Lead optimization in the discovery of novel incretin mimetics

A half-life extended GLP-1 mimetic was nominated on receptor binding scores and pharmacokinetic data, and it did not survive development. Pyxis™ analysis of cell extracts from the same screening experiment returns two dose-dependent signatures under the candidate that are absent under all three control peptides: depletion of nucleoside di- and triphosphates, and accumulation of arginine with its immediate precursor.

Lead optimization and attrition

Lead optimization ranks molecules within a series before a program commits to a candidate. Binding and potency panels are the common basis for that ranking because they are inexpensive and fast to run, and molecules that score well advance to candidate selection and to the functional and safety work that follows.

Attrition data indicate limits to this approach. Approximately half of phase II and phase III failures are attributed to lack of efficacy [1], so molecules clear optimization-stage assays and then fail on the property those assays are intended to predict. Among first-in-class small-molecule approvals between 1999 and 2008, phenotypic strategies accounted for 28 drugs and target-based strategies for 17 [2].

~50%

of phase II and phase III failures are attributed to lack of efficacy [1].

Selection criteria for this candidate

The candidate was a half-life extended GLP-1 mimetic delivered as an Fc fusion and engineered to engage more than one incretin receptor. Selection was based on receptor binding together with half-life data. Binding scored moderate to high against the GLP-1, GIP, and glucagon receptors, and the Fc format provided the intended circulation time. The molecule did not survive development.

Two properties of this class limit what binding data can establish. At the GLP-1 receptor the relationship between agonist binding and insulin release is non-monotonic, and agonists that retain the receptor at the plasma membrane produce greater long-term insulin release than standard cAMP potency assays predict [3]. Half-life extension introduces a second limitation, in that modifications extending circulation time can reduce functional potency while binding is retained. A glycosylated GLP-1 analogue fusion engineered for stability lost 350-fold in potency at the human GLP-1 receptor relative to its non-glycosylated parent [4], and reduced association rate from steric bulk is documented for the format.

The two selection criteria therefore interact. Half-life extension is the modification most likely to have reduced function in this molecule, and half-life was one of the two properties under optimization. Binding is retained when function is lost, so it does not report that reduction. A molecule can score higher on both criteria while its functional activity declines.

Biochemical readouts of cellular function

Biochemical measurements sit downstream of receptor engagement, signalling, and trafficking, and report the cellular consequence of exposure. Among 1,520 drugs profiled by non-targeted metabolomics, 26 percent inhibited cell growth and 86 percent produced measurable intracellular metabolic changes [5], which indicates the sensitivity difference between viability endpoints and metabolic ones.

Pyxis, the Matterworks co-scientist for omic data, reads raw files directly and draws on the Large Spectral Model, a foundation model trained on more than ten billion raw spectra. Each sample is interpreted into a de novo biochemome: an annotated inventory of the small molecules and lipids present, with concentrations attached, produced without per-dataset feature engineering and without a curated target list. For this study the analysis returned 1,314 analytes with untargeted biological concentrations in µM concentration units from an overnight experiment.

Nucleotide phosphate depletion

Cells were treated overnight with four incretin peptides at 100 nM and 1 µM: an approved GLP-1 receptor agonist, GIP, glucagon, and the half-life extended candidate. All six nucleoside di- and triphosphates quantified in the extracts increased under the three control peptides and decreased under the candidate, with dose in both directions. ATP decreased approximately ten-fold against the mean of the controls and ADP approximately three-fold. UTP, UDP, CDP, and GDP followed the same direction.

Figure 1. Nucleoside di- and triphosphate concentrations. Values in µM across vehicle, the three control peptides at two doses each, and the half-life extended candidate at two doses.

10×

decrease in ATP under the candidate against the mean of all control peptides, dose-ordered and absent across the seven control conditions.

Purine catabolism

Decreased triphosphate concentrations are consistent with reduced synthesis or with increased degradation. To distinguish these, Pyxis examined the adenine nucleotide catabolites. AMP was approximately unchanged while inosine, hypoxanthine, and xanthine increased, the direction expected for adenine nucleotide degradation under energy stress [6]. The pattern indicates consumption of the adenylate pool with disposal of the products.

Figure 2. Adenine nucleotide catabolism. Log2 fold change at 1 µM against the mean of all control peptides, ordered along the catabolic pathway.

Arginine and urea cycle intermediates

A second set of changes was confined to the arginine axis. Argininosuccinate, the immediate precursor of arginine, increased approximately fifty-fold at the top dose, while citrulline one step upstream changed by less than two-fold. Arginine reached 125 µM against 17 µM in vehicle. Three arginine disposal products also increased: creatinine approximately eight-fold, 4-guanidinobutyrate approximately twelve-fold, and asymmetric dimethylarginine approximately three-fold.

Figure 3. Arginine axis metabolite concentrations. Values in µM on a log scale across all nine conditions.

Six analytes along one pathway changed in the same direction, at the same doses, under one peptide. The distribution of those changes, with accumulation at argininosuccinate and arginine and little movement at citrulline, is consistent with flux entering the arginine arm faster than it is cleared.

Protein degradation as an alternative explanation

Free arginine accumulation admits an alternative explanation. The candidate is the only Fc fusion in the panel and delivers more protein than the control peptides. Proteolysis of that protein would release free arginine, methylated arginines, and dipeptides with no cellular pathway involved, and several of the species that increased are consistent with such a process. Our scientists raised this objection.

Pyxis addressed it using species that discriminate between the two sources. Argininosuccinate and creatinine are not protein constituents and cannot be released by degrading a dosed protein. Both are among the most dose-dependent species in the set, with larger changes than the species proteolysis could supply. Cellular pathway activity accounts for both groups. This is poly-intelligence in practice: an objection from human scientists resolved by returning to independent evidence within the same measurement.

Figure 4. Fold change by species origin. Log2 fold change against vehicle at 1 µM, grouped by whether the dosed Fc fusion could supply the species directly.

Implications for lead selection

At the optimization stage this profile supports exclusion of the molecule. Two dose-dependent signatures are confined to a single peptide of the four tested, and both indicate cellular dysfunction. Nothing in the profile corresponds to productive receptor engagement, and the three control peptides on the same plate establish the profile associated with engagement in these cells.

Selection for this program used two criteria, binding and half-life, neither of which reports functional consequence. Chemotyping supplies a third measurement that does, at a per-sample cost compatible with an optimization loop. With that measurement available, a series can be ranked on functional consequence alongside binding and persistence, and molecules carrying this signature fall in the ranking.

This candidate failed downstream. The signature reported here was available from an overnight experiment at the stage where the series was still under evaluation, and a program using it might have advanced a different molecule.

Across the 1,007 analytes quantified in every sample, unsupervised clustering does not separate the candidate from the controls, and at 100 nM its whole-profile distance is closest to vehicle. The changes reported here are confined to two pathways, so a global distance measure applied to the same matrix does not recover them.

Limitations

The MS2 sample sets contain one sample per condition. No statistical test is available, and the fold changes reported here are single-sample observations. Their support comes from dose ordering within each peptide, absence across seven control conditions, and consistency among analytes within two pathways. This does not substitute for a significance test. Argininosuccinate falls below fragmentation-confirmable level outside the top-dose sample, so its baseline is a model estimate and the fifty-fold value should be treated as an order of magnitude. Replicate treatments and a matched analysis of the conditioned media are required for confirmation.

Annotation review

Annotations from the biochemome are treated as hypotheses. Each identification is assessed against accurate mass, isotope pattern, retention behaviour, and fragmentation, and surviving calls are graded by confidence. In this dataset that assessment retained argininosuccinate, which carries a fragmentation match at 0.94 similarity and a mass error of 0.03 ppm in the sample where it accumulates. It excluded several annotations inconsistent with the experimental system, among them a withdrawn anti-inflammatory drug, an agricultural fungicide, and a plasticizer, each with low fragmentation support.

For the New Insight Series, we asked Pyxis to revisit datasets where a decision had already been taken and to look for information the original workflow did not access. Here a molecule advanced on binding and half-life data, and subsequently failed, carries in its biochemome a coupled energetic and nitrogen-handling signature that was available at the stage where the series was being ranked.

Data: internal LC-MS acquisition from a peptide therapeutic screening collaboration, Z-HILIC and reversed-phase methods, MS1 and MS2, high-resolution Orbitrap. Cell extracts from a pancreatic adenocarcinoma line, nine treatment conditions, overnight exposure. All figures derive from the Pyxis reanalysis of those raw files. The partner organization and the contract research organization are not named at their request. Enquiries: info@matterworks.ai.

References

Harrison, R. K. Phase II and phase III failures: 2013-2015. Nat. Rev. Drug Discov. 15, 817-818 (2016). doi:10.1038/nrd.2016.184
Swinney, D. C. & Anthony, J. How were new medicines discovered? Nat. Rev. Drug Discov. 10, 507-519 (2011). doi:10.1038/nrd3480
Jones, B. et al. Targeting GLP-1 receptor trafficking to improve agonist efficacy. Nat. Commun. 9, 1602 (2018). doi:10.1038/s41467-018-03941-2
Engineering of a GLP-1 analogue peptide/anti-PCSK9 antibody fusion for type 2 diabetes treatment. Sci. Rep. 8 (2018). doi:10.1038/s41598-018-35869-4
Schuhknecht, L. et al. A human metabolic map of pharmacological perturbations reveals drug modes of action. Nat. Biotechnol. 43, 1996-2008 (2025). doi:10.1038/s41587-024-02524-5
Atkinson, D. E. The energy charge of the adenylate pool as a regulatory parameter. Interaction with feedback modifiers. Biochemistry 7, 4030-4034 (1968). doi:10.1021/bi00851a033
Ferro, L. S. et al. A scalable approach to absolute quantitation in metabolomics. bioRxiv (2024). doi:10.1101/2024.09.09.609906
Asher, G., Delmar, M. C., Campbell, J. M., Geremia, J. & Kassis, T. LSM1-MS2: a foundation model for MS/MS, encompassing chemical property predictions, search and de novo generation. chemRxiv (2024). doi:10.26434/chemrxiv-2024-k06gb-v3