🤖 machine learning

Trustworthy Protein-Ligand Binding Affinity Prediction via Reliability-Aware Multi-Engine Fusion

The paper introduces RELIABLE-BA, a novel evidential framework that fuses multiple docking engines with context-dependent reliability estimates to provide trustworthy, well-calibrated uncertainty measures for protein-ligand binding affinity prediction, thereby enabling the filtering of low-confidence pairs to significantly improve prediction accuracy in drug discovery.

Original authors: Yongchan Hong, Defu Cao, Wenjin Liu, Thomas Ku, Jordy Homing Lam, Emily Nguyen, Willie Neiswanger, Vsevolod Katritch, Yan Liu

Published 2026-07-21

📖 3 min read☕ Coffee break read

CC BY 4.0

Original authors: Yongchan Hong, Defu Cao, Wenjin Liu, Thomas Ku, Jordy Homing Lam, Emily Nguyen, Willie Neiswanger, Vsevolod Katritch, Yan Liu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a detective trying to solve a mystery: how tightly does a tiny key (a drug molecule) fit into a specific lock (a protein in your body)? If the fit is perfect, the drug works; if it's loose, it fails. This is the heart of "computational drug discovery," a field where scientists use computers to predict these fits before ever mixing chemicals in a lab. It saves time and money because testing every possible drug in a real lab is slow and expensive.

However, computers aren't perfect. Scientists have built many different "prediction engines" (like different detective teams) to guess how well a drug fits. Sometimes, Team A says the fit is great, while Team B says it's terrible. The old way to handle this was to just take the average of all their guesses. But that's like asking five friends for directions and just walking in the middle of the road, ignoring that one friend might be looking at a map from 1990 while another is using a GPS. The problem is, we didn't have a good way to know which friend to trust for this specific lock. We needed a system that could not only give an answer but also say, "I'm 90% sure about this one, but I'm totally guessing on that next one."

This is where the new paper comes in. The researchers, led by Yongchan Hong and colleagues, created a smart system called RELIABLE-BA. Think of it as a super-smart "Chief Detective" who listens to all the different prediction engines (the "experts") but doesn't just average their answers. Instead, the Chief Detective looks at the specific lock and key in question and asks, "Hey, Expert A, you're usually great with this type of lock, but Expert B, you tend to get confused by this shape."

The system works in three clever steps. First, it treats each prediction engine as an "evidential expert." Instead of just giving a single number, each engine gives a range of possibilities, admitting how unsure it is. Second, the system learns a "reliability score" for each engine based on the specific chemistry of the drug and protein. If an engine is usually bad at a certain type of job, the system quietly lowers its voice in the final decision. Third, it combines all these adjusted opinions into one final, super-calibrated prediction.

The results are impressive. When they tested this new system on huge databases of known drug-protein pairs (called PDBbind and BDB2020+), it didn't just predict the answers well; it was much better at knowing when it was right and when it was wrong. In fact, when the system was allowed to only show its most confident predictions (ignoring the ones it was shaky on), it reduced the prediction error by up to 25.7%. This means scientists could throw away the risky guesses and focus only on the high-quality ones, saving a massive amount of time.

The team also tested this on real-world, tricky targets like the SARS-CoV-2 virus's main protease and the 5HT2A receptor (a target for mental health drugs). In these tests, RELIABLE-BA outperformed the individual engines and other methods, proving it can handle difficult, real-life scenarios. The paper suggests that by using this "reliability-aware" approach, we can make drug discovery safer and more efficient, ensuring that when a computer says a drug might work, we have a much better reason to trust it.

Technical Summary: Trustworthy Protein–Ligand Binding Affinity Prediction via Reliability-Aware Multi-Engine Fusion

Problem Statement
Accurate prediction of protein–ligand binding affinity is a cornerstone of computational drug discovery, yet modern docking engines frequently produce conflicting predictions without indicating which result to trust. While consensus scoring and ensemble methods improve mean accuracy, they typically treat all predictions identically, lacking interpretable confidence measures or uncertainty decomposition. Crucially, existing approaches ignore the chemical context of specific protein–ligand pairs, failing to distinguish between aleatoric uncertainty (inherent data noise) and epistemic uncertainty (model ignorance). Furthermore, single-model uncertainty quantification (UQ) methods cannot capture the additional epistemic uncertainty arising from disagreement among diverse engines, which rely on different assumptions, scoring functions, and training data.

Methodology: RELIABLE-BA
The authors propose RELIABLE-BA (RELIABiLity-aware Evidential fusion for Binding Affinity), an evidential framework designed to fuse multiple docking engines while accounting for context-dependent reliability. The framework operates in three primary steps:

Evidential Expert Modeling: Each docking engine is modeled as an evidential expert. Instead of outputting a single point estimate, each engine $i$ predicts the parameters of a Normal–Inverse-Gamma (NIG) distribution, $(\gamma_i, \nu_i, \alpha_i, \beta_i)$ , which captures both the predicted mean and the uncertainty. This allows for a closed-form decomposition of uncertainty into aleatoric (data noise) and epistemic (model uncertainty) components.
Reliability-Aware Scaling: A reliability network, conditioned on shared molecular embeddings (derived from ESM-2 for proteins and ChemBERTa for ligands), learns a trustworthiness score $r_i(x) \in (0, 1)$ for each engine based on the specific protein–ligand context. These scores modulate the epistemic uncertainty of each expert via a scaling transformation. Specifically, the reliability score scales the evidence parameters ( $\nu$ and $\alpha$ ) to increase epistemic uncertainty for less reliable experts in a given context, while preserving the expert's predictive mean ( $\gamma$ ).
MoNIG Aggregation: The reliability-scaled experts are fused using a Mixture of Normal–Inverse-Gamma (MoNIG) aggregation. This employs a closed-form summation operator that combines the NIG distributions, accumulating both within-expert uncertainty and inter-engine disagreement. The fusion rule ensures that conflicting predictions from different engines result in higher epistemic uncertainty in the final output.

The model is trained end-to-end using a joint objective that minimizes the negative log marginal likelihood (NLL) at both the individual expert level and the fused MoNIG level, with regularization to prevent excessive evidence accumulation during large prediction errors.

Key Contributions

First Multi-Engine Evidential Framework: RELIABLE-BA is presented as the first multi-engine binding affinity framework to combine evidential uncertainty quantification with context-dependent reliability modeling.
Reliability-Scaling Mechanism: The authors introduce a novel mechanism that scales epistemic uncertainty based on learned reliability scores without altering the predictive means of individual experts, enabling coherent fusion via MoNIG.
Selective Prediction Capability: The framework enables risk-aware decision-making by filtering predictions based on confidence, demonstrating the ability to reduce prediction error significantly when retaining only high-confidence pairs.

Experimental Results
The authors evaluated RELIABLE-BA on the PDBbind 2020 benchmark (using a time-split protocol) and the independent BDB2020+ test set. Additional validation was performed on the SARS-CoV-2 $M^{pro}$ dataset and the 5HT2A receptor (a GPCR).

Point Prediction: RELIABLE-BA achieved competitive point prediction accuracy, attaining the lowest Mean Absolute Error (MAE) of 0.8247 and highest $R^2$ of 0.7017 on the PDBbind test set, outperforming both individual engines and traditional aggregation baselines (e.g., consensus scoring, deep ensembles). On the independent BDB2020+ set, it achieved an MAE of 0.722.
Uncertainty Calibration: The model demonstrated substantially improved uncertainty calibration compared to baselines. On PDBbind, it reduced Expected Calibration Error (ECE) by 71.2% compared to standard evidential regression.
Selective Prediction: By discarding low-confidence predictions, RELIABLE-BA reduced the MAE by up to 25.7% when retaining only the top 10% most confident predictions.
Ablation Studies: Experiments confirmed that context-dependent reliability modeling is critical; removing it or using uniform weights degraded both accuracy and uncertainty quality. Furthermore, increasing the number of heterogeneous engines (from 2 to 4) consistently improved both accuracy and calibration.
Case Studies: The method maintained strong performance on the SARS-CoV-2 $M^{pro}$ dataset and the 5HT2A receptor, demonstrating applicability to clinically relevant targets and robustness under distribution shift.

Significance and Claims
The paper claims that RELIABLE-BA offers a principled path toward trustworthy AI-guided drug discovery. By providing calibrated uncertainty estimates that reflect both individual engine limitations and inter-engine disagreement, the framework allows practitioners to prioritize high-confidence predictions for experimental validation, thereby reducing costly false positives. The authors position RELIABLE-BA as a practical approach for uncertainty-aware virtual screening, particularly for challenging therapeutic targets like GPCRs where conformational flexibility and limited training data representation pose significant modeling difficulties. The work emphasizes that the framework is modular, allowing for the integration of diverse, non-redundant docking engines to enhance robustness in large-scale screening pipelines.

Technical Summary: Trustworthy Protein–Ligand Binding Affinity Prediction via Reliability-Aware Multi-Engine Fusion

More like this