🤖 AI

Confidence is the key: how conformal prediction enhances the generative design of permeable peptides

This paper introduces an uncertainty-aware generative framework that integrates conformal prediction with reinforcement learning to enhance the reliability and efficiency of designing permeable cyclic peptides by ensuring generated molecules remain within the predictive model's domain of applicability.

Original authors: Laura van Weesep, Sunay Chankeshwara, Leonardo De Maria, Florian David, Ola Engkvist, Gökçe Geylan

Published 2026-05-08

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Laura van Weesep, Sunay Chankeshwara, Leonardo De Maria, Florian David, Ola Engkvist, Gökçe Geylan

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 master chef trying to invent a new, magical soup that can pass through a locked door (a cell membrane) to cure a disease. You have a super-smart robot assistant (the Generative Model) that can mix and match thousands of ingredients (amino acids) to create new soup recipes.

However, there's a problem: you also have a Taste Tester (the Predictive Model) who tells you if a soup will pass through the door. The trouble is, this Taste Tester is only an expert on soups they've tasted before. If you ask them to judge a soup made of completely new, weird ingredients they've never seen, they might guess confidently, but they could be totally wrong. They might say, "This will definitely work!" when in reality, it's a disaster.

This paper is about teaching the robot chef how to work with a Wise Mentor (Conformal Prediction) who knows exactly when the Taste Tester is guessing and when they are sure.

The Problem: The Confident Guess

In the past, the robot chef just listened to the Taste Tester's confidence score. If the tester said, "I'm 90% sure this soup works," the robot would keep making that soup.

The Issue: Sometimes the tester is just guessing because the soup is too weird. The robot keeps making these "confident but wrong" soups, wasting time and ingredients. It's like the robot keeps trying to open a door with a key that looks like a key but is actually a spoon, because the lock-picker said, "It looks like a key!"

The Solution: The "Wise Mentor" (Conformal Prediction)

The authors introduced Conformal Prediction (CP). Think of CP as a Wise Mentor who sits next to the Taste Tester.

The Mentor doesn't just say "Yes" or "No."
The Mentor says: "I am 80% confident this soup will work," OR "I am not confident enough to say either way."
If the Mentor says, "I'm not sure," the robot chef knows to stop trying that specific recipe and try something else that looks more familiar and safe.

The Experiment: Three Ways to Listen

The researchers tested different ways for the robot chef to listen to this new Mentor system:

The "Raw" Approach (Old Way): The robot just listened to the Taste Tester's raw confidence.
- Result: The robot made lots of soups that looked good on paper, but many were actually "fake" (unreliable predictions). It was fast, but the quality was shaky.
The "Harsh" Approach: The robot only accepted a recipe if the Mentor was 100% sure it was a "Pass" and 100% sure it wasn't a "Fail."
- Result: This was too strict. The robot got stuck, couldn't learn, and stopped making soups because it was too afraid of making a mistake.
The "Soft" Approach (The Winner): This was the clever middle ground.
- If the Mentor was super sure, the robot got a big reward (1 point).
- If the Mentor was kind of sure (e.g., "It's probably a pass, but I'm not 100%"), the robot got a small reward (0.5 points).
- If the Mentor was clueless, the robot got zero points.
- Result: This worked best! The robot learned quickly. It found the "sweet spot" where it could invent new, permeable soups without wandering into the dangerous territory of "weird ingredients" where the Taste Tester is just guessing.

The Results: What Did They Find?

Reliability: By using the "Soft" approach, the robot generated fewer "fake" soups. It focused on recipes that the Mentor actually trusted.
Speed: The robot found good recipes faster than the old method because it didn't waste time chasing false leads.
The "Length" Rule: They noticed that the Mentor was very good at judging soups with 6, 7, or 10 ingredients (because the Mentor had tasted many of those before). But if the robot tried to make a soup with 12 ingredients, the Mentor got confused and couldn't give good advice. This taught the robot when to stop trying certain types of soups.

The Bottom Line

This paper shows that when you use AI to design new medicines (like peptide drugs), you shouldn't just trust the AI's "gut feeling." You need a system that tells you how sure the AI is.

By adding this "confidence check" (Conformal Prediction) into the learning loop, the AI becomes a smarter, more cautious explorer. It doesn't just run blindly into the unknown; it knows when to stick to the safe path and when it's safe to try something new. This leads to better drug designs that are actually likely to work, rather than just looking good on a computer screen.

Technical Summary: Confidence is the key: how conformal prediction enhances the generative design of permeable peptides

Problem Statement
Generative models coupled with reinforcement learning (RL), such as REINVENT and PepINVENT, are powerful tools for de novo molecular design. However, their reliance on predictive models as scoring functions introduces a critical vulnerability: the domain of applicability. When RL agents explore chemical space, they often propose molecules that lie outside the training distribution of the predictive model. In these regions, predictions become unreliable, potentially steering the optimization toward high-reward but high-uncertainty chemical spaces. This issue is particularly acute for cyclic peptides, which show therapeutic promise but are underrepresented in training data compared to small molecules. The challenge lies in distinguishing between genuinely permeable designs and artifacts of model extrapolation, where high predicted probabilities may lack calibration and reliability.

Methodology
The authors propose an uncertainty-aware generative framework that integrates Conformal Prediction (CP) directly into the RL loop for the design of cell-permeable cyclic peptides.

Generative Framework: The study utilizes PepINVENT, a transformer-based model trained on semi-synthetic data (natural and non-natural amino acids in CHUCKLES notation). The RL process involves masking amino acid positions in input cyclic peptides and generating replacements.
Predictive Model: An XGBoost classifier trained on the CycPeptMPDB database (6,876 cyclic peptide entries) predicts passive membrane permeability (PAMPA assay data). The model uses Extended Connectivity Fingerprints (ECFP) as descriptors.
Uncertainty Quantification: Instead of using raw model probabilities, the authors employ Mondrian Inductive Conformal Prediction (ICP). This framework calibrates predictions to provide valid confidence estimates for individual instances. It generates two p-values: $P_1$ $P_{1}$ (confidence the molecule is permeable) and $P_0$ $P_{0}$ (confidence the molecule is non-permeable).
- A prediction is considered "conformally efficient" (reliable) if $P_1 > 0.2$ and $P_0 < 0.2$ (corresponding to an 80% confidence level).
- An Aggregated Conformal Predictor (ACP) using 10 ICP models was used to improve efficiency.
Scoring Strategies: The study evaluated six scoring functions within the RL loop:
1. Raw Model (RM): Probability of the permeable class (baseline).
2. CP P1: Maximizing the p-value for the permeable class.
3. CP 1-P0: Maximizing the complement of the non-permeable p-value.
4. CP P1-P0: Maximizing the difference between the two p-values.
5. Harsh Function: A binary reward (1 or 0) only if both $P_1 > 0.2$ and $P_0 < 0.2$ .
6. Soft Function: A step function providing a partial reward (0.5) if either condition is met, and a full reward (1) if both are met.

The RL loop optimizes these scores over 350 epochs for 150 test queries (cyclic peptides with 6–12 residues), masking 1–4 positions per query.

Key Results

Limitations of Raw Probabilities: While the RL agent successfully increased raw model probabilities (from 0.51 to 0.87), the number of conformally efficient predictions remained low. High raw probabilities did not guarantee that the molecules fell within the predictor's applicability domain, as $P_0$ often remained above the significance threshold.
Superiority of CP-Based Scoring:
- Single P-value objectives (P1 or 1-P0) showed marginal improvements and failed to guarantee joint confidence.
- The "Soft" scoring function outperformed all other methods. It achieved the fastest convergence (reaching the target of 50% conformally efficient permeable predictions in ~150 epochs) and generated the highest fraction of molecules with confident permeable labels.
- The "Harsh" function was learnable but less efficient in early exploration due to sparse rewards.
Validity vs. Efficiency: The soft scoring function generated fewer total unique valid molecules compared to the raw model or single P-value methods, but a significantly higher ratio of those molecules were conformally efficient permeable predictions. This indicates that the raw model inflates hit rates with unreliable predictions, whereas the CP-soft approach targets reliable hits.
Length Dependence: Performance varied by peptide length. The soft scoring function effectively optimized peptides of lengths 6, 7, 8, and 10 (well-represented in training data). For lengths 9, 11, and 12 (less represented), the model struggled to reach high confidence levels within 350 epochs, demonstrating the framework's ability to signal when a design objective exceeds the predictor's applicability domain.

Significance and Claims
The paper claims to present the first framework that integrates conformal prediction directly into reinforcement learning-guided generative design. The authors argue that this approach:

Bridges the Gap: It connects the inherent uncertainty of predictive models with the strategic exploration of RL, preventing the agent from exploiting unreliable regions of chemical space.
Enhances Reliability: By rewarding designs based on calibrated confidence rather than raw probability, the framework ensures that generated peptides are not only high-scoring but also statistically valid within the model's domain.
Improves Efficiency: The "soft" scoring strategy accelerates convergence to high-reward, trustworthy chemical spaces compared to traditional uncertainty-unaware methods.
Practical Utility: The method provides a tunable mechanism (via significance levels) to control the trade-off between exploration and the strictness of the applicability domain, offering a robust strategy for molecular design tasks where predictive reliability is paramount.

The study concludes that while the specific optimal scoring function may vary by task, the synergistic effect of combining CP, RL, and generative models offers a generalizable path toward more reliable and efficient de novo molecular design.