From Binary Interaction Labels to Continuous Property Regression

This direction is split into other directions: binding affinity regression. Its last version stays readable and citable here.

Version 1 of 191 papers109 conceptsfrozen 2026-10-04

In scientific machine learning, researchers often begin by framing complex physical phenomena as binary classification tasks: Does this drug bind to this protein? Is this material a superconductor? Is this printed circuit functional? While these "yes/no" labels are easier to collect and model, they represent a lossy compression of reality.

A growing body of research across drug discovery, materials science, and genomics argues that binary labels create a "resolution gap" that limits clinical and engineering utility. The emerging direction is a shift toward continuous property regression, where models predict the exact magnitude of an interaction—such as binding affinity, dissociation constants, or physical constants—rather than just the existence of a state.

The Binary Trap and the Resolution Gap

Binary classification is often a pragmatic choice driven by the nature of high-throughput experimental data. For example, Emily K. Makowski and colleagues note that deep-sequencing methods for antibody screening typically produce categorical data. However, this abstraction obscures the continuous spectrum of binding strength. As Jo‐Lan Chung and Philip E. Bourne pointed out early on, binary labels cannot provide the thermodynamic information necessary for molecular design.

This "resolution gap" is particularly visible in structural modeling. Martina Milighetti and Benny Chain found that while coarse structural features can identify if a T cell receptor binds to an antigen, they often lack the precision to quantify how strongly it binds. Without this quantitative nuance, models cannot distinguish between a weak, clinically irrelevant interaction and a potent therapeutic lead—a limitation echoed by Hansaim Lim in the context of drug off-target identification.

Why Regression Improves Learning

The move toward regression isn't just about getting more precise numbers; it actually changes how models learn. Fabian H. Reith and Guido Davidzon demonstrated that training a neural network to predict continuous quantitative values (like biomarker intensity) before applying a threshold for classification provides a "richer gradient" for the model.

In a binary setup, the model only knows if it was right or wrong. In a regression setup, the model receives feedback on how far it was from the true value. This granular feedback allows the model to capture subtle variations that are lost in discrete labeling. This is why Tong He and Martin Ester developed SimBoost, which prioritizes continuous binding affinities to provide more reliable drug-target insights than traditional "yes/no" models.

Overcoming the Data Scarcity Barrier

The primary hurdle for this transition is that high-quality regression data is significantly harder to obtain than binary data. To bridge this gap, researchers are employing several key strategies:

  1. Transfer Learning and Pre-training: Joon-Hyuk Ko and Sunghoon Kim use massive protein sequence databases to pre-train models, giving them a "chemical intuition" that helps them perform regression even on small labeled datasets. Similarly, Ke Liu uses masked language modeling on crystal structures to move toward predicting continuous physical variables like critical temperatures.
  2. Mechanistic and Physical Priors: Rather than relying on "black-box" deep learning, Aviezer Lifshitz and M. Eldar use physically interpretable models to calculate binding energies across a continuous spectrum. This allows them to capture the influence of low-affinity sites that threshold-based binary methods typically ignore.
  3. Active Learning: Daria Balashova and Philippe A. Robert are using active learning to strategically pick which experimental data points to collect, specifically targeting the most informative continuous values to minimize the cost of moving beyond binary labels.

The Engineering Payoff

The shift from classification to regression has immediate practical consequences. In materials science, Xiaojie Zhang and James M. Rondinelli argue that while classifying materials is a start, high-precision regression for constants like band gaps is required for actual device engineering. In manufacturing, Mehran Abbasi Shirsavar notes that predicting whether a printed sensor is "functional" is less useful than predicting its exact electrical resistance for fine-tuned process control.

By treating scientific properties as the gradients they are, researchers like Hexin Bai and Xiaofeng Qian are moving machine learning from a tool for simple screening to a tool for precise, quantitative engineering. As Fuli Feng concludes, capturing the continuous spectrum of nature is the key to unlocking the next generation of biological and physical discovery.

Papers in this version

Showing 60 of 91 papers.

History & citationVersion 1 · 2026-10-04

Split into other directions · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_from_binary_interaction_labels_to_contin_6e8bd201,
  title = {From Binary Interaction Labels to Continuous Property Regression},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/from-binary-interaction-labels-to-continuous-property-regression/rdr_6e8bd201253d429a9261306748d4e87f},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-19},
}

This exact version (version 1, frozen text):

@misc{lacuna_from_binary_interaction_labels_to_contin_6e8bd201_v1,
  title = {From Binary Interaction Labels to Continuous Property Regression},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/from-binary-interaction-labels-to-continuous-property-regression/rdr_6e8bd201253d429a9261306748d4e87f/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-19},
}