that specific site. This approach:

Allows multiple valid construction trajectories for the same molecule.
Acts as a form of data augmentation and regularization during training.
Enables decentralized termination, where attachment points are closed independently until the molecule is fully assembled.

4. Geometry-Aware Architecture

The model utilizes a decoder-only transformer architecture. To incorporate 3D structural information without explicitly generating coordinates at every step, the architecture integrates a global attention bias based on the Euclidean distance between fragment centroids. These centroids are derived from force-field-relaxed (e.g., UFF) intermediate structures. The attention mechanism is modulated by a distance-based kernel, allowing the model to condition predictions on physically plausible intermediate geometries.

5. Training and Conditioning

Training Objective: The model maximizes the log-likelihood of a molecule under the order-agnostic rollout factorization, conditioned on target properties. An auxiliary feed-forward network is trained to predict a suitable starting fragment based on target properties.
Implicit Conditioning: Target properties are provided as inputs during training, but no auxiliary property losses are applied. The model learns property alignment organically from the data distribution.
Multi-Property Control: The system supports conditioning on a subset of 12 physicochemical and structural properties (e.g., logP, QED, SAS, molecular weight). Missing properties are completed using a Gaussian Mixture Model (GMM) prior trained on the empirical joint distribution of the training data.

Key Results

Unconditional Generation

Evaluated on a subset of the ZINC dataset (~200,000 molecules), MolMiner demonstrates:

Validity: 100% chemical validity by construction, with 0% fragmentation (unlike diffusion-based baselines like GDSS which produced fragmented outputs in ~39% of cases).
Diversity and Novelty: High uniqueness (98.5%) and novelty (99.9%), comparable to state-of-the-art fragment-based (HierVAE) and diffusion-based (GDSS) models.
Distributional Fidelity: The model closely matches reference distributions for 12 properties. Minor deviations in molecular weight, molar refractivity, and TPSA were observed, attributed to a bias toward early termination in the rollout process, which was partially corrected via loss reweighting.

Conditional Generation

Calibration: MolMiner achieves low Mean Absolute Error (MAE) and high $R^2$ across 12 properties without explicit property losses. The model successfully learns structure-property relationships.
Multi-Property Control: The model supports simultaneous conditioning on all 12 properties. While a "specialist" model (trained on a single property) performs slightly better on that specific metric, the "generalist" 12D model maintains high controllability across all axes, accepting a small per-property cost for the ability to satisfy complex, multi-constraint specifications.

Targeted Multi-Property Generation

In experiments targeting specific "property windows" (e.g., Lipinski's Rule of 5, CNS drug criteria, 3D-rich-rigid scaffolds):

MolMiner significantly outperforms unconditional baselines (HierVAE, GDSS, and unconditional MolMiner).
Hit Rate Lift: Conditioning lifted hit rates by up to 5.25× over the unconditional baseline and 3.5× over the training distribution itself.
Bias Override: The model successfully steered generation into regions of chemical space it naturally under-represented (e.g., 3D-rich-rigid molecules), demonstrating that conditioning can override intrinsic model biases.

Significance and Claims

The paper claims that MolMiner unifies dynamic geometry, symmetry handling, order-agnostic generation, and scalable multi-property conditioning within a single framework. Its significance lies in:

Controllability without Auxiliary Losses: Demonstrating that high-dimensional conditional control can emerge organically from the data distribution without requiring explicit property-based loss functions or oracle evaluations during generation.
Operational Relevance: Providing a practical tool for HTS pipelines that can generate molecules satisfying simultaneous, multi-property constraints, effectively overriding the model's intrinsic biases to explore specific chemical spaces.
Interpretability and Co-Design: The stepwise, fragment-based assembly process supports interpretability and opens pathways for human-in-the-loop design, where users can intervene during the rollout to constrain fragments or attachment sites.

The authors position MolMiner not as a replacement for downstream optimization, but as a robust generative foundation that ensures chemical validity, structural diversity, and reliable property control for initial candidate generation.