TLDR: TESSERA is a new method that uses a Mixture-of-Experts model combined with conformal prediction to provide reliable, adaptive, and informative uncertainty estimates for AI predictions. It’s particularly effective in risk-sensitive areas like drug discovery, ensuring accurate uncertainty even when data distributions shift, which is a common challenge for existing methods. TESSERA achieves near-nominal coverage, efficient interval widths, and strong adaptivity, outperforming many baselines.
In the rapidly evolving landscape of Artificial Intelligence (AI) and Machine Learning (ML), models are increasingly deployed in critical, risk-sensitive fields such as autonomous driving, medical diagnostics, and drug discovery. In these domains, mistakes can carry significant costs, making reliable and informative uncertainty quantification (UQ) absolutely essential. However, current ML methods often struggle to provide consistent coverage on new data, produce overly broad and unhelpful prediction intervals, or assign uncertainties that don’t accurately reflect actual errors, especially when the data distribution shifts.
A particularly challenging area is protein-ligand affinity (PLI) prediction in drug discovery. This field is characterized by heterogeneous assay noise, vast and imbalanced chemical spaces, and frequent encounters with out-of-distribution (OOD) data, where models must generalize beyond their training experience. Traditional UQ methods often fall short in these complex scenarios, leading to unreliable predictions.
Introducing TESSERA: A Novel Approach to Trustworthy Uncertainty
To address these critical limitations, researchers have introduced a novel uncertainty quantification method called TESSERA, which stands for Trustworthy ExpertSplit-conformal with Scaled Estimation for Efficient Reliable Adaptive intervals. TESSERA aims to provide per-sample uncertainty with a reliable coverage guarantee, offering prediction intervals that are both informative and adaptive, meaning their widths accurately track the absolute error of the prediction.
The core innovation of TESSERA lies in its ability to unify the diversity of a Mixture of Experts (MoE) model with the rigorous calibration of conformal prediction. This combination allows TESSERA to deliver uncertainties that are trustworthy, tight, and adaptive, making it highly suitable for selective prediction and downstream decision-making in drug discovery and other high-stakes applications.
How TESSERA Works
TESSERA operates on two main principles. First, it uses a Mixture-of-Experts (MoE) backbone to learn two distinct signals of difficulty: expert disagreement and per-expert variance. Expert disagreement quantifies the lack of consensus among the routed experts, serving as a proxy for epistemic uncertainty (uncertainty due to the model’s own limitations or lack of knowledge). Per-expert variance heads capture the spread of predictions at the expert level, representing aleatoric uncertainty (uncertainty inherent in the data itself, like noise).
Second, TESSERA applies a single split-conformal calibration to these model-aware difficulty scales. Conformal prediction is a powerful, distribution-free framework that, when applied, ensures finite-sample coverage guarantees for the prediction intervals. This means that the intervals are statistically guaranteed to contain the true value a certain percentage of the time. Crucially, TESSERA’s intervals widen where the model is uncertain (high disagreement or variance) and remain tight where experts agree, making them highly adaptive and informative.
The method was rigorously evaluated on protein-ligand binding affinity prediction under both independent and identically distributed (i.i.d.) and scaffold-based out-of-distribution (OOD) splits, comparing it against strong UQ baselines.
Key Advantages and Findings
The research highlights several significant contributions of TESSERA:
- Reliable, Informative, and Adaptive Uncertainty: TESSERA consistently achieves near-nominal coverage, meaning its prediction intervals reliably contain the true values. It also offers competitive widths, ensuring the intervals are as tight as possible while remaining valid. Furthermore, its strong adaptivity means the intervals accurately reflect the local difficulty of each prediction, widening for harder cases and tightening for easier ones.
- Robustness Under Distribution Shift: A major challenge for many UQ methods is maintaining performance when the test data differs significantly from the training data. TESSERA’s MoE-decomposed uncertainty heuristics effectively concentrate disagreement in OOD regions, and the conformal calibration translates this into coverage-guaranteed intervals, making it robust to distribution shifts.
In comparative tests, TESSERA attained near-nominal coverage and demonstrated the best coverage–width trade-off, as measured by the Coverage–Width Criterion (CWC). It also maintained competitive adaptivity, indicated by the lowest Area Under the Sparsification Error (AUSE). This means TESSERA’s uncertainty ranking is very close to an ideal oracle, effectively identifying points where the model is likely to be wrong. Size-Stratified Coverage (SSC) further confirmed that TESSERA’s intervals are ‘right-sized,’ expanding when data is scarce or noisy and remaining tight when predictions are reliable.
While other methods like Monte Carlo Dropout and RIO-GP might produce narrower intervals, they often suffer from systematic under-coverage, rendering their uncertainties uninformative. Classical Conformal Prediction, while providing marginal coverage, often yields constant-width intervals, lacking the adaptivity needed for heteroscedastic real-world data.
Also Read:
- Evaluating Language Models on Real-World Uncertainty with OPENESTIMATE
- Foundation Models: Charting a New Course for Scientific Exploration
Future Directions
The researchers acknowledge that while TESSERA offers significant advancements, there are areas for future exploration. The precise disentanglement of aleatoric and epistemic uncertainty remains a complex challenge, requiring further validation with synthetic controls. Future work also includes testing TESSERA under different types of distribution shifts, such as protein-conditioned splits, and exploring shift-aware conformal variants to further enhance its guarantees under covariate shift.
In conclusion, TESSERA represents a significant step forward in uncertainty quantification, providing a robust and adaptive framework for AI systems in fields where reliable predictions are paramount. For more in-depth technical details, you can refer to the full research paper here.
