Antibody-drug conjugate engineering: from design to efficacy and safety

Abstract

Antibody–drug conjugates (ADCs) represent a rapidly expanding class of targeted cancer therapeutics that combine the high selectivity of monoclonal antibodies with the potent cytotoxic activity of small-molecule drugs. Their clinical success relies on the simultaneous optimization of multiple interdependent parameters, including antigen selection, antibody engineering, linker chemistry, and payload pharmacology, which limits the effectiveness of traditional empirical approaches. In this context, recent advances in artificial intelligence (AI) and computational biophysics are transforming the rational design of ADCs. AI enables large-scale integration of genomic, transcriptomic, and proteomic data to identify tumor-selective, surface-accessible antigens and to support patient stratification strategies. Deep learning models enhance antibody engineering by predicting structure, affinity, stability, and developability, while generative algorithms accelerate affinity maturation and specificity optimization. Computational prediction of linker design and conjugation sites improves plasma stability, controlled payload release, and drug-to-antibody ratio, whereas graph-based neural networks facilitate the selection and optimization of cytotoxic payloads with favorable potency, membrane permeability, and bystander effects. Complementary molecular dynamics simulations provide atomistic insight into antibody conformation, linker flexibility, and payload interactions, enabling a deeper mechanistic understanding of ADC stability and function. At the translational level, hybrid physiologically based pharmacokinetic–AI models and digital twin simulations enable virtual evaluation of tumor penetration, systemic exposure and safety, ultimately supporting dose optimization and more efficient clinical development. Although this review places particular emphasis on the application of these approaches in oncology, emerging ADC strategies beyond cancer—including autoimmune, neurodegenerative, cardiovascular, and metabolic diseases—are also discussed. Together, these computational strategies represent a convergence of machine intelligence and molecular biology that is poised to fundamentally transform ADC development, enabling safer, more precise, and effective therapies across oncology and beyond.

Introduction

Antibody–drug conjugates (ADCs) represent a rapidly evolving class of targeted therapeutics that is reshaping the oncology landscape.^1,2 The underlying concept is deceptively simple: a monoclonal antibody directed against a tumor-associated antigen is used to deliver a highly potent cytotoxic agent selectively to malignant cells, thereby widening a therapeutic window that would be unacceptably narrow were the same payload administered systemically.¹ In practice, therapeutic activity depends on the successful completion of an extended delivery cascade—antigen engagement at the cell surface, internalization and intracellular trafficking, linker cleavage or antibody catabolism, payload release into the cytoplasm, and, for membrane-permeable payloads, diffusion into neighboring antigen-negative cells—each step of which can fail independently.^3,4 The clinical realization of this concept has been protracted. Decades separated the original proposal from the first approvals, and early agents were limited by unstable linkers, heterogeneous conjugation and off-target toxicity.² Successive generations addressed these liabilities through site-specific conjugation, more stable cleavable chemistries, controlled drug-to-antibody ratios (DAR) and payloads with distinct mechanisms of action, culminating in agents that have redefined treatment standards across several tumor types.³ The current generation has also overturned assumptions once regarded as settled, most notably the requirement for high antigen expression, following the demonstrated activity of trastuzumab deruxtecan in HER2-low and HER2-ultralow disease.^4,5 This has questioned the assumption that the main mechanism of action of ADCs is by an improvement of the therapeutic window.⁶

The field has consequently entered a phase of rapid expansion. A succession of regulatory approvals across hematologic and solid tumors, a preclinical and clinical pipeline comprising several hundred candidates, a global market projected to surpass $30 billion by 2028, and individual licensing transactions exceeding $10 billion together underscore the clinical and commercial momentum of this modality.^1,2 Design strategies have diversified in parallel, encompassing bispecific constructs that engage two antigens or two epitopes of the same antigen, dual-payload architectures intended to pre-empt resistance, immunomodulatory targets such as PD-L1, and non-internalizing designs that exploit proteolytic activity within the tumor microenvironment.^4,7,8 Yet attrition remains high, and the reasons are structural rather than incidental. Rational ADC design requires the coordinated integration of target biology, protein engineering, medicinal chemistry and pharmacology, and optimal constructs arise only from the simultaneous selection of an appropriate antigen, fine-tuning of antibody affinity and internalization kinetics, engineering of linkers that combine plasma stability with efficient intracellular cleavage, and incorporation of payloads with sub-nanomolar potency and adequate physicochemical properties.^3,4 These parameters are interdependent, and each design decision entails a trade-off between efficacy, safety and manufacturability across a combinatorial parameter space that empirical, component-by-component optimization cannot realistically traverse.⁸ Dose-limiting toxicities frequently prove to be payload-mediated and antigen-independent, and are therefore poorly anticipated by target-centric preclinical strategies.⁹

It is against this background that artificial intelligence (AI) and computational biophysics have begun to alter the design paradigm.^6,7 By integrating multi-omic datasets with historical experimental outputs, machine learning models can streamline target prioritization and support patient stratification; protein language models and structure prediction frameworks inform antibody engineering, developability assessment and linker–payload pairing; and generative architectures now propose candidate sequences and molecules rather than merely ranking existing ones.^6–8 Complementing these data-driven methods, physics-based simulation provides atomistic insight into antibody conformation, conjugation-site accessibility, linker flexibility and payload interactions without dependence on training data, and is therefore particularly valuable in a field where curated datasets remain sparse.¹⁰ At the translational level, hybrid physiologically based pharmacokinetic–AI models and digital twin simulations permit virtual evaluation of tumor penetration, systemic exposure and toxicity, supporting dose optimization and more efficient trial design.¹¹ Collectively, these approaches have the potential to shorten development timelines and reduce costs while enhancing design precision and translational success, and thereby to reduce the high attrition characteristic of ADC development.^6,9 Important caveats nonetheless apply, including sparse and unbalanced training data, limited interpretability, poor calibration, and a near-absence of prospective validation.

In this review, we examine how computational and AI-driven approaches are being applied to each key component of ADC engineering design, highlighting recent developments and real-world examples drawn from both academia and industry. We begin with antigen discovery, covering target identification from integrated surfaceome and multi-omic data, tumor heterogeneity, epitope mapping, antigen internalization and neoantigen prediction. We then address AI-enhanced antibody design, including affinity maturation, specificity optimization and developability, followed by linker design and conjugation-site engineering and by payload optimization with respect to potency, hydrophobicity, bystander activity and dual-payload combinations. A dedicated section considers physics-based computational frameworks and their complementarity to machine learning. We subsequently discuss the prediction of efficacy and safety, encompassing physiologically based pharmacokinetic (PBPK) modeling, tumor penetration, toxicity prediction and digital twin trials, and devote separate attention to resistance and temporal dynamics, which we regard as the most consequential unexplored computational frontier in the field. Finally, we consider emerging ADC applications beyond oncology and critically appraise the limitations, validation requirements and regulatory considerations that will determine whether these computational engineering strategies fulfill their translational promise.

AI for antigen discovery

Target identification

Identifying a tumor antigen with high expression on cancer cells and minimal expression in normal tissues is a critical first step in ADC development.^12,13 AI may be used for antigen discovery by integrating large-scale multi-omics datasets to systematically identify and prioritize candidate targets with a level of speed and precision beyond conventional approaches.^14–16 Candidate antigens are ranked according to differential expression, tissue specificity, and cell-surface localization to maximize tumor selectivity.^15,17,18

Additional parameters that can be integrated include spatial transcriptomics, functional genomics such as gene set enrichment analyses, single-cell sequencing, and clinical data on patient outcomes or treatment responses¹⁹ (Fig. 1). An excellent review describing the different methodologies is reported by You et al.¹⁷ Additional parameters, e.g., glycosylation pattern, internalization, and/or degradation of the extracellularly-exposed target antigen, could be later implemented to improve selection of the ADC target. Current computational frameworks also enable integration of protein degradation rates, recycling kinetics, and receptor turnover data, further refining the prioritization of ADC targets with favorable antigen dynamics.^20,21

Fig. 1.

AI-driven discovery and evaluation of ADC targets. a Computational workflow for identifying tumor antigens suitable for antibody–drug conjugate (ADC) development. Multi-omic data (genomics, transcriptomics, proteomics) are combined through machine learning–based integration and feature selection; candidate antigens are then characterized structurally using protein structure prediction tools (AlphaFold, ESMFold), and clustering and classification approaches support patient stratification. Candidates are prioritized on the basis of expression level and internalization capacity, with digital pathology and AI-based image analysis providing orthogonal validation of antigen distribution in tissue. b Key biological and molecular criteria governing ADC target selection: high tumor specificity, minimal expression in normal tissues, efficient internalization following antibody binding, and balanced target recycling—together determining the potency and selectivity of payload delivery

A critical limitation of transcriptomics-only approaches is the imperfect correlation between mRNA abundance and actual cell-surface protein density, which is the parameter that directly governs ADC binding and internalization efficiency.^22,23 Mass spectrometry-based surface proteomics, including cell surface capture (CSC) technology and proximity-labeling strategies such as TurboID and APEX2, enable direct quantification of glycoproteins on the plasma membrane of living tumor cells, providing copy-number estimates that are inaccessible from RNA data alone.^24,25 Integration of these proteomic datasets with transcriptomic and genomic resources—as exemplified by the Cell Surface Protein Atlas (CSPA) covering 41 human cell types—allows systematic identification of targets with confirmed surface localization and high differential expression between tumor and normal tissues.²⁴ Recent advances in microscaled surface proteomics have further extended this approach to cryopreserved primary patient samples and low-input clinical material, facilitating direct translation from discovery to patient-level target validation.²⁶ Together, these orthogonal datasets strengthen target prioritization frameworks by grounding candidate selection in measured protein abundance rather than inferred expression.

As a proof-of-concept for the computational in silico strategy, investigators profiled the surfaceome of KRAS^G12C-mutant non–small cell lung (NSCLC) adenocarcinomas and identified cell-surface proteins with therapeutic potential as novel antibody targets.²⁷ This type of strategy has also been used to identify CAR-T and ADC targets in breast cancer.²⁸ Such data can inform the selection of ADC targets, as well as guide indication prioritization and clinical development strategies for individual ADC candidates. In line with this, the anti-HER3 ADC patritumab deruxtecan (HER3-DXd) is currently in clinical development for patients with EGFR-mutant NSCLC who have progressed following Epidermal Growth Factor Receptor (EGFR) tyrosine kinase inhibitor therapy.²⁹ A classical in silico evaluation of data from patients confirmed the presence of HER3 in that patient population.³⁰ Preclinical in vivo models demonstrated marked sensitivity to anti-HER3 ADCs, underscoring the translational potential of this therapeutic approach.^31,32 This example illustrates a broader translational paradigm: computational surfaceome profiling identifies candidate targets, preclinical in silico validation confirms biological rationale, and clinical development is subsequently guided by biomarker-enriched patient selection strategies. Of note, this strategy is based on the presence of biomarker correlations that would need to be validated in prospective studies to be clinically implemented.

Beyond expression-level criteria, target prioritization must account for functional properties that directly modulate ADC delivery. Antigen shedding—the proteolytic release of the ectodomain into the extracellular space— creates soluble decoy molecules that can sequester circulating ADCs before they reach tumor cells, reducing effective tumor exposure.³³ Mathematical modeling has shown that the impact of shedding is context-dependent.³⁴ AI-driven frameworks that integrate shedding rates, receptor turnover, and copy-number data can therefore refine the rank-ordering of candidate targets beyond static expression metrics alone.

Equally important is the concept of target druggability in the antibody context: not all surface proteins that are differentially expressed present accessible, high-confidence epitopes for antibody engagement. Glycosylation, steric shielding by the glycocalyx, proximity to the membrane, and protein crowding can all mask potential binding sites in the native tumor microenvironment.³⁵ AlphaFold3, with its extended capacity to model glycosylated proteins and protein–ligand complexes, now enables computational prediction of epitope accessibility in the context of post-translational modifications, providing a more realistic assessment of antibody engageability prior to experimental campaign initiation.^36,37 Current AI prediction models, however, remain primarily trained on unmodified protein structures and have limited systematic ability to capture how glycan microheterogeneity dynamically shields epitopes, which remains an important area for future development.³⁷

Concomitant expression of surfaceome receptors has traditionally been assessed using transcriptomic analyses, as exemplified by studies of EGFR and MET co-expression in NSCLC.³⁸ Similarly, EGFR and c-MET are frequently co-expressed across multiple solid tumors—including colorectal, gastric, esophageal, and head and neck cancers— and this co-expression pattern has directly informed the rational development of bispecific ADCs such as PRO1286, designed to simultaneously engage both receptors and overcome resistance to EGFR-directed monotherapy, and SDP01873, which pairs c-Met with HER3 to address EGFR tyrosine kinase inhibitor-resistant disease.^39,40 AI enables rapid and large-scale interrogation of receptor co-expression patterns, revealing clinically actionable target combinations. By identifying co-expressed or functionally synergistic antigens, these approaches inform the rational design of bispecific antibodies, guided to target both membrane proteins and other multi-targeted therapeutic strategies.⁴¹ Such strategies have the potential to augment therapeutic efficacy, analogous to the clinical benefit observed with co-targeting PD-L1 and Vascular Endothelium Growth Factor Receptor (VEGFR).⁴² In addition, the use of immunomodulator receptors as the base for the development of ADCs has been observed with those targeting PD-L1. As an example PDL1V (PF-08046054) is an ADC targeting PD-L1 with vedotin-based payload that has entered phase I in NSCLC.⁴³ Similarly, HLX43 is being assessed in a phase I study across multiple solid tumor types.⁴⁴

Table 1 provides a comprehensive list of ADC targets, cancer types, and current clinical stage.

Table 1.

Antibody–drug conjugate targets, associated agents and regulatory status

Target	ADC (INN; brand/development code)	Developer(s)	Payload class	Principal tumor type(s)	Development status	Approved indication(s) — year/region
A. Approved antibody–drug conjugates
HER2	Trastuzumab emtansine (Kadcyla; T-DM1)	Roche/Genentech	DM1 (maytansinoid)	Breast	Approved	HER2+ metastatic breast cancer (2013 US; 2013 EU); early breast cancer with residual invasive disease after neoadjuvant therapy (2019 US; 2020 EU)
HER2	Trastuzumab deruxtecan (Enhertu; T-DXd)	Daiichi Sankyo & AstraZeneca	DXd (topoisomerase I inhibitor)	Breast, gastric, NSCLC, pan-tumor	Approved	HER2+ breast (2019 US; 2021 EU); HER2+ gastric (2021 US; 2023 EU); HER2-low breast (2022 US; 2023 EU); HER2-mutant NSCLC (2022 US); HER2 IHC3+ pan-tumor, tissue-agnostic (2024 US); HER2-low and HER2-ultralow breast (2025 US/EU)
HER2	Disitamab vedotin (Aidixi; RC48)	RemeGen	MMAE	Gastric, urothelial	Approved (China only)	HER2+ locally advanced/metastatic gastric cancer (NMPA 2021); HER2+ locally advanced/metastatic urothelial carcinoma (NMPA 2021, conditional). Not approved by FDA or EMA
HER2	Trastuzumab rezetecan (Aiweida; SHR-A1811)	Jiangsu Hengrui Pharmaceuticals	Topoisomerase I inhibitor	Breast, gastric, biliary tract	Approved (China only)	HER2-expressing solid tumors (NMPA 2025). Not approved by FDA or EMA
HER2	Trastuzumab botidotin (Shutailai; A166)	Sichuan Kelun-Biotech	Auristatin-class	Breast	Approved (China only)	HER2+ metastatic breast cancer (NMPA 2025). Not approved by FDA or EMA
TROP2	Sacituzumab govitecan (Trodelvy)	Gilead Sciences	SN-38 (topoisomerase I inhibitor)	Breast (TNBC and HR + /HER2 − )	Approved	Metastatic TNBC (2020 US; 2021 EU); HR + /HER2− metastatic breast cancer (2023 US/EU). Urothelial carcinoma accelerated approval voluntarily withdrawn (2024 US) after TROPiCS-04 did not meet its primary endpoint
TROP2	Datopotamab deruxtecan (Datroway; Dato-DXd, DS-1062a)	Daiichi Sankyo & AstraZeneca	DXd (topoisomerase I inhibitor)	Breast, NSCLC	Approved	HR + /HER2− metastatic breast cancer (Jan 2025 US; Jan 2025 EU); EGFR-mutant advanced NSCLC (2025 US, accelerated); 1 L unresectable/metastatic TNBC in patients not candidates for PD-1/PD-L1 inhibitors (May 2026 US; EU CHMP positive opinion 2026)
TROP2	Sacituzumab tirumotecan (Jiatailai; sac-TMT, SKB264/MK-2870)	Sichuan Kelun-Biotech/Merck & Co.	Belotecan-derived topoisomerase I inhibitor	Breast, lung	Approved (China only); global Phase III	Advanced TNBC (NMPA 2024). Not approved by FDA or EMA
Nectin-4	Enfortumab vedotin (Padcev)	Astellas & Pfizer (formerly Seagen)	MMAE	Urothelial carcinoma only	Approved	Locally advanced/metastatic urothelial carcinoma (2019 US; 2022 EU); 1 L with pembrolizumab (2023 US; 2024 EU)
FOLR1 (FRα)	Mirvetuximab soravtansine (Elahere)	ImmunoGen/AbbVie	DM4 (maytansinoid)	Ovarian (platinum-resistant)	Approved	FRα-positive platinum-resistant ovarian cancer (2022 US, accelerated; full approval 2024 US; 2024 EU)
c-Met (MET)	Telisotuzumab vedotin (Emrelis; ABBV-399, Teliso-V)	AbbVie	MMAE	NSCLC (non-squamous)	Approved	c-Met-high (IHC 3+ in ≥50% of tumor cells), EGFR wild-type, non-squamous NSCLC after prior systemic therapy (May 2025 US, accelerated). Confirmatory Phase III TeliMET NSCLC-01 ongoing
EGFR	Becotatug vedotin (Meiyouheng; MRG003)	Lepu Biopharma	MMAE	Nasopharyngeal carcinoma	Approved (China only)	Recurrent/metastatic nasopharyngeal carcinoma (NMPA 2025). Not approved by FDA or EMA
EGFR × HER3 (bispecific)	Izalontamab brengitecan (Yizekang; BL-B01D1, iza-bren)	Sichuan Baili Pharmaceutical/Bristol Myers Squibb	Topoisomerase I inhibitor (Ed-04)	Nasopharyngeal carcinoma, esophageal squamous cell carcinoma	Approved (China only); global Phase III	Recurrent/metastatic NPC (NMPA, 22 Jun 2026)—first bispecific ADC approved worldwide; ESCC (NMPA, Jul 2026). Not approved by FDA or EMA
CD30	Brentuximab vedotin (Adcetris)	Pfizer (formerly Seagen)/Takeda	MMAE	Hodgkin lymphoma, ALCL	Approved	Classical Hodgkin lymphoma and systemic ALCL (2011 US; 2012 EU); frontline cHL, PTCL and cutaneous T-cell lymphoma indications added subsequently
CD79b	Polatuzumab vedotin (Polivy)	Roche/Genentech	MMAE	Diffuse large B-cell lymphoma	Approved	R/R DLBCL (2019 US; 2020 EU); 1 L DLBCL with R-CHP (2023 US/EU)
CD33	Gemtuzumab ozogamicin (Mylotarg)	Pfizer	Calicheamicin	Acute myeloid leukemia	Approved	CD33 + AML (2000 US, accelerated; voluntarily withdrawn 2010); re-approved at fractionated dosing (2017 US; 2018 EU)
CD22	Inotuzumab ozogamicin (Besponsa)	Pfizer	Calicheamicin	B-cell precursor ALL	Approved	R/R B-cell precursor acute lymphoblastic leukemia (2017 US/EU)
CD19	Loncastuximab tesirine (Zynlonta)	ADC Therapeutics	SG3199 (PBD dimer)	Diffuse large B-cell lymphoma	Approved	R/R DLBCL after ≥2 prior lines (2021 US, accelerated; 2022 EU, conditional)
BCMA	Belantamab mafodotin (Blenrep)	GSK	MMAF	Multiple myeloma	Approved (re-approved after withdrawal)	Monotherapy in R/R multiple myeloma after ≥4 prior lines (2020 US/EU, accelerated); US licence revoked Mar 2023 after DREAMM-3 failed. Re-approved in combination with bortezomib + dexamethasone after ≥2 prior lines (Oct 2025 US); combination approvals also granted in the EU, UK, Japan, Canada, Switzerland and Brazil (2025)
Tissue factor (TF)	Tisotumab vedotin (Tivdak)	Pfizer (formerly Seagen) & Genmab	MMAE	Cervical carcinoma	Approved	Recurrent/metastatic cervical cancer (2021 US, accelerated; converted to full approval 2024 US on the basis of innovaTV 301)
B. Investigational or discontinued antibody–drug conjugates (selected)
HER2	Trastuzumab duocarmazine (SYD985)	Byondis/medac (EU commercialization partner)	vc-seco-DUBA (duocarmycin)	Breast	Phase III (TULIP) completed; not approved in any jurisdiction	None. FDA Complete Response Letter (May 2023); EU marketing authorization application (Jivadco) withdrawn 12 Sept 2023, CHMP provisional opinion being that it could not have been authorized
HER3	Patritumab deruxtecan (HER3-DXd; U3-1402)	Daiichi Sankyo & Merck & Co.	DXd (topoisomerase I inhibitor)	Breast, NSCLC	Phase III; not approved	None. FDA Complete Response Letter (2024) on third-party manufacturing grounds; BLA subsequently withdrawn
FOLR1 (FRα)	Luveltamab tazevibulin (STRO-002)	Sutro Biopharma	Hemiasterlin (SC209)	Ovarian, endometrial	Phase II/III; not approved	None
CEACAM5	Tusamitamab ravtansine (SAR408701)	Sanofi	DM4 (maytansinoid)	NSCLC, colorectal	Discontinued (Dec 2023) after CARMEN-LC03 futility analysis	None
Mesothelin	Anetumab ravtansine (BAY 94-9343)	Bayer	DM4 (maytansinoid)	Mesothelioma, ovarian, pancreatic	Phase II; development largely discontinued	None
EGFR	Depatuxizumab mafodotin (ABT-414)	AbbVie	MMAF	Glioblastoma	Discontinued (2019) after INTELLANCE-1 interim futility	None
c-Met (MET)	Telisotuzumab adizutecan (ABBV-400, Temab-A)	AbbVie	Topoisomerase I inhibitor	Colorectal, NSCLC, gastro-esophageal	Phase III	None
DLL3	Rovalpituzumab tesirine (Rova-T)	AbbVie	SC-DR002 (PBD dimer)	Small-cell lung cancer	Discontinued (2019) after TAHOE and MERU failures	None
GPNMB	Glembatumumab vedotin (CDX-011)	Celldex Therapeutics	MMAE	TNBC, melanoma	Discontinued (2018) after METRIC failed its primary endpoint	None
STEAP1	Vandortuzumab vedotin (DSTP3086S)	Genentech/Roche	MMAE	Prostate	Phase I; discontinued	None
ROR1	Zilovertamab vedotin (MK-2140; VLS-101)	Merck & Co. (formerly VelosBio)	MMAE	DLBCL, mantle cell lymphoma	Phase III (waveLINE program)	None
LRRC15	ABBV-085	AbbVie	MMAE	Sarcoma	Phase I; discontinued	None
SEZ6	ABBV-706	AbbVie	Topoisomerase I inhibitor	SCLC, neuroendocrine tumors	Phase I	None
PD-L1	PDL1V (PF-08046054; formerly SGN-PDL1V)	Pfizer (formerly Seagen)	MMAE	NSCLC, HNSCC	Phase I/II	None
PD-L1	HLX43	Shanghai Henlius Biotech	Topoisomerase I inhibitor	NSCLC	Phase II	None
Globo H	OBI-999	OBI Pharma	MMAE	Breast, gastrointestinal	Phase I/II	None

ALCL anaplastic large-cell lymphoma, ALL acute lymphoblastic leukemia, AML acute myeloid leukemia, DLBCL diffuse large B-cell lymphoma, DXd exatecan derivative, EMA European Medicines Agency, ESCC esophageal squamous cell carcinoma, FDA US Food and Drug Administration, HNSCC head and neck squamous cell carcinoma, MMAE monomethyl auristatin E, MMAF monomethyl auristatin F, NMPA National Medical Products Administration (China), NPC nasopharyngeal carcinoma, NSCLC non-small-cell lung cancer, PBD pyrrolobenzodiazepine, R/R relapsed or refractory, TNBC triple-negative breast cancer

The identification of novel surfaceome targets also provides a foundation for the development of bispecific antibodies capable of simultaneously engaging two distinct tumor epitopes avoiding the limitations of tumor heterogeneity.^45,46 These molecules are increasingly advancing into clinical development and represent a promising platform for the next generation of ADCs. As illustrated in Table 2, most constructs are designed to co-target tumor-associated antigens, with HER2 and EGFR emerging as the most frequently utilized targets, followed by HER3 and TROP2—reflecting their high prevalence in solid tumors and their established relevance in ADC development. Among the candidates that have progressed into clinical evaluation, representative antigen combinations include EGFR×HER3 (BL-B01D1), biparatopic HER2 × HER2 constructs (ZW49 and MEDI4276), and TROP2×HER3 (JSKN016).⁴⁶ These bispecific ADCs (bsADCs) typically incorporate cytotoxic payloads such as microtubule inhibitors (e.g., auristatins, tubulysins, maytansinoids) or topoisomerase I inhibitors, commonly linked via cleavable linkers and with drug-to-antibody ratios (DAR) typically in the range of 2–8.⁴⁶

Table 2.

Representative bispecific and biparatopic ADCs in clinical development or approved

Drug/Program (INN)	Target 1	Target 2	Payload	Linker type	DAR	Company	Indication(s)	Clinical stage/key trials	Ref.
BL-B01D1 (izalontamab brengitecan; iza-bren)	EGFR	HER3	Topoisomerase I inhibitor (Ed-04)	Cleavable (tetrapeptide-based)	~8	SystImmune/Sichuan Baili (Biokin); BMS ex-China	Solid tumors: NPC, ESCC, NSCLC, TNBC, urothelial	Approved (China, NMPA): r/m NPC (22 Jun 2026); r/m ESCC (Jul 2026) — first bsADC approved worldwide. Phase III ongoing (BL-B01D1-303 NCT06118333; -305 ESCC; -307 TNBC). FDA BTD in EGFR-mutant NSCLC	270,271
JSKN003 (anbenitamab repodatecan) — INN added	HER2 (ECD2)	HER2 (ECD4)	Topoisomerase I inhibitor	Cleavable (glycan site-specific)	~4	Alphamab Oncology/JMT-Bio (CSPC), China rights	HER2+ and HER2-low breast; platinum-resistant ovarian (PROC); gastric; CRC	Phase III: HER2 + BC (JSKN003-301, vs T-DM1), HER2-low BC (-302), PROC (-306). Phase II: gastric, CRC. FDA BTD + Fast Track (PROC); ODD (GC/GEJ); NMPA BTD (PROC, CRC)	272
JSKN016	TROP2	HER3	Topoisomerase I inhibitor	Cleavable (glycan site-specific)	4	Alphamab Oncology	HER2-negative breast cancer/TNBC; lung cancer	Phase III in TNBC (JSKN016-301; first patient dosed Mar 2026). Phase I JSKN016-101 (NCT06592417)	273
TQB2102	HER2 (ECD2)	HER2 (ECD4)	Topoisomerase I inhibitor	Cleavable (enzyme-cleavable)	6	Chia Tai Tianqing	HER2+ and HER2-low breast cancer (incl. neoadjuvant); HER2-expressing solid tumors	Phase III (2 L HER2 + BC vs T-DM1; HER2-low BC vs chemotherapy). Phase II neoadjuvant HER2 + BC (NCT06198751); Phase I/II NCT06115902	274
AZD9592 (tilatamig samrotecan)	EGFR	c-MET	Camptothecin-derived topoisomerase I inhibitor (samrotecan, AZ14170132)	Cleavable (proprietary)	~6	AstraZeneca	NSCLC, HNSCC, colorectal cancer	Phase I/II EGRET (first-in-human), monotherapy and in combination with osimertinib	275
REGN5093-M114	MET (epitope 1)	MET (epitope 2)	Maytansinoid M24 (as linker-payload M114)	Cleavable (protease)	~3.2	Regeneron	MET-overexpressing NSCLC	Phase I/II (NCT04982224), monotherapy and with cemiplimab	276
IMGN151 (opugotamig olatansine)	FRα (epitope 1)	FRα (epitope 2)	Maytansinoid (DM21)	Cleavable (stabilized peptide)	3.5	ImmunoGen/AbbVie	Ovarian, endometrial, cervical cancer (broad range of FRα expression)	Phase I (IMGN151-1001, first-in-human dose escalation/optimization/expansion)	277
DB-1419	B7-H3 (CD276)	PD-L1	Topoisomerase I inhibitor (P1003)	Cleavable (maleimide tetrapeptide)	~8	DualityBio	Advanced/metastatic solid tumors, including PD-L1-resistant disease	Phase I/IIa first-in-human (NCT06554795); FDA IND and Australian CTN cleared	278
M1231	MUC1 (tumor-associated, hypoglycosylated)	EGFR	Hemiasterlin analog (SC209)	Cleavable (ValCit-PABA)	~4	Merck KGaA/Sutro Biopharma	Solid tumors (NSCLC, ESCC)	Phase I (NCT04695847); enrollment complete, no clinical results published — program status to be confirmed	279
ZW49 (zanidatamab zovodotin)	HER2 (ECD2)	HER2 (ECD4)	N-acyl sulfonamide auristatin (ZD02044)	Cleavable (protease)	~2	Zymeworks	HER2-expressing solid tumors	Phase I; clinical development formally discontinued Q2 2024 following portfolio prioritization	280
MEDI4276	HER2 (subdomain 2)	HER2 (subdomain 4)	Tubulysin AZ13599185 (MMETA)	Cleavable (protease; maleimidocaproyl)	4	AstraZeneca/MedImmune	HER2+ breast and gastric cancer	Phase I; discontinued — narrow therapeutic index with dose-limiting hepatic toxicity	281

BC breast cancer, bsADC bispecific antibody–drug conjugate, BTD breakthrough therapy designation, CRC colorectal cancer, CTN clinical trial notification, DAR drug-to-antibody ratio, ECD extracellular domain, ESCC esophageal squamous cell carcinoma, FRα folate receptor alpha, GC/GEJ gastric/gastro-esophageal junction, HNSCC head and neck squamous cell carcinoma, IND investigational new drug, NPC nasopharyngeal carcinoma, NSCLC non-small cell lung cancer, ODD orphan drug designation, PROC platinum-resistant ovarian cancer, r/m recurrent or metastatic, scFv single-chain variable fragment, SEED strand-exchange engineered domain, TNBC triple-negative breast cancer, T-DM1 trastuzumab emtansine

Several bsADC candidates have now advanced to late-phase clinical development, and izalontamab brengitecan became the first bispecific ADC to obtain regulatory approval in 2026, although most programs remain at early clinical or preclinical stages. Beyond dual tumor-antigen targeting, emerging design strategies also incorporate internalization receptors, tumor microenvironment-associated targets, or immune-related molecules to enhance tumor selectivity, cellular uptake, and immune engagement.⁴⁷ Together, these approaches highlight the rapidly expanding and increasingly sophisticated design landscape of bsADCs.

Current AI frameworks for target identification have predominantly been trained and validated on a limited number of tumor types, particularly breast cancer, NSCLC, and hematologic malignancies, raising questions about their generalizability to tumor contexts with distinct antigen landscapes—such as pancreatic, colorectal, or rare cancers with sparse training data.^9,48 Pan-cancer surfaceome atlases, when combined with tumor-type-specific molecular features, can extend the applicability of these models and reveal shared vulnerabilities across cancer lineages. Efforts to build interoperable multi-omics datasets that span histologies and geographic patient populations will be critical to avoid model bias toward well-represented cancer types.⁴⁹

A further underexplored dimension in target selection is the anticipation of resistance. Targets that are functionally essential—i.e., whose loss imposes a fitness cost on the tumor cell—are less likely to be downregulated as a mechanism of acquired resistance than dispensable surface proteins. AI approaches that integrate functional genomics data, such as CRISPR-based essentiality screens and dependency maps, into the target prioritization workflow can identify antigens that combine favorable expression profiles with low resistance potential, thereby selecting for targets where sustained clinical activity is more probable.⁵⁰ Incorporating resistance-informed logic into the earliest stages of target selection represents a conceptual advance beyond purely expression-based criteria and is expected to improve the durability of clinical responses.

Tumor heterogeneity

Tumor heterogeneity remains a major barrier to durable responses with targeted therapies, including ADCs, and even despite the bystander property present in some ADCs.^51,52 Intra-tumoral variation in antigen expression can facilitate clonal escape and drive therapeutic resistance.^51,53 AI offers a scalable solution to this challenge by leveraging large-scale transcriptomic or single-cell RNA sequencing (scRNA-seq) datasets to map tissue histology and antigen distribution at single-cell resolution across diverse patient cohorts. AI algorithms can assess whether target antigens are expressed homogeneously and robustly enough within tumors to support ADC development, while simultaneously identifying minor subpopulations with distinct expression profiles or molecular vulnerabilities.^54–57

For instance, Lareau et al.⁵⁸ proposed that single-cell genomics atlases should be leveraged to map antigen expression across all human cell types in a data-driven manner, enabling systematic de-risking of ADC targets prior to clinical testing; building on this, Nix et al.⁵⁹ demonstrated that scRNA-seq outperforms bulk RNA-seq and immunohistochemistry in resolving intra-tumoral antigen co-expression patterns, facilitating the identification of optimal TAA combinations for multi-targeting strategies such as bispecific ADCs.

These insights can inform patient stratification strategies, guide the development of companion diagnostics, prioritize targets with the highest likelihood of clinical benefit, and potentially select the best therapy based on the histologic image.^60,61 Moreover, AI-driven models, including deep learning architectures, could predict temporal shifts in antigen expression during disease progression or under therapeutic pressure, enabling the design of treatment strategies aimed at mitigating resistance. For instance, Cheng et al.⁶² developed PROFET, a deep learning framework that reconstructs continuous gene expression trajectories from longitudinal scRNA-seq data, enabling the prediction of cell state transitions and phenotypic shifts induced by therapeutic pressure—identifying surface marker changes in drug-resistant subpopulations that would otherwise be invisible to static transcriptomic snapshots. Complementing this computational approach, clinical evidence has already demonstrated the functional relevance of such predictions: HER2 antigen downregulation under T-DXd treatment has been shown to mediate acquired resistance, which could subsequently be overcome by switching to a TROP2-directed ADC sharing the same payload—illustrating how anticipating temporal antigen shifts can directly guide resistance-mitigating treatment sequencing strategies.⁶³ Finally, heterogeneity in the expression of tumor antigens may be identified by AI, and combination therapies designed to target different populations of cells expressing distinct sets of antigens can be suggested to be the basis for the development of bsADCs. Notwithstanding these advances, it is important to recognize that such approaches remain largely investigational and have yet to be systematically integrated into target identification pipelines.

Mapping optimal epitopes for antibody recognition

Rational antibody design heavily relies on identifying accessible epitopes.¹⁰ To this end, in silico tools such as AlphaFold can be employed to generate high-confidence structural models of the ectodomain of target proteins.⁶⁴ In addition, exposed regions can also be investigated and visualized at molecular level through PyMOL, ChimeraX, or Mol*, in conjunction with quantitative metrics like solvent-accessible surface area (SASA).^65–67 Candidate epitopes are further evaluated for structural stability using per-residue confidence scores (pLDDT) and for sequence conservation across tumor and normal tissues to ensure both binding specificity and clinical translatability.⁶⁴ These studies may help in the identification of antigen epitopes that can likely produce an antibody response. Yet, the tertiary structure of the antigen, critical in many cases in raising antibodies to the protein, may not be easily reproduced by short protein sequences. Moreover, exposure of a region on the surface of the antigen does not guarantee accessibility/antigenicity, which requires experimental lab confirmation. Of note although these methodologies are very widely integrated in the current approaches for antigen selection, a unique computational-specific framework has not been reported.

Antigen internalization

Another factor that may limit the efficacy of an ADC is the internalization of the antigen target, especially after binding to the ADC. Most of the ADCs raised depend on internalization of the ADC-antigen target for antitumoral efficacy, with some exceptions. One of them is represented by ADCs that present an acid-sensitive linker. These ADCs may release the payload in intracellular acidic compartments, such as lysosomes, but also in the extracellular medium, provided that the latter is sufficiently acidic, as it happens in poorly oxygenated tumoral regions.⁶⁸

Predictive models for antibody-induced internalization remain limited and are currently in the investigational phase. Experimental assays continue to represent the gold standard for assessing receptor internalization upon antibody engagement. However, emerging machine learning (ML) approaches trained in available literature offer a promising alternative. These models leverage features such as protein class, trafficking motifs, and endocytic behavior, integrating data from resources like UniProt, SurfaceomeDB, and The Cell Surface Protein Atlas, alongside structural modeling and molecular dynamics (MD) simulations to forecast internalization potential.^69–71 While internalization of antigens appears to be an important component of the action of ADCs, it is worth indicating that most membrane proteins internalize.⁷² Moreover, the clinical efficacy of ADCs in settings of low antigen expression, as is the case for trastuzumab-deruxtecan (T-DXd),⁵ conceptually challenges the value of antigen-mediated ADC internalization in the antitumoral effect of ADCs.

The mechanism underlying T-DXd efficacy in HER2-low and HER2-negative settings has now been characterized: it has been reported that antitumoral activity in these contexts is mediated primarily by extracellular proteases within the tumor microenvironment—particularly cathepsin L (CTSL)—that cleave the tetrapeptide linker extracellularly, releasing the membrane-permeable DXd payload for diffusion into surrounding tumor cells independently of receptor binding or internalization.⁷³This extracellular protease-dependent mechanism is computationally tractable: spatial proteomics and transcriptomic models can be used to map CTSL expression across the tumor microenvironment, and its integration into ADC distribution simulations provides a mechanistic basis for predicting bystander efficacy as a function of tumor protease landscape—a dimension of ADC pharmacology that static internalization models cannot capture.⁷³

Antibody internalization is a critical determinant of ADC efficacy but cannot be predicted by binding affinity alone, as it depends on dynamic antibody–receptor interactions that govern endocytosis. Computational approaches, including molecular dynamics simulations, have demonstrated that antibodies with moderate binding energies and flexible interaction profiles promote receptor clustering and efficient internalization, whereas excessively high-affinity binding may impair uptake by restricting the conformational rearrangements required for membrane trafficking.^11,55 In addition, multivalent engagement enhances receptor crosslinking, stabilizes complexes, and facilitates membrane deformation necessary for endocytosis.⁷⁵

A further and clinically critical downstream step is lysosomal payload export to the cytoplasm, which is mediated by specific lysosomal membrane transporters and is particularly important for ADCs bearing non-cleavable linkers. SLC46A3, a proton-coupled lysosomal membrane transporter, has been identified as essential for the cytoplasmic translocation of maytansinoid catabolites following lysosomal proteolysis of T-DM1; loss of SLC46A3 expression results in lysosomal payload sequestration and substantially diminished cytotoxicity, and has been validated as a mechanism of both innate and acquired resistance to non-cleavable ADCs across multiple tumor types and targets including HER2, EPHA2, and BCMA.^76,77 Consistent with this, reduced lysosomal proteolytic activity—through decreased cathepsin expression or impaired lysosomal acidification via V-ATPase downregulation—has been demonstrated as an independent mechanism of T-DM1 resistance, preventing the release of the active DM1 catabolite Lys-SMCC-DM1 from lysosomal proteolysis and thereby abolishing cytotoxic payload delivery to the cytoplasm.^78,79

These computational predictions have been experimentally validated, supporting their utility in guiding antibody design for improved cellular uptake and ADC performance. Importantly, optimal internalization must be considered in conjunction with linker and payload design: cleavable linkers should ensure plasma stability while enabling efficient intracellular release, and payload physicochemical properties, particularly hydrophobicity, must be carefully balanced to avoid non-specific antibody interactions, reduced antigen binding, and impaired payload accessibility.¹¹ Collectively, these findings highlight that efficient ADC function arises from the integration of dynamic antibody binding, receptor biology, and optimized linker–payload combinations that ensure controlled release, minimal steric hindrance, and sustained target engagement.

Neoantigen prediction

Beyond canonical surface proteins, AI can be used to predict neoantigens, i.e., peptides derived from tumor-specific mutations that are presented on major histocompatibility complex (MHC) molecules.⁸⁰ While neoantigens are primarily explored in the context of cancer vaccines and T-cell therapies, similar AI-driven strategies can help identify cancer-unique markers that guide ADC development to target cells exposing neoantigens.⁸¹ Certain models, such as NetMHCpan-4.1, MHC flurry, and ImmuneApp, now enable rapid prediction of peptide–MHC binding and immunogenicity, significantly accelerating the discovery of tumor-exclusive epitopes that could be co-opted for antibody-based or bispecific ADC strategies.^82–84 Recent reviews highlight how combining immunogenomic approaches (using NGS-derived peptidomes to predict HLA binding) with immunopeptidomic strategies (mass spectrometry validation of MHC-bound peptides) provides a powerful framework for translating computational predictions into therapeutic applications.⁸⁵ Importantly, this concept is already advancing clinically, as some companies have developed platforms that isolate and engineer TCRs against patient-specific neoantigens, directly operationalizing these predictive strategies into cell therapy pipelines.⁸⁶ While primarily focused on T cell therapeutics, such approaches illustrate how neoantigen-derived epitopes could also be leveraged to design next-generation ADCs or bispecific constructs, broadening the actionable space beyond classical surfaceome targets.

Table 3 provides a list of AI strategies for target identification, including the application area, the AI method, and the contribution to the ADC design.

Table 3.

List of AI strategies for surfaceome target identification in ADC development

AI Application Area	AI/ML Methods	Use Case/Description	Benefit for ADC Development
Surfaceome Prediction	Supervised learning, ensemble classifiers	Predict which genes/proteins encode surface-localized proteins based on sequence features	Prioritize membrane proteins not yet annotated as surface targets49
Multi-Omics Integration	Data fusion, neural networks	Integrate transcriptomic, proteomic, and glycomic data to define tumor-specific surfaceomes	Identify tumor-enriched surface targets with high specificity282,283
Target Prioritization Algorithms	Scoring models, decision trees	Rank candidate proteins based on expression, internalization, druggability, etc.	Select optimal ADC targets with favorable therapeutic index50
Protein Localization Prediction	Deep learning (e.g., DeepLoc, SignalP)	Predict subcellular localization from sequence and structural features	Filter for proteins reliably localized to the plasma membrane284,285
Single-Cell Data Mining	Unsupervised clustering, graph AI	Analyze scRNA-seq/Spatial data to identify tumor-specific surface protein expression	Capture tumor heterogeneity; avoid off-tumor toxicity286
Epitope Mapping	AI-guided structure prediction (AlphaFold, Rosetta)	Identify accessible extracellular epitopes for antibody binding	Design antibodies that effectively bind tumor-restricted domains64,287
Antibody-Antigen Binding Prediction	Molecular docking + ML	Predict antibody binding affinity and specificity to novel targets	Accelerate antibody generation and validation for surface antigens288,289
De Novo Target Discovery	Generative models (e.g., GANs, VAEs)	Generate synthetic candidates or suggest unannotated membrane proteins	Expand target space beyond known surfaceome203,290
Literature & Patent Mining	Natural language processing (NLP)	Automatically extract potential targets from biomedical texts and patents	Identify overlooked or emerging ADC targets from published sources291,292

AI-enhanced antibody design

Once a target antigen is selected, the next challenge is engineering an antibody that binds the target with high affinity and specificity, while also being stable and manufacturable (Fig. 2). An article describing approaches to improve specificity is reported elsewhere.¹⁰ In conventional antibody development, optimizing an antibody often requires iterative rounds of mutagenesis and experimental screening (e.g., panning display libraries), which sample only a tiny fraction of possible variants.¹⁰ AI-based methods trained on vast antibody sequence and structural datasets, including RESP, AntBO, and IgCraft, RFdiffusion can predict residue mutations that enhance binding affinity, specificity, stability, solubility, and low immunogenicity, and in some cases generate fully novel antibody designs, enabling optimization in a single modeling framework.^87–89 Below, we describe key approaches in antibody engineering.

Fig. 2.

AI-assisted pipeline for the generation of antibody–drug conjugates. A four-stage framework linking data to optimized ADC design. (1) Input data span genomics and transcriptomics (RNA-seq, single-cell omics), proteomics (single-cell proteomics, CITE-seq), spatial biology (imaging mass cytometry, CODEX), clinical data (patient cohorts, biomarkers), and biochemical and in vitro data (binding, cytotoxicity, internalization). (2) AI models applied to these data include deep learning architectures (graph and convolutional neural networks), generative models (variational autoencoders, generative adversarial networks), protein structure prediction (AlphaFold, ESMFold), classical machine learning (regression, classification), and cloud and high-performance computing infrastructure supporting scalable, distributed workflows. (3) AI-driven ADC design encompasses target identification (surfaceome analysis, tumor mapping, biomarker stratification), antibody engineering (stability, aggregation and immunogenicity prediction, affinity maturation), linker optimization (cleavage kinetics, drug-to-antibody ratio optimization, off-target minimization), payload engineering (bystander effect modulation, toxicity prediction, SAR-guided design), dual-target strategies (in silico epitope binding, multi-epitope engagement, synergy prediction), and mechanistic integration (molecular dynamics, molecular docking, structural modeling, PBPK and quantitative systems pharmacology analysis, in silico ADME). (4) Translational outcomes comprise improved efficacy through enhanced tumor selectivity and potency, reduced toxicity via off-target minimization, strategies to overcome resistance, patient stratification for precision oncology, and accelerated development timelines — converging on an optimized ADC

Affinity maturation and specificity optimization

AI-driven generative modeling is transforming the engineering of antibody complementarity-determining regions (CDRs) to enhance binding affinity and specificity. Generative architectures, including variational autoencoders (VAEs), generative adversarial networks (GANs), and transformer-based protein language models, are trained on large-scale datasets of antibody–antigen complexes to learn sequence features associated with high-affinity binding.¹⁹ These models can propose novel CDR variants predicted to improve antigen engagement while preserving structural compatibility and developability. Achieving high specificity is equally critical in antibody development, as off-target binding and polyspecificity remain key liabilities. Recent advances in machine learning have enabled the decoding of sequence determinants of specificity, providing strategies to reduce nonspecific interactions. For example, neural networks trained on heavy-chain variable region sequences successfully predicted polyspecificity.⁸⁷ Building on this, Makowski et al. developed classifiers trained on deep-sequenced CDR-mutated libraries of emibetuzumab, which generalized beyond the training set and identified mutations that simultaneously improved specificity and maintained affinity.^88,90

While the approaches above focus on optimizing existing CDR sequences, a conceptually distinct and more ambitious paradigm has recently emerged: the de novo generation of full antibody scaffolds targeting user-specified epitopes entirely from scratch, without reliance on a pre-existing binding scaffold. RFdiffusion, a diffusion-based deep learning model originally developed for general protein design, has been fine-tuned on thousands of antibody–antigen complexes to design novel CDR loops with inherent shape complementarity to a chosen epitope. Combining RFdiffusion backbone generation with ProteinMPNN sequence design and yeast display screening, Bennett et al. demonstrated the de novo generation of VHHs, scFvs, and full antibodies binding disease-relevant epitopes—including influenza hemagglutinin and C. difficile toxin B—with cryo-EM–confirmed atomic-level precision.⁸⁹ This work establishes that AI-guided design can, at least in selected cases, bypass the need for immunization or random library screening, representing a potentially transformative advance for ADC development where precise epitope targeting is a prerequisite. Active learning frameworks such as ALLM-Ab couple fine-tuned protein language models with multi-objective optimization to simultaneously improve binding affinity and developability, with performance benchmarked against deep mutational scanning datasets—establishing a scalable computational-experimental loop for iterative antibody engineering.⁹¹

State-of-the-art models created with AlphaFold2, Rosetta-based pipelines, and language model predictors like ESMFold, can now robustly model Fab domains and, in many cases, their complexes with the antigens.^36,64,90 These tools enable visualization of binding geometries, interface complementarity, and regions with low-confidence predictions (e.g., per-residue pLDDT scores), guiding targeted mutagenesis and sequence redesign. More recently, HelixFold-Multimer has outperformed prior frameworks, offering refined epitope definition and structural insights.⁹² In addition, Meta’s ESMFold enables rapid Fab structure prediction, providing confidence estimates suitable for large-scale structural screening.⁹⁰ Collectively, these strategies allow in silico exploration of CDR variants to refine affinity and specificity, effectively accelerating antibody design ahead of experimental testing.

An important advance not captured in prior frameworks is AlphaFold3 (AF3), which extends its predecessor by employing a diffusion-based generative architecture capable of modeling not only protein–protein interfaces but also complexes involving glycans, small molecules, and ions.⁷⁰ For affinity maturation and ADC-relevant antibody design, AF3 offers improved capacity to model glycosylated antigen epitopes, predict how post-translational modifications on the target influence antibody engagement, and assess structural compatibility of the antibody–linker–antigen interface. However, it is important to note that AF3 still struggles to capture glycan microheterogeneity and dynamic shielding effects, and its stochastic sampling behavior means that reliable performance for antibody–antigen docking typically requires multiple independent runs rather than a single prediction.⁹³ When sufficient sampling is performed, success rates for near-native complex prediction improve substantially, but this computational overhead remains a practical limitation for large-scale screening campaigns.

In the ADC context, the stakes of polyspecificity extend beyond general therapeutic safety: off-target binding to antigens expressed on normal tissues can result in direct cytotoxic payload delivery to healthy cells—a distinctly more serious liability than for unconjugated antibodies. Computational assessment of cross-reactivity with normal tissue antigens must therefore be integrated into affinity maturation workflows, not merely as a post-selection filter but as a simultaneous design objective. AI models trained on species-ortholog and paralog datasets can flag candidate CDR variants that, while improving tumor antigen affinity, inadvertently acquire binding to structurally related epitopes on non-target proteins.⁹⁴ Additionally, AI-guided image analysis can identify subtle structural motifs—including glycosylation patterns and receptor clustering domains—that distinguish the tumor-specific antigen conformation from its normal-tissue counterpart, providing a structural basis for specificity engineering at the interface level.⁹⁵ This dimension of specificity optimization—distinguishing tumor-expressed from normal-tissue-expressed protein conformations—is particularly relevant for antigens with low but non-zero normal tissue expression, which constitute the majority of current ADC targets.

Given the increasing prominence of bispecific ADCs (bsADCs) discussed in the target identification section, it is notable that computational tools for bispecific antibody engineering present unique challenges beyond those of monospecific optimization. Engineering two functional binding arms on a single scaffold requires simultaneous optimization of two distinct CDR sets, management of VH–VL pairing preferences across arms, and avoidance of interdomain steric conflicts that could impair either binding activity or manufacturability. A recent AI-driven framework for common light chain (cLC) engineering—a key requirement for producing well-behaved bispecific IgGs —successfully identified cLC designs for 7 out of 10 therapeutic targets tested, reducing the experimental screening burden from thousands of variants to only a handful per target and achieving >80% purity at biologically relevant expression levels.⁹⁶ These emerging computational strategies for bsADC engineering are still nascent compared to monospecific tools, but their rapid development signals that the computational-experimental gap for this important therapeutic format is closing.

However, despite their promise, all these frameworks share a fundamental limitation: the CDR-H3 loop—the primary determinant of antigen binding—remains the most structurally diverse and difficult region to model accurately, with single static predictions failing to capture biologically relevant conformational variability.⁹⁷ AlphaFold-Multimer is further constrained by the absence of co-evolutionary signals at antibody-antigen interfaces; benchmarking across 427 nonredundant antibody-antigen complexes showed that while the latest AlphaFold version improved near-native modeling success to over 30%, a previous version achieved only ~20%, and increased sampling was required to reach ~50% success—highlighting that reliable performance is not yet achieved under standard conditions.⁹⁸ Rosetta-based pipelines, while superior for local CDR refinement, are computationally expensive and similarly limited by CDR-H3 inaccuracy, and their small training datasets restrict generalization to novel complexes.⁹⁹ ESMFold trades accuracy for speed, correctly predicting only ~41% of dimeric structures and performing poorly on multi-chain systems—limiting its utility beyond rapid single-chain screening.¹⁰⁰ Critically, none of these models has been prospectively validated in an experimental antibody development workflow, as existing benchmarks reflect retrospective performance on known structures, leaving a significant gap between computational prediction and actionable experimental guidance.¹⁰¹

Emerging strategies aim to address the CDR-H3 sampling problem by treating this loop as a conformational ensemble rather than a single static structure—an approach rooted in physics-based MD and enhanced sampling methods discussed in detail in the Physics-Based Computational Frameworks section below. Briefly, metadynamics, gREST, and hybrid AlphaFlow/HADDOCK workflows have shown improved success rates for antibody–antigen complex modeling compared to single-prediction outputs,^100,102,103 complementing sequence-level generative models with the conformational resolution required for reliable structure-guided affinity engineering.

Developability and stability

Therapeutic viability requires antibodies with favorable manufacturing properties in addition to high affinity. Desired developability traits include solubility, thermal and conformational stability, low aggregation propensity, minimal immunogenicity, and robust expression in mammalian systems.¹⁰⁴ AI and ML models are increasingly applied to predict these parameters directly from sequence data, enabling early identification of liabilities and reducing attrition.⁸⁷

PROPERMAB (PROPERties of Monoclonal AntiBodies), developed by Regeneron Pharmaceuticals, is a versatile integrative computational framework for large-scale in silico developability prediction of mAbs, combining sequence-based features, structure-based descriptors derived from AlphaFold-predicted Fv models, and pretrained protein language model embeddings within a unified ML pipeline. Applied to HIC retention time—a proxy for hydrophobicity and aggregation risk—and high-concentration viscosity, PROPERMAB demonstrated that structure-derived features can be rapidly predicted directly from sequences by pre-training simple surrogate models, enabling scaling to repertoire-level datasets without experimental input¹⁰⁵

Raybould et al. established a developability framework for antibodies that is conceptually analogous to Lipinski’s Rule of Five for small molecules, yet adapted to the structural and biophysical complexity of biologics. While Lipinski’s rules define drug-likeness through simple physicochemical thresholds (e.g., molecular weight, lipophilicity, hydrogen bond donors/acceptors) to predict oral bioavailability, the TAP guidelines capture higher-order features such as CDR length, surface hydrophobicity, and charge distribution that govern stability, aggregation, and manufacturability in antibodies.^106,107 Both frameworks rely on empirical distributions derived from successful drugs and function as early-stage filters rather than strict rules. However, whereas Lipinski operates largely at the level of molecular composition, Raybould et al. introduce a structure-aware paradigm, reflecting the importance of spatial organization and surface properties in biologics. Together, they illustrate a unifying principle: successful therapeutics occupy a constrained physicochemical space, although the defining parameters differ fundamentally between small molecules and antibodies.^106,107 With this framework in mind, these heuristics have been integrated into ML pipelines for systematic library screening. In parallel, regressors have been developed to predict quantitative traits such as melting temperature and aggregation propensity, while classification models can forecast expression titers based on CDR and framework composition.¹⁰⁸ These predictive frameworks allow early engineering of stabilizing mutations and removal of destabilizing residues prior to laboratory production, accelerating optimization cycles and increasing the probability of success. For a comprehensive overview of AI-driven methods addressing developability, viscosity, aggregation, and solubility, the reader is referred to Varun Dewaker et al.¹⁰⁹

A critical and often under-addressed developability challenge is immunogenicity—the risk of inducing anti-drug antibody (ADA) responses that reduce efficacy or cause adverse events. T-SCAPE, a deep learning framework integrating multidomain pretraining across diverse biological information, achieves state-of-the-art prediction of T-cell epitope immunogenicity for specific pMHC pairs and delivers top-tier performance in assessing the ADA-inducing potential of therapeutic antibodies without requiring MHC allele inputs as a prerequisite, representing a practical advance for antibody developability screening.¹¹⁰

Despite notable progress in AI-based developability prediction, systematic benchmarking across multiple protein models shows that no single tool consistently outperforms others across all key traits—including expression, thermostability, immunogenicity, aggregation, and polyreactivity. Consequently, the field still lacks a universally validated integrated prediction framework, and experimental characterization remains essential for lead candidate selection.¹¹¹

Table 4 compares Lipinski’s Rules for Chemical Entities with Raybould’s Antibody Developability Guidelines.

Table 4.

Comparison: Lipinski’s rules vs Raybould’s antibody developability guidelines

Feature	Lipinski’s Rule of Five (Small Molecules)	Raybould et al.’s Developability Guidelines (Antibodies)
Purpose	Predict oral bioavailability and drug-likeness	Predict developability of therapeutic antibodies (e.g., aggregation, viscosity, manufacturability)
Molecule type	Small molecules	Antibodies (biologics)
Main parameters	- Molecular weight <500 Da - LogP <5 - H-bond donors ≤5 - H-bond acceptors ≤10	- Total CDR length - Hydrophobic surface patch (PSH) - Positive charge patch (PPC) - Negative charge patch (PNC) - Fv charge symmetry parameter (FvCSP)
Source of features	Molecular formula and structure	Antibody sequence and 3D structure
Focus	Permeability, solubility, and bioavailability	Aggregation, viscosity, immunogenicity, and expression/stability
Applicability domain	Small molecule drug screening	Antibody engineering and selection
Interpretation	Fewer violations = higher chance of success as oral drug	Outlier features = higher risk of poor developability
Computational tools	Cheminformatics tools (e.g., RDKit)	Therapeutic Antibody Profiler (TAP)
Limitation	Does not apply to biologics	Requires structure modeling or experimental data

This table contrasts druglikeness criteria for small molecules with developability principles for therapeutic antibodies. Lipinski’s rules rely on simple physicochemical properties to predict oral bioavailability, whereas Raybould et al. use structure- and sequence-based features such as CDR length, hydrophobicity, and charge distribution to assess aggregation, stability, and manufacturability. Both frameworks are empirically derived and function as early-stage filters rather than strict rules. However, they differ in complexity, reflecting the distinct biophysical constraints of small molecules versus biologics

Linker design

The linker is a central determinant of ADC performance, as it bridges the antibody to its cytotoxic payload. Optimal linkers must remain stable during systemic circulation to prevent premature toxin release, yet be selectively cleavable within the tumor microenvironment to ensure efficient payload delivery at the site of disease^13,112 (Fig. 2). Achieving this balance requires precise control over chemical and biological properties. In silico approaches can predict linker stability, hydrolysis rates, and cleavage kinetics, while simultaneously identifying optimal conjugation sites to maximize payload delivery and minimize off-target effects, aggregation, and immunogenicity.^11,83

Linkers are broadly categorized as cleavable or non-cleavable. Cleavable linkers, such as acid-labile hydrazones, protease-sensitive dipeptides, or disulfides reducible by intracellular glutathione, are engineered to release their payload in tumors. Non-cleavable linkers, in contrast, rely on lysosomal degradation of the antibody to liberate the active drug.^112,113 In the next section, we explore computational and AI approaches for linker optimization and payload release in ADC development.

AI-enabled conjugation site engineering

The site and method of conjugation exert a critical influence on ADC stability, pharmacokinetics, and therapeutic efficacy. Traditional conjugation approaches, such as stochastic modification of lysine or interchain cysteine residues, often result in heterogeneous drug-to-antibody ratios (DARs) and variable functional performance. Site-specific conjugation strategies, including engineered cysteine insertion (THIOMAB technology as an example), noncanonical amino acid incorporation, and glycan remodeling, enable tighter control of DAR while preserving antibody structure and antigen affinity.^114–116 The precise location of drug attachment can significantly impact in vivo behavior, influencing clearance, immunogenicity, and tumor penetration. Some studies have demonstrated that the position of engineered cysteine residues on antibodies critically determines the stability of ADCs. Data suggest that maleimide linkages are generally more stable than disulfide ones and that stable conjugation sites exist across both heavy and light chains.¹¹⁷ Additionally, in vitro stability translates to in vivo performance and to be transferable across different payloads and antibodies.¹¹⁷ Complementing this, hiding payload inside the IgG Fc cavity significantly enhances the therapeutic index of ADCs, which reveals that structural positioning of the payload strongly impacts ADC efficacy and safety.¹¹⁸ The authors found that hydrophilic linkers are essential for in vivo activity and that conjugation to terminal sialic acid yields greater stability and potency than galactose.¹¹⁸ Together, these studies establish that both precise conjugation site selection and strategic payload positioning are key determinants of ADC stability, efficacy, and overall druggability.^117,118

Glycan remodeling at the conserved Fc Asn297 N-glycosylation site represents a particularly attractive conjugation strategy because it exploits a defined, structurally remote position that minimally perturbs antigen binding while yielding highly homogeneous ADCs. Chemoenzymatic glycan remodeling approaches—employing endoglycosidases such as EndoS2 to strip and rebuild the Fc glycan with azido- or alkyne-functionalized sugar donors—enable strain-promoted cycloaddition-based site-specific conjugation applicable across all IgG subclasses and multiple linker-payload combinations.¹¹⁹ A recent biparatopic HER2-targeting ADC assembled via this glycan conjugation strategy demonstrated superior in vitro stability, in vivo safety, and antitumor efficacy compared with conventional cysteine-conjugated counterparts, directly attributing the improved therapeutic index to the homogeneity and remote positioning of the Fc glycoconjugation site.¹²⁰

Structure-prediction platforms such as AlphaFold2 and Rosetta-based ML frameworks can approximate antibody tertiary structure and surface accessibility, guiding the selection of optimal sites for engineered cysteine insertion.^64,121 For unnatural amino acid incorporation (pAcetylphenylalanine as an example), deep generative models such as ProteinMPNN and Bayesian optimization algorithms predict structurally permissive sites for site-specific labeling. These methods account for codon usage, solvent accessibility, and conformational constraints, balancing chemical functionality with structural integrity and low immunogenicity.^122,123

Drug-to-antibody ratio optimization

DAR, defined as the average number of cytotoxic payloads conjugated per antibody, is a critical determinant of ADC efficacy, safety, and pharmacokinetics. Most approved ADCs have DARs in the range of 2–4, reflecting a compromise between maximizing tumor cell kill and minimizing systemic toxicity.¹³ While increasing DAR can enhance potency through higher payload delivery per antibody molecule, elevated DARs often result in reduced stability, accelerated plasma clearance, and increased risk of off-target toxicity due to premature payload release or aggregation.¹²⁴ Thus, the optimization of DAR remains a central objective in ADC engineering.

The biophysical basis of DAR-dependent clearance is now well established: hydrophobic payloads at high DAR values cause ADCs to adopt surfactant-like amphiphilic structures that self-aggregate in vivo, driving hepatic uptake and accelerated catabolism in a hydrophobicity-dependent manner.¹²⁵ This relationship between DAR, aggregate propensity, and plasma clearance is not linear and is strongly modulated by payload and linker hydrophobicity— meaning that the same DAR value can yield acceptable PK with a hydrophilic payload or linker but unacceptable clearance with a hydrophobic one. PEGylation of the linker is the most clinically validated strategy to decouple DAR from hydrophobicity: incorporating methyl-PEG24 side chains into dipeptide linkers has been shown to enable the production of DAR 8 ADCs with MMAE—a payload not generally considered compatible with high DAR—that retain full biophysical stability, prolonged plasma half-life, and superior animal tolerability compared to conventional DAR 4 constructs with the same payload.¹²⁶ Two approved ADCs (sacituzumab govitecan and loncastuximab tesirine) already incorporate PEGylated linker architectures to enable higher DAR loading without the clearance penalties typical of non-PEGylated high-DAR constructs, establishing a clinical precedent for this design principle.¹²⁵

Computational approaches are now being deployed to identify the optimal DAR for specific antibody–linker–payload combinations.

Recent data suggest the important role that biophysics can have in the evaluation of DAR. Particularly, molecular dynamics and SILCS, enable systematic evaluation of multiple site/DAR combinations by predicting structural flexibility, interaction patterns, and payload accessibility. For example, simulations show that different conjugation sites and DARs significantly alter antibody conformation and payload exposure, influencing drug release and efficacy.¹²⁷

At the systems pharmacology level, quantitative systems pharmacology (QSP) models are increasingly being applied to define the optimal DAR range for specific ADC-antigen-tumor combinations by integrating mechanistic descriptions of ADC binding, internalization, intracellular trafficking, and payload release with PK predictions across a range of DAR values. A recently published QSP model of a pyrrolobenzodiazepine-based ADC targeting BCMA (MEDI2228) integrated in vitro mechanistic data with sensitivity analysis to quantify how DAR-dependent payload loading translates to intracellular drug concentrations and cytotoxic outcomes in myeloma cells—providing a framework for in silico DAR optimization that could be applied prospectively to novel ADC programs before in vivo testing.¹²⁸ Complementing this, a PBPK model for a CLDN18.2-targeting ADC, developed and qualified with clinical PK data using the open-source PK-Sim platform, demonstrated that target-mediated drug disposition (TMDD) and deconjugation rate parameters—both of which are DAR-dependent — critically determine the systemic exposure of both intact ADC and free payload, establishing PBPK modeling as a practical tool for DAR-dependent dose prediction ahead of clinical trials.¹²⁹ ML approaches are now also being applied directly to DAR prediction.

Payload optimization: AI-guided design of cytotoxic warheads

Payload release: a determinant of ADC efficacy and safety

The therapeutic efficacy and safety of ADCs critically depend on payload release. Upon antigen binding, ADCs are internalized via receptor-mediated endocytosis and routed to lysosomes, where proteolytic degradation of the antibody and cleavage of the linker liberate the cytotoxic payload.¹³⁰ Alternatively, release of the payload in acidic extracellular environments is possible in the case of acid-sensitive linkers. Precise temporal and spatial control of these processes is essential to ensure on-target cytotoxicity while minimizing off-tumor effects.^13,131

The kinetics of intracellular release are governed by a complex interplay between linker chemistry, subcellular trafficking dynamics, and payload properties, each influencing the extent and timing of drug activation.^132,133 Leveraging computational-based analysis of endosomal pH gradients and lysosomal protease activity, together with rational linker design, could enable the creation of pH-sensitive ADCs that selectively release their payload upon endocytosis, minimizing off-target toxicity and enhancing tumor-specific cytotoxicity.^131,132 Recently, models capable of designing optimal linkers have been reported. For example, Linker-GPT is a Transformer-based deep learning model that uses self-attention, transfer learning, and reinforcement learning to generate novel, synthetically accessible ADC linkers.¹³⁴

Payload optimization: tailoring potency and pharmacochemistry in ADC Design

The payload, i.e., the cytotoxic moiety delivered by an ADC, serves as a lethal effector once internalized by cancer cells. Given that only a fraction of the administered dose reaches the tumor site, payloads must either be extraordinarily potent at low concentrations or possess excellent physicochemical properties that are compatible with conjugation, systemic circulation, and controlled cellular release^1,13,135 (Fig. 2). Moreover, payload characteristics such as hydrophobicity, membrane permeability, and efflux potential are key determinants of both efficacy and toxicity, including the capacity to induce a bystander effect.¹

Classic ADC payloads include natural product-derived microtubule inhibitors like auristatins and maytansinoids, and DNA-damaging agents, such as calicheamicin or duocarmycins.¹³⁶ These agents disrupt critical cellular processes but must be finely tuned to avoid aggregation and off-target release.^4,7

Computational tools and AI now support payload development in transformative ways. In silico structure-activity relationship (SAR) models assess how chemical modifications impact potency, cell permeability, and resistance mechanisms such as efflux transporter activity.¹³⁷ Moreover, simulations of payload release kinetics and predicted bystander cytotoxicity inform rational design of linker–payload combinations, together with tumor characteristics. Thus, tumors with an expected high degree of heterogeneity in the expression of the ADCs antigen should be targeted with ADCs capable of killing all tumoral cells with independence of the expression of the ADC target.

Emerging AI platforms also facilitate the exploration of novel payload classes with unique mechanisms, such as immune-stimulatory or targeted protein degradation agents (PROTACs), thus broadening the therapeutic arsenal of ADCs.^1,138 An example is the identification of CDG0501, a potent and conjugable GSPT1 degrader, using the Rosetta For Molecule Glue (RFMG) model combined with a CRBN-based degrader library.¹³⁹ This computational approach enabled the selection of a degrader suitable for antibody conjugation, supporting its use as a next-generation ADC payload.

Core scaffold optimization

Each payload class contains a core scaffold responsible for its cytotoxic mechanism. Monomethyl auristatin E (MMAE) and F (MMAF), derived from dolastatin, inhibit tubulin polymerization, while SN-38 and deruxtecan (DXd) inhibit topoisomerase I.^1,13 In one example, minor structural alterations such as methyl or carbamate additions to auristatins, enhanced membrane permeability and bystander effect.¹⁴⁰ Likewise, stabilizing the lactone ring of camptothecin analogs improved potency and chemical durability.¹⁴¹ AI could then help in creating analogs with group substitutions/additions, which can improve a chemical core to generate more sophisticated payloads.

Emerging payload modalities are expanding ADCs beyond cytotoxicity toward immune modulation and targeted protein degradation. Immune-stimulating antibody conjugates (ISACs) deliver TLR or STING agonists to activate tumor-localized immunity, requiring stable linkers and controlled payload properties to limit systemic toxicity.¹³⁸ In parallel, PROTAC-ADCs enable intracellular protein degradation through ubiquitin–proteasome pathways after internalization.¹⁴² These constructs introduce challenges related to payload size, polarity, and linker design, often favoring cleavable strategies. Overall, optimizing linker stability, payload chemistry, and antibody properties is critical to maximize efficacy and safety.

Linker attachment chemistry

Payloads must possess or be engineered to include functional groups for linker conjugation, commonly primary amines, carboxylates, or phenolic hydroxyls. Modifying these groups without compromising biological activity is a key challenge.¹⁴³ For example, derivatizing the C20-hydroxyl of SN-38 to a carbamate improves plasma stability while retaining intracellular efficacy.^1,144 Classification models further assist in identifying “linker-ready” variants, as exemplified by the exatecan derivative DXd, optimized for stable conjugation and potent cytotoxicity.¹

Hydrophobicity and bystander activity

Payload polarity directly influences cellular diffusion and the bystander killing effect, a critical consideration in heterogeneous tumors as this has been considered as a potential mechanism of resistance.³³ Hydrophobic payloads like MMAE can diffuse into neighboring antigen-negative cells, enhancing efficacy but increasing risk of off-tumor toxicity. In contrast, hydrophilic payloads like MMAF typically remain cell-confined.^8,13 The linker connecting the drug to the antibody increases molecular size, limiting diffusion to neighboring cells.¹⁴⁵ AI platforms trained on retention time and cytotoxic diffusion profiles can suggest modifications to tune permeability.⁸ For example, polar substitutions may reduce off-target spread, while masking hydrophilic groups can improve transmembrane transit.⁸ For instance, in silico computer simulations have been used to model how payload ionization state and linker architecture influence membrane permeability and bystander killing capacity, demonstrating that charged molecules encounter resistance at the hydrophobic membrane core while linker addition further limits diffusion by increasing molecular size—providing a rational computational basis for tuning polarity to control bystander spread.¹⁴⁵ Complementing this, Guo et al.¹⁴⁶ employed graph attention networks with comprehensive molecular characterizations to develop a bystander scoring (B score) model, enabling the rational, data-driven identification of ADC payloads with optimized bystander killing potential for use in antigen-heterogeneous tumors.¹⁴⁶

Dual payloads

Dual-payload ADCs, antibody–drug conjugates that carry two distinct cytotoxic agents on a single antibody backbone, are gaining attraction in oncology for their potential to target cancer through complementary mechanisms and overcome resistance.^147–149 Early evidence demonstrates that dual-payload ADCs achieve synergistic efficacy, outperforming co-administration of single-payload variants in heterogeneous tumor models by enhancing tumor penetrance and overcoming pathway-specific resistance. Use of ML trained on datasets of preclinical activity for payloads could predict which combination could have more synergistic potential in a particular tumor type. Preclinical efforts, combining payloads with different mechanisms of action, highlight the feasibility of this strategy to augment efficacy and overcome resistance.^147–149

The field has now moved decisively from concept to clinical reality. In 2025, the first dual-payload ADCs entered human trials: KH815 (Chengdu Kanghong Biotech), a TROP2-directed ADC combining a topoisomerase I inhibitor and an RNA polymerase II inhibitor at a DAR of 7.5, entered Phase I in patients with advanced solid tumors (NCT06885645).¹⁵⁰ Shortly thereafter, IBI3020 (Innovent Biologics), a CEACAM5-directed dual-payload ADC developed on the proprietary DuetTx® platform, completed first-in-human dosing (NCT06946446).¹⁵¹

Beyond combining classical cytotoxins, emerging dual-payload strategies are exploring orthogonal combinations that integrate direct tumor killing with immune modulation: pairing cytotoxic agents with toll-like receptor (TLR) agonists, STING agonists, or other immunostimulatory payload. These immunostimulatory ADC hybrids represent a convergence of the ADC and ISAC paradigms and could be particularly impactful in immunologically cold tumors.¹⁵²

Computational and AI-driven approaches are positioned to be particularly valuable in this space. ML models trained on datasets of preclinical payload activity, multi-omics tumor profiles, and drug synergy databases— incorporating graph network approaches and biomolecular interaction networks—can predict which payload combinations are likely to be synergistic in specific tumor contexts. Graph neural network architectures such as DumplingGNN could in principle be extended to score payload pairs^153,154 Despite the excitement, it must be acknowledged that clinical validation of dual-payload ADCs remains entirely absent: KH815 and IBI3020 are the only programs with human data, and results addressing efficacy, tolerability of combined payloads, and the mechanistic rationale for synergy in patients are awaited.

Physics-based computational frameworks for rational ADC design

MD simulations have been shown to provide submolecular resolution insights into monoclonal antibody (mAb) structure and dynamics, capturing conformational plasticity and stability of specific domains, antigen-binding orientation, or aggregation propensities under near-physiological conditions.^127,155,156 In the ADC context, MD simulations enable the quantitative dissection, spanning from nanosecond to millisecond trajectories depending on the molecular-level resolution, of how linker chemistry, conjugation site positioning, and DAR may influence payload accessibility, interference, and antibody-antigen engagement.^157,158 Accurate biophysical modeling at atomistic resolution requires empirically validated force fields such as CHARMM36m,¹⁵⁹ OPLS-AA¹⁶⁰ and AMBER ff19SB¹⁶¹ for modeling the protein sequences (including explicit water and ions^162–164), and CGenFF,¹⁶⁵ OPLS^166,167 and GAFF¹⁶⁸ their corresponding force fields, respectively, for simulating the payloads and linkers. For the Fc glycan simulations,¹⁵⁵ Amber-based force fields usually employ GLYCAM06,¹⁶⁹ OPLS simulations use its recent reparameterization to carbohydrates and other functional groups^166,170, whereas CHARMM36m models glycans through the specific CHARMM Carbohydrate force field.¹⁷¹ Importantly, despite the high computational cost of these simulations, all-atom MD excels in providing realistic conformational ensembles of protein domains and their underlying molecular interactions^{127,161,165,172–175} rather than static coordinates to capture the intrinsic flexibility of both mAbs and conjugated warheads.

While all-atom MD provides atomistic resolution detail, coarse-grained (CG) approaches, approximating different groups of atoms as single interaction sites,^176–179 have been used to probe antibody and ADC-relevant phase behavior on longer timescales and larger system sizes, generally inaccessible through atomistic simulations. CG simulations such as the 12-bead model¹⁸⁰ have been used to investigate intermolecular mAb self-association and stability under formulation-relevant conditions, providing mechanistic insights into aggregation propensities and viscosity changes at high concentrations.¹⁸¹ In the ADC context, CG models enable exploration of how conjugation-induced modifications in antibody surface properties affect colloidal stability and higher-order self-assembly, complementing atomistic predictions of linker–payload behavior.¹²⁷ In that sense, accurate high-resolution CG force fields such as MARTINI-based frameworks^182,183 have been recently used to characterize protein−excipient interactions of arginine and glutamate excipients, using the Fab domains of the therapeutic mAbs trastuzumab and omalizumab as model systems.¹⁸⁴ By bridging the resolution gap between biomolecular individual interactions and larger macromolecule ensembles,¹⁸⁵ CG simulations allow systematic mapping of formulation effects,¹⁸⁶ multivalent interactions,^187,188 and mesoscale phase behavior,¹⁸⁹ thereby offering a scalable route to de-risk ADC design prior to experimental validation, complementing atomistic predictions at significantly larger sizescales.¹⁰ Furthermore, for simulating intrinsically disordered regions (IDRs), even more coarse-grained protein force fields, such as those of residue-resolution level, including the Mpipi,¹⁹⁰ HPS-Urry, CALVADOS2, or the Mpipi-Recharged, among others, might be applied to systematically explore the conformational ensemble and intermolecular interactions established by protein IDRs in conjunction with structured globular domains.^177,178,191

Computational hybrid methods, such as the Site-Identification by Ligand Competitive Saturation (SILCS) approach, have been proposed to integrate grand canonical Monte Carlo with classical MD simulations to generate pre-computed grid free energy maps (“FragMaps”) of functional group binding affinities across the mAb surface.¹⁹² Remarkably, given the computational challenges of MD simulations of full mAbs for each ADC being considered, these maps enable rapid, yet quantitative evaluation of linker and payload binding landscapes without rerunning full-system MD simulations for each new ADC variant. In its biological adaptation, SILCS-Biologics can also guide the prediction of excipient-driven viscosity changes, map protein–protein interaction hotspots, and quantify effective protein charges across different formulation conditions.¹⁹³ As SILCS is based on pre-computed FragMaps from SILCS simulations, calculations on different ADC variants with different linker types and conjugation sites can be rapidly performed.¹²⁷ As an example, SILCS FragMaps can be calculated individually on the Fab and Fc, with a given warhead, allowing different possible binding sites and binding affinities of the warhead to be predicted. Moreover, SILCS can also be used to determine whether the linker or payload may influence interactions of the mAb with its antigen based on the payload-specific physicochemical binding to the CDRs. Conjugation site selection can also be refined through solvent accessibility profiling,¹⁹⁴ structure-based docking, and MD-driven identification of dynamically exposed reactive residues,^195,196 with pK_a prediction informing nucleophile reactivity.^197,198 The content in the SILCS FragMaps can also be used to guide excipient selection for facilitating formulation development.^199,200 MD-based payload–mAb and payload–payload interaction analyses can also quantify how hydrophobicity, DAR, or conjugation geometry modulate antigen affinity and warhead exposure, in which metrics such as root mean square deviation (RMSD) and SASA analyses can be highly informative for ultimate linker-release optimization.^8,127

In particular, feature attribution methods, such as SHAP (SHapley Additive exPlanations) and LIME^201,202 can be employed to quantify the contribution of molecular descriptors (e.g., hydrophobicity, charge distribution patterning, DAR or linker properties) to understand RFdiffusion predictions,^89,203, thereby providing insight into the physicochemical drivers and molecular interactions governing ADC phase behavior. In addition, combining deep learning methods, such as ColabFold for protein structural prediction, with physics-based simulation approaches (e.g., molecular dynamics simulations) offers a hybrid strategy to enhance interpretability by grounding predictions in established biophysical principles. Notably, their results suggest that excessively strong binding affinity of the fragment antigen-binding (Fab) domain may hinder internalization, as highly stable interactions with receptors—particularly within tight junction environments—can kinetically trap antibodies in the extracellular matrix.

Specific examples include the study by Corrada and Colombo,²⁰⁴ in which atomistic MD simulations are employed to investigate the energetic and dynamic mechanisms underlying antibody affinity maturation, specifically in variants of the bevacizumab antibody. This approach focuses on how mutations acquired during maturation influence not only the binding interface but also the internal flexibility and conformational dynamics of the antibody, ultimately enhancing antigen affinity. Moreover, Prass et al.²⁰⁵ use atomistic simulations to predict the viscosity of highly concentrated antibody solutions, a critical parameter for pharmaceutical formulation. Their work analyses protein–protein interactions and identifies surface regions that promote strong intermolecular associations, which are directly linked to increased viscosity. Furthermore, Llombart et al.⁷⁴ map the binding energy landscape of different antibody clones with the membrane protein JAM-A to provide computational predictions of antibody internalization. They find that internalizing mAbs exhibit a unique membrane-oriented contact topology that promotes cooperative receptor–receptor interactions, lowering the energetic barrier for early endocytic events. Their results establish molecular dynamics–guided clonal selection as a predictive framework for optimizing internalizing therapeutic antibodies and provide mechanistic insight into how antibody binding reshapes membrane-proximal receptor energetics to drive endocytosis. Altogether, these studies demonstrate how all-atom molecular simulations can be applied to both functional optimization and developability assessment of therapeutic antibodies and ADCs. Moreover, simulations using these force fields can provide the rational basis for selecting and optimize conjugation sites, linker architectures, and drug–antibody ratios (DARs) that balance systemic stability with effective intracellular drug release, an issue which has been shown experimentally to strongly influence their therapeutic index.^116,125

Taken together, these complementary physics-based platforms employing MD simulations can be applied to novel ADC architecture without dependence on prior training data, such as ML-based approaches (whose accuracy heavily relies on the available amount of training data), offering mechanistic and molecular resolution rationale that can substantially compress design timelines and de-risk translation (Fig. 3) Nevertheless, the limitations of computational physics-based approaches such as multiscale MD simulations remain in: (1) their high computational cost for large mAbs; (2) force-field accuracy; and (3) simulation convergence, which often require advanced sampling methods.²⁰⁶ However, emerging solutions, including enhanced-sampling MD methods,²⁰⁷ highly efficient GPU-parallelized pipelines,²⁰⁸ and machine-learning integration approaches (e.g., graph neural networks leveraging physics-derived structural features) aim to accelerate discovery-to-clinic transition.

Fig. 3.

Physics-based computational frameworks for rational ADC design. Three complementary molecular simulation approaches. (1) All-atom molecular dynamics (MD) resolves structural dynamics, conformational plasticity, antigen engagement, linker chemistry, conjugation site behavior and drug-to-antibody ratio (DAR) effects at atomic resolution, using biomolecular force fields (CHARMM36m, AMBER ff19SB, OPLS-AA, GLYCAM06). (2) Coarse-grained MD trades atomic detail for accessible length and time scales, enabling assessment of aggregation propensity, viscosity, surface properties, protein–membrane interactions, multivalency and intrinsically disordered regions (IDRs). (3) Hybrid MD/Monte Carlo approaches, exemplified by SILCS and SILCS-Biologics, map binding landscapes, excipient effects, conjugation site selection and solvent accessibility, and quantify the impact of warhead and linker attachment on antigen binding. Together these methods support accelerated translation from discovery to the clinic

Efficacy and safety predictions: AI in ADC pharmacology

Before an ADC enters clinical trials, developers must gain confidence that it will be effective at shrinking tumors and safe for patients at therapeutic doses. Traditionally, this is assessed through extensive in vitro and in vivo studies, followed by human trials.^19,209 These experiments are time-consuming, expensive, and sometimes, except for clinical trials, poorly predictive of human outcomes¹⁵⁹. Integrating quantitative systems pharmacology with machine learning algorithms trained on large-scale clinical and preclinical ADC datasets, these models could capture pharmacokinetic behavior, predict tumor–payload interactions, and anticipate adverse event spectra (Fig. 4). In the following section, we describe in detail some of the areas where computational approaches can be implemented to optimize the process.

Fig. 4.

Multi-scale AI-supported framework for predicting ADC efficacy and safety. Four linked modeling stages spanning systemic to individual-patient scales. (1) Physiologically based pharmacokinetic (PBPK) modeling simulates absorption, distribution, metabolism and excretion across parameterized compartments using physiological blood flow and tissue volume data, informing tissue exposure, clearance and dose optimization. (2) Tumor penetration and distribution modeling captures extravasation, interstitial diffusion and target binding, resolving the spatial gradient from high to low ADC penetration within the tumor as a function of tissue-specific parameters, binding kinetics and ADC clearance. (3) Toxicity prediction integrates antigen expression in healthy tissues, antigen–ADC complex internalization dynamics and FcRn-mediated recycling kinetics to anticipate on-target/off-tumor toxicity and estimate the therapeutic index alongside predicted tumor response. (4) Digital twins extend these outputs to virtual patient cohorts that capture inter-individual variability in physiological and tumor-specific parameters, supporting individualized dosing, response prediction and patient selection for clinical trials

Physiologically based pharmacokinetic (PBPK) modeling

Physiologically based pharmacokinetic (PBPK) models are mechanistic mathematical models that simulate drug absorption, distribution, metabolism, and excretion (ADME) via compartments corresponding to defined organs or tissues, parameterized using physiological blood flow and tissue volume data.²¹⁰

ADCs pose additional complexity because they are large proteins that follow antibody kinetics (with features like target-mediated drug disposition when they bind antigen, possible internalization, etc.) and also small-molecule kinetics for the released payload.^11,211 AI helps by calibrating and refining these PBPK models. Neural network approaches and parameter-fitting algorithms can leverage sparse experimental pharmacokinetic (PK) data, such as from rodent models or in vitro systems, to predict human PK profiles. Hybrid AI–mechanistic strategies integrate machine learning into PBPK models to fill gaps with hard-to-measure parameters like tumor permeability or payload clearance.²¹²

For instance, Lu et al. developed a Neural-ODE PK model trained on sparse clinical data from trastuzumab emtansine (T-DM1), outperforming traditional population PK models in predicting time-series concentration profiles.²¹² Hybrid mPBPK–ML frameworks integrating decision tree algorithms have further enabled prediction of ADC target occupancy from molecular features such as isoelectric point and binding affinity, with neural networks trained on over 300 ADC constructs demonstrating strong performance in predicting clearance (CL) and volume of distribution (Vd).⁹⁵ At the translational level, a PBPK model for MMAE-based ADCs successfully predicted whole-body pharmacokinetics across species, supporting cross-species extrapolation and human dose prediction from preclinical data.²¹³

Tumor penetration and distribution

While the ADC is parenterally injected into the blood, it must penetrate the tumor mass and bind to cancer cells throughout the tumor. Tumor tissues can be difficult to penetrate due to high interstitial pressure, heterogeneous blood supply, and the so-called “binding site barrier” where antibodies get trapped by antigens on cells near blood vessels, preventing deeper diffusion.¹⁹ It has been suggested that high-affinity antibodies can become sequestered near tumor entry points, limiting uniform distribution throughout the tissue.²¹⁴ However, this effect is difficult to explain due to the fact that plasma concentrations of the ADCs are much higher than the saturating doses. Therefore, even if the ADCs are captured by cells expressing high levels of the antigen, an excess of the ADC should be available to interact with other cells.

The Krogh cylinder model provides the principal mathematical framework for quantifying these competing forces. In this geometry, each tumor capillary is approximated as a cylinder surrounded by a radially symmetric tissue shell; the ADC extravasates through the capillary wall, diffuses radially outward, and is simultaneously consumed by binding and internalization at each spatial point across the tumor cross-section.²¹⁵ Weddell et al. extended this framework to a clinical setting by coupling a minimal PBPK model with spatially resolved tumor growth inhibition submodules, validating their predictions against observed response rates in T-DM1–treated patients with metastatic breast cancer—demonstrating that tumor penetration is mechanistically linked to clinical efficacy in a quantifiable manner.²¹⁶ Critically, simulations using this model confirm that elevated antigen expression density can paradoxically limit ADC penetration depth by accelerating binding-mediated consumption near the vasculature, whereas at very high plasma concentrations the binding-site barrier is relieved as surface receptors become saturated and excess drug diffuses further into the tissue.^217–220 This has informed the rational design of “carrier dose” strategies—the co-administration of unconjugated antibody to pre-saturate perivascular antigen and improve intratumorally homogeneous ADC distribution—which have been validated in agent-based simulation and preclinical models.^221,222 See Fig. 5.

Fig. 5.

ADC tumor penetration, cancer cell heterogeneity, and a spatially informed computational framework. a Physical and biological barriers to ADC delivery within the heterogeneous tumor microenvironment. Following transvascular transport across a discontinuous basement membrane, ADC distribution is constrained by the binding-site barrier at high-antigen cells adjacent to the vasculature, producing a perivascular penetration gradient; further limitations arise from the extracellular matrix and fibroblast-rich stroma, the necrotic core, and heterogeneous target expression across high-, medium-, low-, and antigen-negative tumor cells interspersed with immune infiltrates. b Computational framework in which digital pathology and spatial multiomics (Path2Space) generate antigen maps, immune infiltrate distributions, and resistant niche annotations from H&E slides. These outputs parameterize two complementary mechanistic models—a spatially resolved PBPK model of ADC transport from the vasculature into tumor tissue, and a cell-resolved (agent-based) model of the tumor microenvironment—whose outputs train a machine-learning surrogate enabling rapid screening of drug-to-antibody ratio (DAR) and binding affinity. Predicted outcomes include tumor penetration, target coverage, and ADC design optimization

Computational-based simulation platforms, incorporating agent-based modeling or machine learning surrogates of partial differential equations, are increasingly used to predict how ADCs distribute within the tumor microenvironment. These models account for spatial variables such as cell architecture, antigen density, antibody diffusivity, and tumor clearance of the ADC.

Agent-based models are particularly suited to capturing the stochastic, cell-level heterogeneity of solid tumors that continuum PDE models cannot represent. Calopiz et al. used a validated agent-based framework to test four different payloads conjugated to trastuzumab across a range of target expression levels and doses, demonstrating that the optimal carrier dose and the relationship between tumor saturation and ADC efficacy are payload-class dependent—insights directly applicable to ADC design optimization prior to preclinical in vivo testing.²²² A complementary approach involves machine learning surrogates that replace computationally expensive PDE solvers with trained neural networks, enabling rapid, high-throughput evaluation of thousands of ADC parameter combinations—including DAR, affinity, linker stability, and dosing schedule — within the spatial penetration framework, at a fraction of the computational cost of full mechanistic simulations.²²³

Building on this mechanistic foundation, next-generation approaches integrate digital pathology, leveraging whole-slide histology to map antigen distribution and embed spatial context into penetration and binding models. For example, an AI framework was recently shown to simulate ADC distribution in HER2-positive breast tumors, demonstrating that the spatial organization of malignant and stromal populations can critically influence therapeutic efficacy.⁶⁰

A further advance is the integration of spatial multiomics—combining spatial transcriptomics, multiplex immunofluorescence, and computational pathology—to directly map ADC antigen expression, cell-state information, and microanatomical context with treatment outcomes. Such spatially integrated models can identify resistant niches within the tumor, defined as subregions with low antigen density, poor vascularity, or stromal shielding, that are systematically underexposed to ADC therapy and represent a primary source of residual disease after treatment.²²⁴ AI-based tools such as Path2Space can predict spatial gene expression—including antigen expression maps—directly from hematoxylin-and-eosin–stained whole-slide images, enabling the reconstruction of spatially resolved ADC target landscapes from routinely collected diagnostic pathology without requiring dedicated spatial transcriptomics assays.²²⁵ As described in Fig. 5, these approaches are converging toward a vision in which computational tumor penetration models are personalized to individual patient tumor architectures— informed by digital pathology and spatial omics data—enabling pre-treatment simulation of ADC distribution and efficacy as part of a precision medicine framework for ADC clinical development.²²⁴

Toxicity prediction

ADCs can cause a range of toxicities, some related to their on-target action in healthy tissues and some related to off-target effects of the payload, especially once it is released.² However, most of the toxicity can be attributed to the constituting payload, although the dynamics of the mechanism of action of the ADC, including the antibody and linker cannot be overlooked in terms of safety.⁹

Accurate AI-driven safety assessment of ADCs requires input parameters that extend beyond payload chemistry to incorporate mechanistic features such as antigen expression levels in normal tissues, the dynamics of antigen-ADC complex endocytosis, and FcRn-mediated recycling kinetics. For example, evaluating antigen abundance and internalization rates can help predict on-target toxicity, in addition to the proliferation characteristics of cells in different tissues.²²⁶

Importantly, payload toxicity is also influenced by physicochemical characteristics such as membrane permeability and polarity, as exemplified by differences between MMAE and MMAF, which shape cellular uptake, bystander effects and side effects like ocular toxicity.^7,227 In this context, payload biophysical properties, together with internalization dynamics, can be modeled based on target expression levels and the specific membrane features of normal cells, including lipid composition and transport mechanisms.^7,227 Accurate AI-based toxicity prediction, therefore, requires multidimensional input parameters that integrate payload chemistry with antigen distribution, endocytosis kinetics, and tissue-specific cellular properties.

Several AI-driven frameworks have demonstrated the capacity to address multidimensional ADC toxicity prediction. DumplingGNN, a hybrid Graph Neural Network integrating molecular structure and physicochemical features, was specifically validated for ADC payload toxicity prediction—including early-stage safety assessment of Topoisomerase I inhibitor-based payloads—providing interpretable structure–toxicity relationships that complement activity prediction.¹⁵⁴ At the antibody level, machine learning classifiers have been employed to evaluate Fc-silencing mutations such as LALA and LALAPG, which abrogate antibody-dependent cell-mediated cytotoxicity, enabling computational prediction and minimization of on-target, off-tumor toxicity in ADC applications.⁹⁵ More broadly, multimodal deep learning and transformer-based frameworks trained on ADMET datasets are increasingly capable of simultaneously predicting payload pharmacokinetics and organ-specific toxicity endpoints—including hepatotoxicity and cardiotoxicity—by integrating payload chemistry with tissue-specific molecular features, offering a scalable route to in silico safety profiling prior to experimental testing.²²⁸

“Digital twin” clinical trials

One of the most forward-looking applications of AI in drug development is the creation of “digital twins”, computational representations of patients, to simulate clinical trials.²²⁹ In the context of ADCs, a digital twin system incorporates all available data (patient characteristics, tumor biology, ADC mechanistic models, etc.) to predict how a patient would respond to a given ADC and dose.¹³⁷ Some companies have announced the creation of those platforms although no data has been published and several concerns and limitations have been raised.²³⁰ In practice, this might involve generating a virtual cohort of patients with variability in parameters like antigen expression, tumor size, liver and kidney function (affecting drug clearance), etc., and then running simulations to evaluate outcome distributions.

In silico clinical trials (ISCTs) are increasingly recognized as an important component of the Model-Informed Drug Development (MIDD) framework—an approach now institutionalized by the FDA through its MIDD Paired Meeting Program and Project Optimus dose optimization initiative—enabling virtual patient cohort generation, treatment response simulation, patient stratification, and precision medicine approaches that can directly inform regulatory submissions.²³¹ The generation of credible virtual populations requires sampling from distributions of patient characteristics derived from real clinical data, with careful calibration against observed outcomes—a methodological requirement that adds rigor but also demands high-quality input datasets that are currently sparse for ADC-specific applications.^231,232

These in silico trials, in case they could be performed, could help optimize dosing schedules (balancing efficacy and toxicity), select patient subgroups most likely to benefit, and even identify biomarkers of response.²³³

While digital twin and in silico clinical trial frameworks are increasingly proposed as transformative tools in drug development, their application in ADC development remains largely exploratory and should be considered forward-looking. Existing evidence is primarily limited to partial implementations rather than full patient-level simulations. For example, a recent study demonstrated that integrating digital pathology with computational modeling can predict spatial distribution and efficacy of ADCs within tumors, highlighting how tissue architecture and antigen heterogeneity influence therapeutic response.⁴⁰ In parallel, hybrid physiologically based pharmacokinetic (PBPK) and machine learning approaches have been shown to improve prediction of systemic drug exposure from preclinical data, supporting early dose selection strategies.²¹²

PBPK models have increasingly been submitted as pivotal regulatory evidence in oncology drug approvals— among 245 FDA-approved new drugs from 2020 to 2024, 26.5% included PBPK models, with oncology drugs accounting for the highest proportion at 42%—establishing a regulatory precedent for computational model evidence that ADC developers could in principle leverage for dose optimization submissions.²³⁴

However, these models are generally restricted to pharmacokinetics or tumor-level behavior and do not yet capture the full complexity of ADC pharmacology, including toxicity and clinical response.

In 2025, MIDD has matured to the point where it is now recognized as central to regulatory submissions in oncology, rare diseases, and immunology, with the FDA and EMA articulating frameworks that encourage its use in early development and the FDA’s MIDD Paired Meeting Program serving as a cornerstone initiative fostering sponsor–regulator dialog.²³⁵

Broader digital twin frameworks have been proposed to simulate virtual patient cohorts and treatment outcomes, but these remain largely conceptual and lack prospective validation in ADC clinical development.²²⁹ Therefore, although these approaches hold promise for improving trial design and dose optimization, their current utility in ADC development is not yet supported by robust clinical evidence, and further validation is required before routine application.

Resistance mechanisms and temporal dynamics in ADC development: the unexplored computational frontier

Despite the transformative potential of AI and computational approaches across the ADC development pipeline, the prediction of acquired resistance over time remains a critically underserved area that warrants its own dedicated treatment. Resistance to ADCs is a multistep, temporally evolving biological process spanning the entire drug delivery cascade—from antigen downregulation and impaired internalization to defective lysosomal trafficking, upregulation of efflux transporters, and payload detoxification—and its dynamic nature is not yet meaningfully integrated into existing computational frameworks.^236,237 Experimentally, it has been well established that T-DM1 resistance can arise through impaired lysosomal proteolytic activity, with lysosomal pH elevation and defective cathepsin function leading to accumulation of the conjugate without payload release⁷⁸—a mechanism directly tractable by computational pH-gradient and intracellular trafficking models, yet largely absent from current AI design frameworks. In parallel, defective cyclin B1 induction has been identified as a further mechanism of acquired T-DM1 resistance, demonstrating that mitotic checkpoint subversion represents a distinct and pharmacodynamically actionable resistance pathway²³⁸ HER2 intratumoral heterogeneity has been established as a primary driver of resistance to anti-HER2 ADCs, underscoring that antigen expression must be modeled as a spatially and temporally variable parameter rather than a fixed input^51,52 Adaptive resistance adds a further layer of complexity: prior exposure to one targeted agent can compromise the efficacy of subsequent therapies through deregulation of cell death mechanisms, establishing a cross-resistance logic directly applicable to sequential ADC treatment planning.²³⁹ The importance of generating and systematically characterizing primary and secondary resistance models—a prerequisite for training temporally aware AI systems—has been widely recognized, with in vitro and in vivo strategies including cell-line-derived and patient-derived xenograft models enabling the distinction between de novo and acquired refractoriness.²⁴⁰ A critical mechanistic determinant of resistance to non-cleavable ADCs is loss of the lysosomal transporter SLC46A3, required for cytoplasmic export of maytansinoid catabolites, which has been validated across multiple ADC targets and underscores that transporter expression status should be an integral input to computational models of intracellular ADC fate.⁷⁶ Efflux pump upregulation through ABC transporters including P-glycoprotein, ABCG2, and ABCC1 constitutes a convergent resistance mechanism for multiple clinically used payloads including MMAE, DM1, and DM4, whose substrate status can in principle be predicted computationally from physicochemical descriptors.²⁴¹ Clinically, HER2 loss under selective pressure of T-DXd—with retained sensitivity to TROP2-directed ADCs sharing the same payload—exemplifies how antigen dynamics over time govern treatment sequencing decisions that static transcriptomic AI models are ill-equipped to anticipate.⁶³ Targeting an alternative antigen, such as HER3, has been shown to bypass resistance to anti-HER2 ADCs across multiple resistance models, providing a clinical proof-of-concept for AI-guided antigen switching as a resistance-circumvention strategy.³¹ Collectively, these findings establish that resistance to ADCs is not a singular event but an evolving, multi-mechanism process whose trajectory depends on the interplay between antigen dynamics, intracellular trafficking, transporter biology, and signaling adaptation—none of which are adequately captured by the static, cross-sectional AI models currently deployed in ADC design. What is needed are longitudinal, temporally aware computational frameworks: recurrent neural architectures or dynamical systems models trained on serial single-cell transcriptomics, circulating tumor DNA profiling, and spatial proteomic readouts of antigen dynamics, capable of forecasting resistance trajectories before they become clinically manifest.⁷⁹ Integration of CRISPR-based functional genomics screens—which have begun to systematically uncover endolysosomal regulators and novel sensitivity genes—could supply the essentiality data needed to train AI models that distinguish durable from resistance-prone targets at the earliest stages of ADC development.^242,243 Realizing this vision will require the construction of longitudinal ADC resistance datasets, annotated with temporal molecular profiles matched to clinical outcomes, as a prerequisite for AI models capable of anticipating—rather than merely reacting to—the temporal evolution of ADC resistance.^244–246

Expanding the therapeutic scope of antibody drug conjugates beyond oncology

Although ADCs have primarily been developed for oncology, recent advances have extended this technology to the treatment of several other diseases. In the following section we will describe the limited evidence for the time being that supports their use.

Antibody–drug conjugates in autoimmune diseases

ADCs are in clinical development in autoimmune diseases by enabling targeted delivery of immunomodulatory agents to pathogenic immune cell populations. One notable example is ABBV-3373, an ADC composed of an anti-TNF monoclonal antibody conjugated to a glucocorticoid receptor modulator (GRM).²⁴⁷ This construct was designed to selectively deliver glucocorticoid activity to TNF-expressing inflammatory cells, thereby enhancing anti-inflammatory efficacy while minimizing systemic glucocorticoid exposure. Preclinical studies demonstrated that conjugation of the GRM payload to an anti-TNF antibody improved targeted anti-inflammatory activity in models of inflammatory disease, including arthritis.²⁴⁷ In clinical evaluation, ABBV-3373 has progressed to phase II trials for rheumatoid arthritis, where it demonstrated reductions in disease activity compared with historical outcomes associated with adalimumab therapy.^248,249 These findings highlight the potential of ADCs to enhance the therapeutic index of established biologic therapies through targeted intracellular delivery of anti-inflammatory payloads.

Additional preclinical efforts have focused on delivering glucocorticoids selectively to antigen-presenting cells using ADC technology. CD74-targeted ADCs conjugated with glucocorticoid receptor agonists have been developed to exploit the expression of CD74 on B cells and antigen-presenting immune cells. In vitro studies demonstrated that these constructs increase glucocorticoid signaling in CD74-expressing cells and inhibit B-cell proliferation, suggesting a potential strategy to achieve localized immunosuppression while reducing systemic steroid toxicity.^250,251 Collectively, these studies illustrate the emerging potential of ADC platforms to deliver immunomodulatory agents with improved precision, thereby expanding the therapeutic landscape for autoimmune diseases.

A principal challenge that distinguishes non-oncology ADC development from oncology is target identification: unlike tumor antigens, the pathogenic immune cell populations relevant to autoimmune diseases—including autoreactive B cell subsets, disease-associated synovial fibroblast subtypes, or pathologically expanded plasmacytoid dendritic cell populations—are defined by transient, context-dependent surface phenotypes that are not adequately captured by bulk transcriptomics or conventional immunophenotyping. Single-cell multi-omics technologies, including scRNA-seq, CITE-seq, and spatial transcriptomics applied to inflamed patient tissues, are beginning to resolve these disease-restricted cell states with the resolution required for ADC target nomination.^252,253 AI-driven integration of these datasets across patient cohorts, disease stages, and tissue compartments could systematically identify surface antigens that are selectively and stably expressed on pathogenic populations across a broad range of autoimmune conditions.

Antibody–drug conjugates in neurodegenerative disorders

In neurodegenerative diseases, the application of ADCs is an emerging area of research aimed at enabling targeted delivery of therapeutic agents to pathological protein aggregates in the central nervous system. Unlike oncology, where ADCs typically deliver cytotoxic payloads to tumor cells, ADC strategies in neurodegeneration focus on delivering disease-modifying molecules to neurons or glial cells associated with pathogenic protein accumulation. However, most ADC approaches in this field remain in the preclinical stage due to challenges associated with blood–brain barrier (BBB) penetration and the identification of suitable neuronal targets.

One proposed strategy involves ADCs directed against amyloid-β (Aβ), the peptide that aggregates to form extracellular plaques in Alzheimer’s disease.²⁵⁴ Antibodies targeting Aβ have already demonstrated the ability to bind plaque-associated peptides and facilitate immune-mediated clearance. By conjugating these antibodies to therapeutic payloads, ADC platforms aim to enhance the selective delivery of small molecules or modulatory compounds to plaque-associated neurons and surrounding microglia. Such targeted delivery could potentially improve plaque clearance while minimizing systemic exposure and toxicity, thereby increasing the therapeutic index of anti-amyloid therapies²⁵⁴ In addition, antibody-mediated targeting may enable localized modulation of neuroinflammation and synaptic dysfunction associated with amyloid pathology.

A related strategy focuses on targeting pathological tau aggregates, which are a hallmark of several neurodegenerative disorders collectively known as tauopathies, including Alzheimer’s disease. Tau-directed antibodies can bind extracellular tau species that propagate pathology between neurons. Conjugation of therapeutic payloads to these antibodies has been proposed as a means to deliver disease-modifying agents directly to tau-associated neuronal compartments. Such ADC approaches could potentially inhibit tau aggregation, enhance clearance of pathological tau species, or modulate intracellular signaling pathways implicated in neuronal degeneration.²⁵⁴

Beyond Alzheimer’s disease, the ADC and conjugate paradigm is being actively explored in synucleinopathies such as Parkinson’s disease and multiple system atrophy. A particularly instructive preclinical example is SAR446159 (ABL301), a bispecific antibody composed of an α-synuclein aggregate-targeting immunoglobulin fused to an engineered insulin-like growth factor receptor 1 (IGF1R)-binding single-chain variable fragment that acts as a brain-shuttle, exploiting receptor-mediated transcytosis to substantially increase CNS exposure. In preclinical models, SAR446159 demonstrated preferential binding to fibrillar α-synuclein aggregates over monomers by five orders of magnitude, inhibited α-synuclein seeding in primary neurons, enhanced aggregate clearance by microglia, and reduced pathological spread and motor phenotypes in transgenic mice, with a Phase I clinical trial in healthy volunteers is now completed (NCT05756920).²⁵⁵ This brain-shuttle conjugate design directly mirrors the engineering logic of ADCs and illustrates how antibody-cargo conjugation can be leveraged to overcome the BBB independently of cytotoxic payloads. In parallel, AC Immune has disclosed the morADC (Morphomer®-ADC) platform, in which brain-penetrant small-molecule Morphomers targeting pathological aggregates are covalently conjugated to conformation-selective monoclonal antibodies. Over 30 morADC constructs targeting Aβ, tau, and α-synuclein were engineered with drug-to-antibody ratios of 2.5–4.5, achieving 3–6-fold enhanced BBB penetration versus the parent antibody in vitro, a 2.5-fold increase in brain parenchymal exposure in mice, and synergistic anti-aggregation activity exceeding the potency of each component alone as quantified by isobologram analysis.²⁵⁶ Although AC Immune’s morADC platform demonstrated promising preclinical results—including 3–6-fold enhanced BBB penetration and synergistic anti-aggregation activity—the program was discontinued in September 2025 as part of a pipeline reprioritization that included the removal of several preclinical assets and a reduction of approximately 30% of the company’s workforce, underscoring the translational challenges that continue to face ADC approaches in neurodegeneration.²⁵⁷

The BBB penetration challenge is one in computational tools are directly applicable but remain substantially underutilized in the ADC context. ML models trained on BBB permeability datasets—such as the LeiCNS-PK3.0 framework, which integrates quantitative structure–property relationship (QSPR)-predicted brain-to-plasma partition coefficients (Kp,uu,BBB) with physiologically based pharmacokinetic CNS models—can predict the CNS exposure of antibody-cargo conjugates from molecular features prior to synthesis, enabling rational pre-screening of brain-shuttle receptor targets (such as TfR1, LRP1, or CD98hc) and linker chemistries that may optimize transcytosis efficiency and parenchymal retention.^258,259 These CNS PBPK approaches are currently applied primarily to small molecules and naked antibodies; extending them to ADC-specific parameters represents a high-value computational frontier that could substantially reduce the empirical attrition rate in this indication class. The discontinuation of ADC programs reinforce that computational pre-screening of conjugate properties against CNS delivery parameters should be considered an integral component of the design workflow.²⁵⁹

Beyond classical ADCs, a related class of antibody-oligonucleotide conjugates (AOCs) is emerging as a powerful delivery strategy for neurological and neuromuscular diseases, exploiting receptor-mediated endocytosis via anti-transferrin receptor 1 (TfR1) antibodies to achieve tissue-targeted silencing of disease-causing transcripts. Avidity Biosciences has demonstrated that TfR1-directed AOCs achieve greater than 15-fold higher oligonucleotide concentrations in striated muscle versus unconjugated RNA in non-human primates, with pharmacokinetic/pharmacodynamic translation across species.²⁶⁰

Antibody–drug conjugates in infectious and metabolic diseases

Beyond autoimmune and neurodegenerative disorders, the ADC framework is increasingly being explored in infectious and metabolic diseases. In the infectious disease space, antibody-antibiotic conjugates (AACs)—a direct structural analog of ADCs—have been developed to overcome intracellular bacterial reservoirs that evade conventional antibiotics.²⁶¹ The prototype AAC, DSTA4637S, combines an anti-wall teichoic acid antibody with a novel rifamycin-class antibiotic delivered via a protease-cleavable linker; upon phagocytic internalization of opsonized Staphylococcus aureus, the antibiotic is released intracellularly, achieving killing of methicillin-resistant strains that would otherwise persist inside host cells, with Phase I clinical evaluation completed.²⁶² In metabolic disease, early-stage programs are adapting the ADC payload delivery concept to address atherosclerosis and dyslipidemia: LXR agonist–ADCs have been designed to selectively deliver lipid-modulating payloads to macrophages within atherosclerotic plaques, exploiting macrophage-specific surface antigens to circumvent the hepatotoxicity that has historically limited systemic LXR agonist use.²⁶³ Collectively, these programs illustrate that the precision delivery principle underpinning ADCs—targeted payload release at a defined cell population—is a generalizable therapeutic logic extending well beyond oncology, with the breadth of non-oncology applications expanding rapidly as linker and antibody engineering mature.⁷

Across all non-oncology indications discussed above, a structural limitation shared by current ADC programs is the near-complete absence of disease-specific training datasets for ML models—in sharp contrast to oncology, where large, annotated datasets of ADC activity, resistance mechanisms, and patient outcomes have been progressively assembled to support AI-driven design. Building interoperable multi-omics repositories that capture immune cell surface phenotypes in autoimmune disease at single-cell resolution, CNS cell-type-specific antigen expression profiles in neurodegeneration, and bacterial surface antigen variability in infectious disease will be a prerequisite for extending the AI-driven ADC design framework described throughout this review to these emerging indications.²⁶⁴ In parallel, computational models of payload pharmacology—including intracellular release kinetics, receptor internalization dynamics, and bystander effect modeling—will require recalibration for non-oncology cellular contexts, where the target cell biology differs fundamentally from that of tumor cells in terms of proliferation rate, endosomal trafficking efficiency, and lysosomal protease activity. Finally, regulatory frameworks for computational evidence supporting non-oncology biologics development are at an earlier stage of maturity than in oncology, and proactive engagement between computational ADC developers and regulatory agencies will be essential to define acceptable evidentiary standards for AI-assisted target selection and conjugate optimization in these novel disease contexts.²⁶⁴

Limitations

The application of AI and computational approaches across the ADC development pipeline—from antigen discovery and antibody engineering to linker optimization, payload design, and clinical pharmacology—holds transformative promise, yet important challenges limit their current reliability and translational impact. Model performance is constrained by the availability and quality of training datasets, and that is a critical aspect in ML. In oncology, available data are often sparse, biased, or unbalanced across tumor types, limiting predictive accuracy and external validity.^14,137 Predictions may also lack generalizability across different antigens, payload chemotypes, or tumor contexts, raising concerns about model transferability.

A related but distinct limitation, rarely discussed in the ADC context, is model overconfidence: many deep learning architectures generate predictions with high apparent certainty even when operating outside their training distribution or when the underlying data is insufficient to support confident conclusions. Poor calibration—where model confidence scores do not align with actual prediction accuracy—can mislead researchers into pursuing false leads with costly experimental resources, and is particularly dangerous in ADC design where each synthesis-and-test cycle is expensive and time-consuming.²⁶⁵ This problem is especially acute for generative AI and large language model (LLM)-based approaches, which are increasingly being deployed for molecular design and literature synthesis: these models can produce fluent, plausible, and internally consistent outputs that are factually incorrect —a phenomenon termed confabulation—and may propose molecular designs or structure-activity relationships that appear scientifically credible but are not grounded in experimental evidence.²⁶⁶ Uncertainty quantification methods, including ensemble approaches, Bayesian neural networks, and semantic entropy-based confabulation detection, are emerging as essential complements to predictive models to flag low-confidence outputs before they propagate into experimental pipelines, but their systematic adoption in ADC-focused AI tools remains limited.^265,266

Moreover, interpretability lags behind predictive power in many AI frameworks; while deep learning models such as transformers achieve state-of-the-art performance, their “black-box” nature complicates biological insight and regulatory acceptance.^109,267

Explainable AI (XAI) frameworks—including SHAP, LIME, and attention visualization tools—are increasingly deployed to address this gap, but there remains a fundamental tension between model complexity and interpretability: the most accurate models are typically the least transparent, and post-hoc explanation methods provide approximations of model behavior rather than true mechanistic insight.²⁶⁵ Regulatory agencies are responding to this challenge: the FDA’s January 2025 draft guidance on the use of AI to support regulatory decision-making for drug and biological products explicitly requires developers to define the context of use, demonstrate model generalizability, and provide traceable validation evidence—establishing for the first time a formal framework for AI evidence in regulatory submissions for biologics, including potential ADC applications.²⁶⁸

In that sense, the integration of MD computational approaches, which do not rely on explicit training data sets, can strongly complement the molecular and mechanistic insight lacking from some AI frameworks. Crucially, in silico predictions require rigorous experimental validation, as computational pipelines alone cannot yet substitute for in vitro or in vivo testing.²⁰⁹

A specific and underappreciated aspect of this validation gap is the near-complete absence of prospective experimental validation studies for AI models in the ADC field: the overwhelming majority of published benchmarks reflect retrospective performance on known structures or historical datasets, and it remains unclear how these models perform when applied prospectively to genuinely novel ADC candidates outside their training distribution. Community-accepted benchmark datasets and standardized validation protocols—analogous to CASP in protein structure prediction—are urgently needed to enable objective comparison of competing AI models and to build the evidence base required for regulatory and clinical trust⁹⁵

The integration of AI into clinical development workflows remains uneven, with translational adoption hindered by infrastructure gaps, barriers to data sharing, and the absence of standardized regulatory frameworks.^174,190 Federated learning represents one of the most promising technical solutions to the data sharing barrier: by enabling multiple pharmaceutical organizations to collaboratively train shared models without transferring proprietary raw data — sharing only model gradients or distilled knowledge—federated approaches can substantially expand the effective training dataset available to each participant while preserving data confidentiality. The FLuID framework, validated in a real-world collaboration across eight pharmaceutical companies, demonstrated that federated distillation expands model applicability domains and improves predictive performance compared to models trained on any single company’s data alone, without compromising data privacy or security.²⁶⁹ Applied to ADC development, federated learning could enable industry-wide pooling of ADC activity, pharmacokinetics, and safety data—the precise datasets that are currently too sparse within any single organization to train reliable ML models —while respecting the competitive and regulatory constraints that make direct data sharing impractical. Realizing this potential will require agreement on standardized data formats, shared ontologies, and interoperability standards across the ADC research community.²⁶⁹

Figure 6 describes all mentioned limitations and mitigation strategic plans.

Fig. 6.

Limitations of AI in ADC development, corresponding mitigation strategies, and the regulatory context. Six recurring limitations (left) are mapped one-to-one onto the mitigation strategies that address them (center). Data scarcity and bias, arising from sparse and unbalanced tumor-type datasets, may be mitigated by federated learning enabling cross-institutional model training without data transfer. Model overconfidence, manifesting as poor calibration and false leads, is addressed by uncertainty quantification using Bayesian approaches. LLM confabulation—plausible but incorrect generative outputs—can be flagged by semantic-entropy methods that identify low-confidence generations. Black-box opacity, which limits mechanistic insight and creates regulatory risk, is partly addressed by explainable AI (SHAP, LIME, attention visualization). The prospective validation gap, reflecting predominantly retrospective benchmarking, calls for community benchmarks analogous to CASP for AI models in ADC development. Finally, the gap between molecular dynamics and machine learning—mechanistic insight versus data-driven pattern recognition—motivates hybrid AI–MD frameworks that combine both. All six pairings sit within a regulatory context (right) defined by the FDA draft guidance on AI to support regulatory decision-making (January 2025), which emphasizes context of use and traceability, and by the FDA/EMA model-informed drug development (MIDD) program governing the use of computational evidence in regulatory submissions

Addressing these challenges, through the creation of large, interoperable multi-omic datasets, the development of explainable AI models, and the tighter coupling of computational predictions with experimental validation, will be critical to realize the full translational potential of AI in ADC development.^230,268 The emergence of formal regulatory guidance for AI in drug development, while still evolving, signals that the field is maturing toward a stage where computational predictions can begin to contribute meaningfully to regulatory submissions—provided that the validation, transparency, and generalizability standards now being defined are met.²⁶⁸

Conclusions and perspectives

This review has traced the application of computational and AI-driven approaches across the full ADC engineering design cascade. At the level of antigen discovery, machine learning applied to integrated genomic, transcriptomic and proteomic datasets now permits systematic prioritization of tumor-selective, surface-accessible targets, with mass spectrometry-based surfaceome profiling correcting the imperfect correspondence between transcript abundance and cell-surface protein density that has historically confounded RNA-only prioritization.^22–26 Refinements to this logic—incorporating antigen shedding, receptor turnover, epitope accessibility under glycosylation, intratumoral heterogeneity at single-cell resolution, and CRISPR-derived essentiality as a proxy for resistance potential—move target selection beyond static expression metrics.^{7,8,20,21,34,59} In antibody engineering, generative architectures and protein language models optimize affinity and specificity simultaneously with developability, while diffusion-based frameworks have demonstrated de novo generation of binders against user-specified epitopes, in principle uncoupling antibody discovery from immunization and library screening.^{10,89–91,105} Physics-based simulation—all-atom, coarse-grained and hybrid MD/MC—supplies atomistic mechanistic insight that does not depend on training data and therefore complements, rather than duplicates, data-driven inference.^{10,127,192,206} Finally, hybrid PBPK–AI models and digital twin simulations extend prediction to tumor penetration, systemic exposure, toxicity and dose optimization, and the same computational logic is now being exported to autoimmune, neurodegenerative, infectious and metabolic indications.^{129,212,213,234,263} The unifying theme is the replacement of sequential empirical iteration with simultaneous, multi-parametric in silico optimization.

Several obstacles nonetheless separate this promise from routine practice. The most fundamental is data: training sets in the ADC field are sparse, unbalanced across tumor types and payload chemotypes, and heavily weighted toward a small number of well-studied antigens, which constrains external validity and raises unresolved questions of transferability to rare or under-represented tumor contexts.^48,49,137 Compounding this, many deep learning architectures are poorly calibrated and generate confident predictions outside their training distribution; for generative and large language model-based tools the same tendency manifests as confabulation, producing molecular proposals that are fluent and internally coherent yet experimentally unfounded.^265,266 Because each ADC synthesis-and-test cycle is costly, uncalibrated confidence translates directly into wasted resources. Interpretability remains a second constraint, with the most accurate architectures typically the least transparent and post-hoc explanation methods offering approximations of model behavior rather than mechanistic understanding—a tension with direct consequences for regulatory acceptance.¹⁰⁹ Third, validation in this field is almost entirely retrospective: published benchmarks report performance on historical datasets and known structures, and prospective testing against genuinely novel candidates is rare, leaving the real-world predictive value of most models unquantified.^95,101 Fourth, current frameworks are cross-sectional, treating antigen expression, trafficking competence and transporter status as fixed inputs, and are therefore structurally unable to anticipate acquired resistance, which is an evolving, multi-mechanism process spanning antigen loss, defective lysosomal proteolysis, transporter downregulation and efflux induction.^78,236,241 Finally, the data most needed to close these gaps is distributed across organizations whose competitive and regulatory constraints make conventional sharing impractical.

These limitations are tractable, and several concrete remedies are already available. Federated learning offers the most direct solution to data fragmentation: by exchanging model gradients or distilled knowledge rather than proprietary raw data, multi-institutional collaborations can pool ADC activity, pharmacokinetic and safety datasets that no single organization possesses in sufficient volume, with demonstrated gains in applicability domain and predictive performance in real-world pharmaceutical consortia.²⁶⁹ Realizing this will require agreement on shared ontologies, standardized data formats and interoperability standards, together with pan-cancer surfaceome atlases spanning histologies and geographically diverse patient populations to counteract training bias. The validation gap should be addressed through community benchmark datasets and standardized evaluation protocols analogous to CASP in protein structure prediction, coupled with a deliberate shift toward prospective testing in which models nominate candidates before experimental evaluation rather than being scored on retrospective recall.^95,101 Uncertainty quantification—through ensemble methods, Bayesian architectures and semantic entropy-based confabulation detection—should become a default reporting requirement, so that low-confidence outputs are flagged before they propagate into experimental pipelines.^265,266 Interpretability can be pursued along two complementary routes: explainable AI methods applied to predictive models, and systematic integration of physics-based simulation, which yields mechanistic rationale independent of training data. Anticipating resistance requires purpose-built longitudinal resources—serial single-cell transcriptomics, circulating tumor DNA and spatial proteomic profiling annotated to clinical outcome—as the training substrate for recurrent or dynamical-systems models. Throughout, alignment with emerging regulatory expectations regarding defined context of use, demonstrated generalizability and traceable validation evidence should be planned from the outset rather than retrofitted.^235,268

Looking ahead, we consider several directions particularly promising. The first is genuine co-optimization: antigen, antibody, conjugation site, linker and payload are still optimized component-wise, yet their interactions govern the therapeutic index, and multi-objective frameworks that treat the ADC as a single coupled system are likely to yield greater gains than further refinement of any individual module.^8,10 The second is temporally aware modeling. Resistance is the principal determinant of durability, and forecasting its trajectory—antigen switching under selective pressure, loss of lysosomal export capacity, efflux transporter induction—would convert computational design from a static exercise into a tool for treatment sequencing.^244,246 Third, mechanistic surprises continue to reshape design logic and should be modeled explicitly: the demonstration that extracellular protease-mediated linker cleavage drives activity in low-antigen settings implies that the tumor protease landscape, and not internalization alone, should inform target and linker selection.⁷³ Fourth, de novo epitope-directed antibody generation, combined with computational interrogation of receptor co-expression, opens a combinatorial design space for bispecific and dual-payload constructs that is inaccessible to empirical screening.¹⁵³ Finally, the extension of ADC technology beyond oncology will test the generalizability of these frameworks in settings where the therapeutic window and target biology differ substantially from those of cancer. None of this displaces experimental validation, which remains indispensable. The realistic near-term expectation is therefore not autonomous in silico design but a tighter, faster and better-calibrated loop between prediction and experiment—and it is in that loop, rather than in any single algorithm, that the translational value of AI in ADC development will ultimately be realized.

Author contributions

A.O.: conception, analysis, interpretation of data, drafting the work, and revision of the final draft. A.P. and B.G.: conception and revision of the final draft. J.E. and H.T.: conception, analysis, and interpretation of data. C.A-M.: drafting the work. All authors have approved the submitted version and agree to be personally accountable for their own contributions, as well as to ensure that any questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.

Data availability

The datasets supporting the conclusions of this article are included within the article.

Competing interests

A.O. Consultant fee from NMS. Former consultant of Servier, WWIT, and CancerAppy. Former employee of Symphogen. A.O and J.R.E are co-founders of WBT HK. No competing interests to declare in relation to this work.