Traditional molecular machine learning models are typically bound to the specific chemical topologies, sequences, or domains present in their training sets. A model trained on small, synthetic drug-like molecules routinely fails when applied to large, cyclic peptides; similarly, a model trained on single-chain proteins often struggles to predict the spatial arrangements of multi-chain complexes.
Cross-domain molecular generalization overcomes these boundaries. Instead of forcing a neural network to memorize global molecular identities or specific sequence-structure mappings, this paradigm shifts the learning target toward local geometric invariants and universal physical principles. Because fundamental forces—like van der Waals interactions, hydrogen bonding, and electrostatic potentials—operate identically whether they occur in a synthetic drug, a biological protein, or an inorganic crystal, models that capture these local patterns can generalize to entirely unseen chemical domains in a zero-shot manner.
Grounding Models in Local Geometry and Physical Invariants
To achieve transferability across distinct molecular domains, a model must use representations that do not break when the size or composition of the system changes.
For example, in molecular dynamics (MD) simulations, traditional models often rely on internal coordinates (such as specific bond angles and torsions) that are highly sensitive to a molecule's exact topology. To bypass this limitation, Klein et al. developed Timewarp: Transferable Acceleration of Molecular Dynamics by Learning Time-Coarsened Dynamics. By modeling molecular transitions directly in Cartesian coordinates, the authors allowed the network to learn from small molecular building blocks and successfully simulate the dynamics of larger, unseen peptides without system-specific retraining.
[Internal Coordinates] ---> Tied to specific molecular topology (Fails on new domains)
[Cartesian Coordinates] ---> Learns local atomic environments (Generalizes across domains)
This geometric invariance can even bridge the gap between biological and synthetic nanotechnology. Cha et al. demonstrated in Unifying structural descriptors for biological and bioinspired nanoscale complexes that by utilizing composition-agnostic spatial descriptors, a model trained exclusively on biological protein-protein interactions can accurately predict binding sites on inorganic nanoparticles.
When a model views the chemical world through a purely geometric and physical lens rather than a taxonomic one, boundaries between domains dissolve. As shown by Kong et al. in Generalist Equivariant Transformer Towards 3D Molecular Interaction Learning, representing diverse entities (proteins, small molecules, and nucleic acids) as unified geometric graphs allows a model to learn universal atom-level interaction rules. Consequently, the model can perform zero-shot transfer to entirely new domains, such as predicting binding affinities for RNA and DNA using weights trained solely on protein-ligand data.
Zero-Shot Repurposing of Generative Foundation Models
A key benefit of capturing universal physical principles is the ability to perform zero-shot inference on complex tasks that traditionally required expensive, system-specific sampling or optimization.
In the realm of kinetics, Raja et al. demonstrated in Action-Minimization Meets Generative Modeling: Efficient Transition Path Sampling with the Onsager-Machlup Functional that pre-trained generative models (such as diffusion and flow matching) can be repurposed to find high-probability transition pathways between molecular states. By minimizing the Onsager-Machlup action functional using the model's learned score function, they eliminate the need for costly, system-specific retraining when predicting the dynamics of novel chemical sequences.
Similarly, for thermodynamic sampling, Lu et al. introduced STR2STR in Score-based Enhanced Sampling for Protein Molecular Dynamics. By training a diffusion model on the $SE(3)$ manifold of static crystal structures, the authors enabled zero-shot conformational sampling for novel proteins using only a single starting structure, bypassing the need for protein-specific simulation trajectories.
This zero-shot transferability extends to structural biology and materials science foundation models:
- Multi-Chain Docking: Bryant et al. showed in Improved prediction of protein-protein interactions using AlphaFold2 that models trained exclusively on single-chain protein structures can be adapted for multi-chain docking. The model's internal understanding of protein geometry and physical constraints is robust enough to generalize to the spatial arrangement of separate interacting subunits without explicit training.
- Crystal Structure Generation: In Large Language Models Are Innate Crystal Structure Generators, Gan et al. demonstrated that general-purpose large language models possess sufficient latent knowledge of crystallography and chemical bonding from their broad pre-training corpus to propose physically plausible crystal structures without any domain-specific fine-tuning.
Extracting Foundational Chemical Logic from Sequences
Cross-domain generalization does not always require explicit 3D coordinates; it can also be achieved by training language and sequence models on massive datasets to capture the latent "logic" of chemistry.
When faced with multi-property molecule optimization, models often struggle when asked to optimize combinations of traits they have never seen together. To address this, Dey et al. developed \mathttGeLLM³O: Generalizing Large Language Models for Multi-property Molecule Optimization. Their instruction-tuned chemical models exhibit significant zero-shot capabilities when faced with entirely new property combinations, proving that the underlying representation captures a foundational chemical logic rather than rigid task boundaries.
[1D Sequence View] <====== Multi-View Pre-training ======> [3D Geometric View]
|
v
[Robust Out-of-Sample Generalization]
To make these sequence representations even more resilient, researchers combine them with geometric views. Shi et al. demonstrated in Prediction of chemical reaction yields with large-scale multi-view pre-training that aligning 1D SMILES sequences with 3D spatial geometries during pre-training significantly enhances a model's ability to generalize to out-of-sample chemical spaces. This multi-view alignment allows the model to accurately predict reaction yields when encountering novel catalysts or reactants that were completely absent from the training set.
Ultimately, this latent understanding allows models to map directly between entirely different modalities, such as sequences and structures. As shown by Chen et al. in Deep generative model for drug design from protein target sequence, a deep generative model can design potential drug molecules solely from a target protein's amino acid sequence. Because the model learns the underlying rules of target-ligand compatibility, it achieves zero-shot generation for protein targets not encountered during training, bypassing the need for explicit structural data or prior ligand knowledge.