Knowledge-enhanced AI to Supercharge ADC Development for Treatment of Cancer

Originally posted in ADC Review

Targeted therapeutic approaches to cancer are crucial for minimizing off-target toxicity and adverse events while maintaining efficacy against tumor cells. Recently, several different targeted therapeutic approaches have been investigated. Antibody-drug conjugates (ADCs) represent innovative pharmaceutical compounds designed for targeted cancer therapy with a low risk of toxic events. These drug molecules consist of an antibody, or antigen-binding protein, that selectively binds to specific antigens expressed on cancer cells. Attached to the antibody via a linker is a potent cytotoxic drug. Once the antibody binds to the cancer cell’s surface antigen, the ADC is endocytosed, and the drug is delivered directly to the cancer cell, leading to cell death while sparing healthy cells. This targeted approach minimizes systemic toxicity and enhances therapeutic efficacy, making ADCs a promising strategy in the field of precision medicine for cancer treatment. [1][2]

Even though ADCs are being increasingly investigated with over 100 ADCs currently in clinical trials, only 13 ADC treatments are currently approved by the US Food and drug Administration (FDA) – representing a small fraction of available cancer treatments. [2] The complexity of the molecules – consisting of an antibody, a toxic payload, and a linker – leads to several challenges in the development of ADCs. These challenges that relate to the safety and efficacy of the ADCs include quantification of dose-exposure of the ADC, its plasma stability, the impact of post-translational modifications, and drug-antibody ratio (DAR), payload potency, payload organ exposure, and payload toxicity. Given the wide-range of challenges, comprehensive technologies to address these challenges are required.

AI and drug discovery/development

‍
The use of Artificial Intelligence (AI) is becoming increasingly prevalent in drug discovery and development [3][4] and holds the potential to address the challenges in ADC development. AI requires substantial amounts of high-quality data to make reliable predictions of a novel ADC’s safety and efficacy. However, since this therapeutic area is still fairly new, there is limited data available to meet all the challenges of ADC development through AI. In such a scenario, enriching AI with knowledge (hybrid AI) is a very promising approach and provides high-accuracy predictions even in data-limited scenarios. [5]

Such knowledge enrichment of AI models can take several forms. One form is the utilization of knowledge based models to generate data for inputs to AI models to address data limitations. For example, experimental binding affinity predictions between a target and ligands can be enriched with high quality molecular-modeling generated binding affinity data. Another form is the utilization of knowledge-based models to find patterns or quantifications from limited data and use those patterns/quantifications as inputs to the AI models.

An example of that is leveraging Bioinformatics techniques to enrich information from omics data and using that information as inputs to AI models.[6] Yet another form is utilizing knowledge to guardrail AI and improve biological relevance of the AI models. An example of this is using physiologically-based pharmacokinetic (PBPK) or other quantitative models for enabling interspecies translation and using such translation metrics as inputs to AI models to improve their robustness of toxicity predictions from one species to another. [5]

The following describes the use of such hybrid AI models in addressing myriad of challenges in ADC development for targeted cancer therapeutics.

Safety and efficacy

‍
Let us start this discussion with the essential goal: assessing efficacy and safety of novel ADC candidates.

Both the small-molecule payload and the antibody of an ADC can have safety concerns. Payloads can result in substantial hepatotoxic and other cytotoxic concerns. On the other hand, antibodies can lead to undesirable immunogenic responses. For adverse events to occur, the chemical moiety of interest (ADC/antibody or the small-molecule payload) must be present in high concentrations at specific sites within the body. On the other hand, for the ADC to be efficacious, a minimum ADC concentration is required at the site of action, which may result in a narrow therapeutic window. [1]

Hybrid AI is a powerful paradigm to assess comprehensive safety and toxicity aspects of ADCs. The connection between an ADC and specific efficacy and safety outcomes starts with conversion of the chemical moieties (ADC, antibodies, and the small-molecule payload) into ‘descriptors.’ These descriptors are sets of machine-understandable quantitative values that represent the molecular moieties. These knowledge-based descriptors can include physicochemical, structural, and quantum mechanical features of these moieties.

Once the chemical moieties are converted into machine-understandable values, AI models can be trained to predict physiological parameters of the ADCs, antibodies, and the payload such as their clearance rates. These physiological parameters can then be incorporated into knowledge-based mechanistic models to predict concentration profiles of the chemical moieties in plasma and different tissues. [7] AI models, enriched with knowledge-enhanced descriptors, can also be used to predict tumor cell growth inhibition due to the payload across hundreds of cancer cell lines across different types of cancers rapidly. Similar models can also quantify a payload’s off-target toxicity by rapidly assessing its binding to thousands of different off-target proteins. Similarly, immunogenic potential of the parent antibody can also be assessed using such models.

The safety and efficacy of ADCs is further impacted by linker plasma stability and modifications, including post-translational modifications, to the ADC molecule.

Cleavabe vs uncleavable

‍
ADC linkers may be either cleavable or uncleavable. Cleavable linkers take advantage of the different environments that exist between plasma and intracellular environments. These linkers may be acid-sensitive, where they are more likely to hydrolyze at the more acidic pHs of lysosomes (pH 4.8) and endosomes (pH 5.5-6.2) than the slightly basic pH of plasma (pH 7.4). [8] Alternatively, linker cleavage may be catalyzed by molecules in the body such as the antioxidant glutathione (GSH) or the enzyme Cathepsin B, both of which have higher concentrations in cancer cells than in plasma. [9][10] Plasma stability of these cleavable linkers thus plays a critical role in both efficacy and off-target toxicity. For example, if a linker cleaves prematurely and the ADC releases its toxic payload outside the target tumor cells, high off-target toxicity and low efficacy is likely.

Hybrid-AI can provide the ability to rapidly screen different linkers for their likelihood to hydrolyze at higher pHs or interact with enzymes such as human Cathepsin B. For the latter scenario in particular, the BindingDB database [11] contains experimental assays and results Ki, IC₅₀, and EC₅₀ values for over 2,700 small molecules with Cathepsin B. This data may be used as training data to develop AI models to predict the interaction of linkers with this protease.

As mentioned above, the inputs to AI models are of critical importance. At VeriSIM Life,  we curate specialized descriptor sets for each machine learning model. For an AI model to predict the interaction strength of Human Cathepsin B with an ADC, one set of descriptors may then be from the similarity between the linker in question and small molecules that demonstrate significant binding with Human Cathepsin B. Common motifs of molecules that interact strongly with Human Cathepsin B may be identified, and AI model inputs may then be the presence of such motifs in the linkers. However, one must also consider the spatial bulk of both the antibody itself and the payload. An additional set of descriptors that captures the physics of the spatial relationship between ADC and enzyme may be generated by analyzing the ADC’s three-dimensional structure in the region of each linker and payload to determine if there is sufficient room for the enzyme to interact with and cleave the linker. Such descriptors may include the TPSA (topological polar surface area) and the solvent accessible surface area (SASA) of the linker atoms when it is in complex with the entire ADC. Furthermore, the distribution of electrons about the linker will be modified by the presence of the antibody and the payload. Advanced computational chemistry methods may be used to, for example, approximate partial charges of atoms that may be hydrogen bond donors or acceptors that interact with atoms in the Human Cathepsin B binding pocket and further increase the predictive power of the machine learning model. [4]

Generative AI technology can also be used to create synthetic data for chemical moieties (or, their descriptor representations) to enrich the binding interaction data between Human Cathepsin B and the ADC of interest and enable more robust predictions.

Next, let us discuss DAR, conjugation sites, and post-translational modifications of ADCs and how hybrid AI can help assess the impact of those on ADC performance.

Unlike traditional small molecule drug treatments, ADC treatments do not have a discrete active pharmaceutical ingredient with a specific molecular structure. This is due to several effects such as conjugation methods and post-translational modifications (PTMs).

Conjugation methods refer to how a linker and payload are attached to the antibody. Multiple conjugation methods exist, some of which result in a more random distribution of conjugation sites as in lysine and cysteine residue amide coupling reactions, [12] while others are more site-specific such as what is used in ThioMab technology. [13] However, all conjugation methods result to an extent in a variable number of linkers, and thus payloads, being attached to the antibody – an idea known as the drug-antibody ratio or DAR. The heterogeneous products of these conjugation methods is reflected in the DAR of most current clinical stage ADCs of 3.5, which is notably not a discrete integer.[1]

Post-translational modifications often occur during production, purification, storage, and even after treatment administration. These modifications include glycosylation, deamidation, and oxidation and result in variable, heterogeneous structures. (14).

The overall effect of multiple possible conjugation sites, variable DARs, and heterogeneous PTMs is that a single ADC treatment consists of multiple discrete ADC structures, each with their own efficacy potentials and safety concerns.

Given the combinatorial nature of such heterogeneity, the different structures of an ADC may not be able to be isolated and analyzed experimentally, and traditional computational chemistry methods such as molecular modeling become intractably expensive. Hybrid computational chemistry and artificial intelligence approaches offer the ability to rapidly screen such inhomogenous formulations as these methods leverage both knowledge-based chemistry descriptors as well as any limited experimental data available.

One such case in which AI may be particularly effective is by assessing whether a particular ADC may be negatively impacted by a conjugation site too near to an antigen receptor. If a conjugation site is too close to the antigen-binding site, both steric hindrance and electronic effects can influence the interaction strength between the antibody and the antigen receptor. Here, hybrid-AI may be leveraged to inform whether the total net effects of each discrete structure result in an overall efficacy and safety profile for the ADC that may be concerning.

Conclusion

‍
ADCs are targeted therapeutics for the treatment of a variety of cancer. The emergence of ADCs is advancing new treatment options for patients, but also stands to benefit from the enhancement of artificial intelligence in their development. Despite the small amount of data available due to the nascency of the field, the ability to model and make predictions about quantification of dose-exposure of the ADC, its plasma stability, the impact of post-translational modifications and drug-antibody ratio (DAR), payload potency, payload organ exposure, and payload toxicity can be made possible via data enrichment from knowledge based models including ab initio computational chemistry and PBPK models.

References
[1] Beck A, Goetsch L, Dumontet C, Corvaïa N. Strategies and challenges for the next generation of antibody-drug conjugates. Nat Rev Drug Discov. 2017 May;16(5):315-337. doi: 10.1038/nrd.2016.268. Epub 2017 Mar 17. PMID: 28303026. https://doi.org/10.1038/nrd.2016.268
[2] Fu Z, Li S, Han S, Shi C, Zhang Y. Antibody drug conjugate: the “biological missile” for targeted cancer therapy. Signal Transduct Target Ther. 2022 Mar 22;7(1):93. doi: 10.1038/s41392-022-00947-7. PMID: 35318309; PMCID: PMC8941077. https://doi.org/10.1038/s41392-022-00947-7
[3] Paul D, Sanap G, Shenoy S, Kalyane D, Kalia K, Tekade RK. Artificial intelligence in drug discovery and development. Drug Discov Today. 2021 Jan;26(1):80-93. doi: 10.1016/j.drudis.2020.10.010. Epub 2020 Oct 21. PMID: 33099022; PMCID: PMC7577280. https://doi.org/10.1016/j.drudis.2020.10.010
[4] Dara S, Dhamercherla S, Jadav SS, Babu CM, Ahsan MJ. Machine Learning in Drug Discovery: A Review. Artif Intell Rev. 2022;55(3):1947-1999. doi: 10.1007/s10462-021-10058-4. Epub 2021 Aug 11. PMID: 34393317; PMCID: PMC8356896. https://doi.org/10.1007/s10462-021-10058-4
[5] Bhave R, Bello B, Bhatt D, Shea J, Khotimchenko M, Varshney J, Tong W, Baran S. Importance of integrating transcriptomics with quantum mechanics and physiological insights to enable robust DILI predictions in data-limited scenarios with artificial intelligence. [Publication pending]
[6] Bello B, Bundey YN, Bhave R, Khotimchenko M, Baran SW, Chakravarty K, Varshney J. Integrating AI/ML Models for Patient Stratification Leveraging Omics Dataset and Clinical Biomarkers from COVID-19 Patients: A Promising Approach to Personalized Medicine. Int J Mol Sci. 2023 Mar 26;24(7):6250. doi: 10.3390/ijms24076250. PMID: 37047222; PMCID: PMC10094615. https://doi.org/10.3390/ijms24076250
[7] Antontsev V, Jagarapu A, Bundey Y, Hou H, Khotimchenko M, Walsh J, Varshney J. A hybrid modeling approach for assessing mechanistic models of small molecule partitioning in vivo using a machine learning-integrated modeling platform. Sci Rep. 2021 May 27;11(1):11143. doi: 10.1038/s41598-021-90637-1. PMID: 34045592; PMCID: PMC8160209. https://doi.org/10.1038/s41598-021-90637-1
[8] Nolting B. Linker technologies for antibody-drug conjugates. Methods Mol Biol. 2013;1045:71-100. doi: 10.1007/978-1-62703-541-5_5. PMID: 23913142.
[9] Estrela JM, Ortega A, Obrador E. Glutathione in cancer biology and therapy. Crit Rev Clin Lab Sci. 2006;43(2):143-81. doi: 10.1080/10408360500523878. PMID: 16517421.
[10] Gondi CS, Rao JS. Cathepsin B as a cancer target. Expert Opin Ther Targets. 2013 Mar;17(3):281-91. doi: 10.1517/14728222.2013.740461. Epub 2013 Jan 8. PMID: 23293836; PMCID: PMC3587140.
[11]Tiqing Liu, Yuhmei Lin, Xin Wen, Robert N. Jorissen, Michael K. Gilson, BindingDB: a web-accessible database of experimentally determined protein–ligand binding affinities, Nucleic Acids Research, Volume 35, Issue suppl_1, 1 January 2007, Pages D198–D201, https://doi.org/10.1093/nar/gkl999
[12] Brun MP, Gauzy-Lazo L. Protocols for lysine conjugation. Methods Mol Biol. 2013;1045:173-87. doi: 10.1007/978-1-62703-541-5_10. PMID: 23913147.
[13] Junutula JR, Raab H, Clark S, Bhakta S, Leipold DD, Weir S, Chen Y, Simpson M, Tsai SP, Dennis MS, Lu Y, Meng YG, Ng C, Yang J, Lee CC, Duenas E, Gorrell J, Katta V, Kim A, McDorman K, Flagella K, Venook R, Ross S, Spencer SD, Lee Wong W, Lowman HB, Vandlen R, Sliwkowski MX, Scheller RH, Polakis P, Mallet W. Site-specific conjugation of a cytotoxic drug to an antibody improves the therapeutic index. Nat Biotechnol. 2008 Aug;26(8):925-32. doi: 10.1038/nbt.1480. Epub 2008 Jul 20. PMID: 18641636.
[14] Xu X, Huang Y, Pan H, Molden R, Qiu H, Daly TJ, Li N. Quantitation and modeling of post-translational modifications in a therapeutic monoclonal antibody from single- and multiple-dose monkey pharmacokinetic studies using mass spectrometry. PLoS One. 2019 Oct 16;14(10):e0223899. doi: 10.1371/journal.pone.0223899. PMID: 31618250; PMCID: PMC6795451. https://doi.org/10.1371/journal.pone.0223899

‍