Small-Data AI Optimization of Bicyclic Peptide Inhibitors

CD ComputaBio provides software-based computational services to support research and development. We do not offer free software packages.

Overview

The Starting Point: A Fixed Bicyclic Scaffold and Measured Activity

The initial dataset contained 37 phage-selected bicyclic peptides with experimentally determined Ki values ranging from 53 to 7,670 nM. These inhibitors shared three cysteines and two loops, each containing six intervening amino acids. Reaction with 1,3,5-tris(bromomethyl)benzene (TBMB) constrains the peptides through a common chemical scaffold. The design problem therefore varies the loop sequences while keeping the cyclization chemistry and cysteine arrangement fixed.

This controlled chemical context is important. The model does not have to learn how unrelated linkers or ring sizes affect activity. Conversely, it cannot establish those effects from this dataset. A favorable sequence pattern in this family should not automatically be transferred to a different cyclization strategy.

UK18 had already undergone attempted experimental affinity maturation without a clear improvement in potency. The paper asks whether computational analysis can recover useful combinations that are not obvious from the strongest sequence consensus. Its starting information is richer than an arbitrary list of peptides: the molecules are related, selected against the same target, and accompanied by quantitative enzyme-inhibition measurements.

Why Neither Model Was Sufficient on Its Own

The authors initially used random forest regression (RFR) to predict Ki from amino-acid sequences and, in additional modeling, sequence-derived properties. A random forest combines predictions from multiple decision trees. It can accommodate small datasets more readily than training a large neural network from scratch, but that does not eliminate uncertainty. Here, the authors observed high prediction error and overfitting.

They also explored pseudolikelihood maximization direct coupling analysis (plmDCA). This statistical model describes position-specific residue preferences and pairwise sequence dependencies from a multiple sequence alignment. Although the terminology refers to coevolutionary information, the input here is a small experimentally selected peptide family, rather than a large natural evolutionary alignment.

The resulting score is not a measured binding energy or a calibrated Ki. With only a few dozen sequences, not every inferred residue relationship is reliable. The authors therefore use plmDCA to guide exploration of plausible sequences and RFR to prioritize the generated collection. Both models are trained on the same limited experimental evidence; combining them does not create an independent validation dataset.

A Hybrid Workflow with an Explicit Role for Expert Judgment

Monte Carlo sampling uses the plmDCA score as an effective energy to explore candidate sequences. For the initial family, approximately 23,600 sequences were generated. RFR-based filtering produced a subset of approximately 3,000 sequences, from which the authors examined positional amino-acid frequencies. The reported Ki cutoff was 0.92 μM, described in the paper as a 50th-percentile threshold.

These counts describe computational exploration, not thousands of experimentally validated improved inhibitors. More importantly, the synthesized candidates were not simply the highest-ranked raw model outputs. Because the regression model was uncertain, the researchers interpreted the sequence logo, chose representative residue combinations, and used the known UK18–huPA structure where necessary to guide decisions.

Eight initial candidates were synthesized, cyclized with TBMB, purified, and characterized by mass spectrometry before enzyme assays. Seven additional variants then tested the contribution of selected substitutions. This distinction matters when describing the workflow: the model supplies hypotheses, while experimental data and structural reasoning determine which hypotheses become molecules.

Potency Improves Through Combinations and Follow-Up Variants

The first eight designed peptides had Ki values from 7.4 to 154.9 nM. Some improved on UK18, while others were weaker. UK957 reached 7.4 nM and UK961 reached 16.2 nM. Their results supported the value of Gly14 and Pro15 in the second loop, but did not establish that every residue enriched by the computational process was beneficial.

The seven follow-up variants retained the Gly14/Pro15 combination while testing first-loop substitutions. This experimental dissection produced inhibitors with Ki values between approximately 4.27 and 31.4 nM. UK970 was the strongest, reported at about 4.3 nM. Relative to the stated 53 nM starting value, that is approximately a twelvefold reduction in Ki; the paper also describes several variants as at least tenfold more potent than the parental inhibitor.

Ki is an enzyme-inhibition parameter, not a directly measured dissociation constant. These biochemical measurements should not be relabeled as cellular potency, tumor-growth inhibition, or therapeutic efficacy. Figure 2 reports a remeasured UK18 value near 54.7 nM; that small difference from the historical 53 nM benchmark does not alter the principal comparison.

Structures Explain More Than a Better Sequence Score

The authors determined huPA complexes with UK965 and UK970 at 1.64 Šand 1.8 Šresolution, respectively, and compared them with the parental UK18 complex. The optimized peptides occupy the substrate-binding region and engage the enzyme through both loops. Their buried surface areas were 749 Ų and 746 Ų, compared with 730 Ų for UK18.

The interpretation is not simply that a larger interface must bind better. Residue changes alter the conformations of the loops and the network of intermolecular and intramolecular contacts. Pro15 is associated with a turn and rearrangement of the first loop, and the optimized molecules show more compact structures in the complexes. These observations help explain why substitutions must be evaluated in sequence context.

The TBMB scaffold also deserves a nuanced reading. It was held constant and omitted as a varying model feature, but structural analysis suggests hydrophobic packing around the small organic core may contribute to the optimized molecules' organization. The study therefore does not establish that the linker is chemically irrelevant. Nor do crystal conformations alone quantify the full unbound conformational ensemble or the entropy of binding.

Iteration and a Second Family Test the Strategy's Reach

Adding the fifteen newly measured sequences to the original thirty-seven expanded the first-family dataset to fifty-two. Another round of modeling and sampling guided three further candidates. UK971, incorporating Trp at a selected position, reached 6.4 nM, while the other two variants had Ki values of 14.8 and 15.3 nM. The iterative process recovered active alternatives, but did not surpass UK970.

A separate huPA family contained thirty-one sequences, with a different sequence pattern and activities spanning approximately 0.20–51.4 μM. Here, the researchers lacked an informative parental complex structure for design. Five candidates were synthesized using preferences inferred from the combined workflow. UK978 improved potency approximately 2.8-fold relative to UK140, whose Ki was 202 nM. Other candidates were worse, reinforcing the importance of reporting the whole tested set.

The paper additionally assessed sequence recovery in a factor XIIa inhibitor family with a different linker, TATA. After removing two sequences from that dataset, the computational process regenerated one excluded sequence. A similar exercise recovered one excluded huPA-family sequence. These are useful computational checks, but they are not prospective demonstrations of newly improved factor XIIa inhibitors. Optimization across two huPA families and sequence recovery on another target are different levels of evidence.

Selectivity Can Help the Lead and Complicate the Animal Model

UK970 showed very weak inhibition of the other tested trypsin-like proteases, with reported Ki values above 1 mM and more than 250,000-fold selectivity for huPA. The panel included murine uPA as well as related human proteases. Such selectivity is attractive when neighboring enzymes have essential physiological functions, but it creates a translational constraint: a molecule that spares mouse uPA is harder to evaluate for on-target effects in ordinary mouse models.

The second-family lead offers a different tradeoff. UK978 retained murine uPA inhibition at approximately 1.7 μM, compared with 2.6 μM for UK140, while human trypsin inhibition remained near 10.4 μM. Its improved huPA activity therefore increased selectivity relative to trypsin without removing the cross-species activity. This supports planning for future animal studies; the paper does not itself demonstrate therapeutic efficacy in those models.

Practical Lessons for Small-Data Peptide Optimization

The strongest starting point for this approach is a coherent family with consistent quantitative assays, a fixed chemical scaffold, and activity differences large enough to inform modeling. Sequence provenance, units, assay conditions, and the meaning of the activity endpoint are part of the input specification. A collection assembled from incompatible assays would weaken the interpretation of predicted rankings.

For a new project, the method should be treated as an iterative hypothesis-generation workflow. Maintain comparison compounds, test representative combinations rather than only a single top prediction, and record weaker candidates alongside improved ones. A prospective holdout panel and uncertainty assessment would help establish whether prioritization transfers beyond closely related sequences. Those are validation priorities, not performance already demonstrated for arbitrary targets.

The study also sets a clear boundary on scope: it optimizes sequences within established bicyclic scaffolds. Changing linker geometry, cysteine positions, ring lengths, or target class requires new evidence. It does not show that thirty-seven measurements are universally sufficient, that expert interpretation can be removed, or that improved enzyme inhibition automatically yields a developable drug.

Overview of What CD ComputaBio Can Provide

The paper connects to computational support for existing peptide-hit optimization and experimental planning. These service areas address related research tasks; they do not imply replication of the study's potency improvements or validation of this exact model on a client dataset.

References

1. Mazzocato Y, Frasson N, Sample M, et al. Combination of Coevolutionary Information and Supervised Learning Enables Generation of Cyclic Peptide Inhibitors with Enhanced Potency from a Small Data Set. ACS Central Science. 2024;10(12):2242–2252. Published November 20, 2024. https://doi.org/10.1021/acscentsci.4c01428.

* For Research Use Only.

Related Services