Anchor-Based Conformal Prediction Under Noisy Annotations in Single-Cell Data

ICLR 2026 Conference Submission20743 Authors

Keywords: Conformal Prediction, Single-Cell Data, Noise Labels, Machine Learning

Abstract:

Learning predictive models from noisy annotations is a challenge in modern machine
learning, particularly in domains where labels are obtained from multiple
imperfect annotators. In this work, we introduce an anchor-based conformal prediction
framework that provides rigorous uncertainty guarantees even in the presence
of label noise. Our method identifies pseudo-anchors by selecting samples
with strong agreement across annotators, uses these anchors to train a base predictor,
and calibrates top-k conformal sets to ensure valid coverage. This construction
produces prediction sets that are both reliable and compact, while explicitly
accounting for annotation disagreement. Our results show that anchor-guided
conformal prediction attains coverage close to nominal targets while producing
smaller prediction sets and maintaining robustness in the presence of noisy labels.
Although evaluated on single-cell data, the framework more generally offers a
principled way to integrate multiple noisy annotator signals with conformal prediction,
enabling reliable uncertainty estimates under imperfect supervision.
This enables reliable uncertainty estimates in settings where ground-truth labels are scarce, expensive to obtain, or inherently ambiguous, and highlights how conformal methods can be applied to more realistic and noisy supervision scenarios.

Primary Area: applications to physical sciences (physics, chemistry, biology, etc.)

Submission Number: 20743