Few-Shot Molecular Property Prediction

This direction is merged into another direction: labeled data scarcity (data scarcity in drug discovery). Its last version stays readable and citable here.

Version 1 of 134 papers35 conceptsfrozen 2026-10-04

In early-stage drug discovery, researchers often face a "cold-start" problem. While deep learning has revolutionized many fields, it typically requires thousands of labeled examples to achieve high accuracy. In chemistry, obtaining those labels requires expensive, time-consuming "wet-lab" experiments. For a novel biological target or a rare disease, a scientist might only have a dozen known active compounds—a "few-shot" scenario where traditional models fail.

To bridge this gap, a new research direction has emerged: Few-Shot Molecular Property Prediction (FSMPP). As Zeyu Wang and colleagues note in their systematic survey, this paradigm shifts the focus from learning a specific property to "learning to learn" across diverse chemical tasks.

Learning to Learn via Meta-Learning

The most prominent strategy in this field is meta-learning. Instead of training a model to predict a single property (like toxicity), meta-learning models are trained on a distribution of many different chemical tasks. The goal is to learn a "prior" that can be rapidly adapted to a new, unseen task using only a handful of examples.

Chuxu Zhang and John E. Herr pioneered this with Meta-MGNN, which uses Graph Neural Networks (GNNs) combined with self-supervised auxiliary tasks. By learning the fundamental language of molecular graphs across many properties, the model can identify drug candidates with as little as one labeled example.

Similarly, Yaqing Wang introduced Property-Aware Relation networks (PAR). PAR recognizes that different properties depend on different structural features; it dynamically adapts the molecular representation and the relationship graph to the specific property being predicted, preventing the model from getting "stuck" in a static, one-size-fits-all representation.

Parameter-Efficient Adaptation

A major challenge in meta-learning is the "overfitting trap"—with so few examples, a complex model can easily memorize the noise rather than the signal. To solve this, Shiguang Wu and Quanming Yao developed PACIA.

Instead of updating the entire model, PACIA uses lightweight hypernetworks to generate a tiny set of adaptive parameters. These "adapters" modulate the molecular encoder at the task level and the predictor at the query level. This approach is not only more robust against overfitting but is also roughly five times faster than traditional gradient-based meta-learning methods.

In-Context Learning for Chemistry

Taking inspiration from Large Language Models (LLMs), researchers like Christopher Fifty and Jure Leskovec have introduced In-Context Learning (ICL) to molecular prediction. Their model, CAMP, treats property prediction as a sequence modeling task.

In CAMP, the model receives a "context" of a few labeled molecules and a "query" molecule in a single forward pass. It predicts the query's property by attending to the patterns in the context, much like an LLM answers a question based on a provided prompt. This eliminates the need for any gradient updates during inference, allowing for near-instantaneous screening of new biological targets. Gang Liu extended this concept to molecular design with DemoDiff, a graph diffusion transformer that can design high-performing molecules for entirely new tasks just by looking at a few examples.

Infusing Chemical Domain Knowledge

While general-purpose meta-learning is powerful, chemistry has unique rules that models can leverage. Yin Fang and Qiang Zhang developed Knowledge-enhanced Contrastive Learning (KCL), which injects the Periodic Table into the learning process via a Chemical Element Knowledge Graph. By grounding the model in fundamental chemical relationships, it becomes much more efficient at learning from scarce data.

Other researchers focus on hierarchical matching. Ruifeng Li and colleagues introduced UniMatch, which explicitly compares molecules at the atom, substructure, and whole-molecule levels. By matching a new molecule against a small "support set" of known actives across these multiple scales, the model can identify subtle structural similarities that a global representation might miss.

Why It Works: The Shared Chemical Space

The underlying intuition across all these papers is that biological properties are not independent. Toxicity, solubility, and binding affinity all emerge from the same underlying physics and chemistry. By training on thousands of diverse tasks, these models learn the "manifold" of drug-like molecules. When a researcher presents a new task with only five examples, the model doesn't have to learn chemistry from scratch; it only needs to figure out where this new task sits on the map it has already built.

Papers in this version

History & citationVersion 1 · 2026-10-04

Merged into another direction · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_few_shot_molecular_property_prediction_fdab9348,
  title = {Few-Shot Molecular Property Prediction},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/few-shot-molecular-property-prediction/rdr_fdab9348570f48b88f11096d1b5b0299},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_few_shot_molecular_property_prediction_fdab9348_v1,
  title = {Few-Shot Molecular Property Prediction},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/few-shot-molecular-property-prediction/rdr_fdab9348570f48b88f11096d1b5b0299/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}