Development of a Chatbot to answer oncology related questions

We have been developing a AI methods that are specifically focused on answering questions related to oncology drug development.

One challenge for antibody drug conjugates (ADCs) is understanding how we can improved the clinical success rates for ADCs. By understanding why certain ADCs fail in the clinic, perhaps this can help guide the next generation of ADCs, which will hopefully address some of the limitations of the current wave of ADCs.

ADCs are complex molecules consisting of an antibody that binds to a cell surface receptor expressed on cancer cells, a cytotoxic drug and that is covalently attached to the antibody. Once the ADC binds to the cancer cell, it is internalized, releases the cytotoxic drug/payload, inside of the cell where the payload inhibits critical cell functions such as DNA synthesis/replication or tubulin function/cell replication, that ultimately kills the cancer cells.

There are numerous things to also take into consideration such as antibody formats, conjugation strategies, the selection of the payload, DARs, CMC challenges, pharmacokinetics, toxicology challenges to improved the therapeutic window, dose and schedules, clinical designs, patient enrollment strategies, etc.

Given the vast and complex preclinical and clinical data sets for ADCs, we are developing methods to investigate how AI could be used to help analyze the publicly available information and provide some insights into why ADCs have failed in the clinic. One approach is to develop a chatbot that is exclusively focused on answering questions related to ADC development and oncology drug development. The second approach is to develop an AI driven approach that provides a risk assessment for ADCs.

Figure 1 provides a high-level overview of things to consider for ADC development and how AI can be used to help drug developers make better informed decisions.

Figure 1.

We evaluated OpenAI's LLMs (i.e. GPT 5.4-mini, GPT 5.5, etc) and Anthropic's LLM, Claude (i.e. Sonnet, Opus,etc) and integrated these with other tools into our chatbot. We found that Claude was vastly superior to OpenAI's LLMs in providing more detailed answers to more complex questions but OpenAI's LLMs were faster.

One question we asked was "Why do ADCs fail during clinical development?" Here are the responses to that question...

🔑 Commonality 1: The Vast Majority Fail at Phase 1 — Before Efficacy Can Be Established

Tallying where the ADCs failed during clinical development:

Highest Stage	% of Total
Phase 1 only	~67%
Phase 1/2	~13%
Phase 2	~14%
Phase 2/3	~2%
Phase 3	~4%

~80% of all clinically halted ADCs never progressed past Phase 1/2. This is the single most important structural insight from the data. The consistent Phase 1 attrition signals that dose-limiting toxicities (DLTs) in the first-in-human dose-escalation phase are the proximate cause of most discontinuations — programs are stopped before there is even a reliable efficacy signal to assess.

The five Phase 3 failures represent especially instructive cases: Rovalpituzumab tesirine (DLL3-PBD, SCLC), Vadastuximab talirine (SIGLEC3-PBD, AML), BAT8001 (HER2-maytansine, breast), Trastuzumab duocarmazine/SYD985 (HER2-DUBA, breast), and SYSA1801 (CLDN18.2-MMAE, gastric) — all reached Phase 3 and were stopped either due to safety halts or failure to outperform standard of care.

🔑 Commonality 2: Payload Class Is the Dominant Failure Driver

Payload Distribution Across Discontinued/Terminated Clinical ADCs

Payload Class	Approximate Count	Key Examples
Auristatin (MMAE / MMAF / variants)	~43	AGS67E, MEDI-547, Glembatumumab vedotin, Ladiratuzumab vedotin, Cofetuzumab pelidotin, Vandortuzumab vedotin, Enapotamab vedotin, PRO1107, SGN-CD48A, SGN-STNV, BAY79-4620, ASG-5ME, many more
Maytansinoid (DM1 / DM4 / variants)	~27	BAT8001, TAA013, Lorvotuzumab mertansine, SAR428926, SAR566658, BIIB015, AVE9633, Cantuzumab ravtansine, Cantuzumab mertansine, LOP628, AMG 172, AMG 595, PCA062, IMGN388, MLN2704, HKT288, LY3076226, Bivatuzumab Mertansine, and others
PBD dimer / PBD monoamide	~18	Vadastuximab talirine, Rovalpituzumab tesirine, ADCT-502, SGN-CD70A, SGN-CD19B, SGN-CD123A, SGN-CD352A, MEDI7247, DHES0815A, DS-9606a, SC-002, SC-004, Tamrintamab pamozirine, Olintatug tesirine, Rolinsatamab talirine, MEDI2228, ADCT-401, Mipasetamab uzoptirine
Duocarmycin (DUBA / seco-DUBA)	~6	Trastuzumab duocarmazine (SYD985), SYD1875, Vobramitamab duocarmazine, BYON4413, MDX-1203
Calicheamicin	~3	PF-06647263, CMB-401, CMD-193
Tubulysin	~4	MEDI4276, DXC-004, BMS-986148, BMS-986183
Non-cytotoxic / unconventional	~8	Trastuzumab imbotolimod (TLR7/8), NJH395 (TLR7), SBT6290 (TLR8), Pertuzumab zuvotolimod (TLR8), ABBV-3373 (glucocorticoid), Adalimumab fosimdesonide (glucocorticoid), Milatuzumab doxorubicin (anthracycline), ORM-5029 (GSPT1 degrader)
Other / Undisclosed	~14	Various undisclosed payloads including PF-06688992, SGN-CD47M, SC-005/007, DB-1202, etc.

The two dominant failed payload classes — auristatins (~43) and maytansinoids (~27) — together account for approximately 57% of all clinical failures. This is profound: these are also the two oldest, most extensively used payload families in the field. Their failure rate at scale reflects a fundamental limitation — they are not sufficiently potent in the heterogeneous solid tumor setting when delivered at the DAR levels (~2–4) achievable with conventional conjugation without causing unacceptable off-tumor toxicity.

The PBD class (~18 entries) has the most alarming failure-to-use ratio: while PBDs represent a small fraction of all ADC programs ever initiated, they constitute ~15% of all clinical failures — reflecting the extreme toxicity of PBD dimers that was ultimately not overcome by any target or linker engineering approach at the DAR 2 level used in most PBD programs.

🔑 Commonality 3: Linker Chemistry Patterns — A Clear Generational Signature of older-generation linker chemistry:

Linker Distribution (Clinical Failures)

Linker Type	Count	Pattern / Failure Mode
Valine-Citrulline (Val-Cit)	~55	Most common linker overall — its over-representation reflects volume of use, not linker failure per se. The Val-Cit + auristatin combination at DAR 2–4 proved insufficient for most solid tumor indications
SPDB / Sulfo-SPDB (disulphide-based)	~10	SAR428926, SAR566658, Cantuzumab ravtansine, BIIB015, AVE9633, Indatuximab ravtansine, Coltuximab ravtansine, IMGN388, IMGN779, Tusamitamab ravtansine — disulphide bonds are susceptible to reductive cleavage by plasma thiols, causing off-target payload release before tumor delivery
SMCC (non-cleavable maleimide)	~10	TAA013, B003, SHR-A1201, AMG 595, LOP628, Laprituximab emtansine, AVID100, PCA062, AMG 224, Naratuximab emtansine — non-cleavable linker prevents bystander killing; requires full lysosomal catabolism for payload release, insufficient in tumors with heterogeneous target expression
MC (Maleimidocaproyl)	~6	MEDI-547, Vorsetuzumab mafodotin, Denintuzumab mafodotin, AGS16F, PF-06263507, Lonigutamab Ugodotin — maleimide retro-Michael exchange transfers payload to serum albumin and cysteine-34 in plasma, increasing systemic drug exposure
Val-Ala (cleavable)	~12	Rovalpituzumab tesirine, SC-002/004, SGN-CD70A, SGN-CD19B, SGN-CD123A, SGN-CD352A, Olintatug tesirine, ADCT-502, MEDI7247, and others — the Val-Ala + PBD combination appears particularly prone to systemic PBD toxicity
SPP / SPDB disulphide variants	~5	Lorvotuzumab mertansine, MLN2704, Bivatuzumab mertansine, Cantuzumab mertansine, MLN2704 — older disulphide linker with plasma instability
AcBut acid-hydrazone	~2	PF-06647263, CMB-401 — original calicheamicin acid-labile linker prone to hydrolysis at near-neutral plasma pH
Fleximer polymer	~2	XMT-1522 (DAR~12), Upinitatug rilsodotin (DAR~10) — extreme-DAR polymer approaches showing unfavourable PK
Undisclosed	~15	Multiple programs with no public linker data

Key observation: The combination of disulphide-based linkers (SPDB/SPP) + maytansinoids accounts for a large fraction of the ImmunoGen/AbbVie/Sanofi-partnered failures. The disulphide is cleaved by glutathione and plasma thiols before sufficient tumor delivery occurs, releasing free maytansinoid into circulation. This design, while enabling conjugation at low DAR (~2–4) with reasonable tolerability, consistently underperformed on efficacy across diverse solid tumor targets.

🔑 Commonality 4: Target Antigen — The Same Problem Targets Appear Repeatedly

Examining the target field reveals a cluster of repeatedly attempted, repeatedly failed antigen targets:

Targets with Multiple Failed Clinical ADC Programs

Target	Failed Programs	Why It Keeps Failing
HER-2	BAT8001, TAA013, MEDI4276, PF-06804103, DHES0815A, ADCT-502, SHR-A1201, NJH395, B003, Trastuzumab duocarmazine, zanidatamab zovodotin, Trastuzumab imbotolimod	Competitive displacement by T-DXd; payload or DAR insufficiency vs T-DXd's 60-80% ORR benchmark
EGFR	Depatuxizumab mafodotin, Laprituximab emtansine, Losatuxizumab vedotin, Serclutamab talirine, AVID100, ABBV-637, DXC-004	On-target skin/GI/corneal toxicity limits dose intensity below efficacious levels
CD70	Vorsetuzumab mafodotin, SGN-CD70A, AMG 172, Vorsetuzumab mafodotin variants, MDX-1203	Normal T/B cell expression; immunosuppression and lymphocyte depletion
SIGLEC3 (CD33)	Vadastuximab talirine, AVE9633, IMGN779, ABBV-787	Normal myeloid progenitor expression causes fatal myelosuppression
IL-3R (CD123)	SGN-CD123A, BYON4413, Lixarkitug samrotecan, VIP943	Normal haematopoietic progenitor expression
Mesothelin	DMOT4039A, BMS-986148, BMS-986183	Normal serosal expression causes pleuritis/peritonitis; heterogeneous tumor expression
CLDN18.2	SYSA1801, SOT102, TQB2103, LM-102, Elatatug vedotin	Competitive displacement by better-engineered programs; early programs had suboptimal efficacy
SLC34A2	Upinitatug rilsodotin, XMT-1592, Lifastuzumab vedotin	Extreme DAR (polymer); prior programs outpaced by improved designs

The repeated failure of the same targets across multiple companies and payload classes (HER-2, EGFR, CD70, CD33, IL-3R) strongly suggests that the biology of these targets — particularly their expression on critical normal tissues — creates a structural ceiling that cannot be overcome simply by switching payloads or linkers. These are not random failures; they represent a systematic incompatibility between target tissue distribution and the achievable therapeutic window of any conventional ADC.

The Most Failure-Prone Target Categories:

1. Normal haematopoietic targets (CD33/SIGLEC3, IL-3R/CD123, CD56, CD44v6, FLT3, cKIT): All express on normal marrow progenitors → inevitable myelosuppression at doses required for tumor cell killing. Multiple programs against each have all failed.

2. Broadly expressed epithelial targets (EGFR, Mesothelin, CA9, CanAg): Expression in skin, GI, renal, or serosal normal tissues causes dose-limiting off-tumor toxicity before adequate tumor exposure.

3. Immune cell targets (CD70, CD37, CD19, CD56): Expression on critical normal immune cells causes unacceptable immunosuppression, cytopenia, or neurotoxicity.

   
   

🔑 Commonality 5: DAR Distribution — The "Too Low" Problem Dominates

Directly tallying DAR values from failed ADCs (where disclosed):

DAR Range	Count of Programs	Interpretation
DAR ≤ 2	~20	SGN-CD123A (1.9 VAD/PBD), ADCT-502 (1.7), SYD1875 (1.7), MDX-1203 (1.25), Vandortuzumab (1.8–2.0), MEDI2228 (2), Iladatuzumab (2), SC-004 (2), SGN-CD19B (2), DHES0815A (2), and many others
DAR 2–4	~55	The majority — conventional first/second-gen ADCs at standard conjugation
DAR 4–8	~20	Including a few higher-DAR approaches
DAR > 8	~3	XMT-1522 (DAR~12), Upinitatug rilsodotin (DAR~10), DS-6157 (DAR 8, GGFG-DXd format — only one with modern payload class; discontinued due to target biology, not DAR)

The concentration of failures at DAR ≤ 2 is particularly revealing. Several programs — especially with PBD payloads — attempted to use extremely potent cytotoxins at very low DAR (1.7–2.0) to control systemic toxicity. The result was that even with payloads hundreds of times more potent than MMAE, the absolute drug delivery to tumor cells was insufficient for solid tumor eradication, while systemic toxicity from the small fraction of unbound or deconjugated payload remained dose-limiting.

The "Goldilocks Zone" confirmed by the data: Virtually all approved ADCs cluster at DAR 3–8, with modern high-DAR ADCs (T-DXd at DAR 8; sacituzumab govitecan at DAR 7.6) succeeding precisely because their combination of cleavable linker, homogeneous conjugation, and membrane-permeable payload enables sufficient tumor killing even at low overall antigen expression.

🔑 Commonality 6: The AbbVie/Stemcentrx and Seagen Legacy — Portfolio-Level Failures

Two company portfolios account for a disproportionate share of the failures, reflecting both scientific and corporate strategic failure modes:

AbbVie/Stemcentrx Cluster (~9 programs halted)

ADC	Target	Payload	Status
Rovalpituzumab tesirine	DLL3	PBD (Tesirine)	Discontinued Phase 3
SC-002	DLL3	PBD (SG3199)	Discontinued Phase 1
SC-004	CLDN6/9	PBD (SG3199)	Discontinued Phase 1
SC-006	RNF43	PBD (SC-DR003)	Discontinued Phase 1
SC-005	TAA (undisclosed)	Undisclosed	Discontinued Phase 1
SC-007	Undisclosed	Undisclosed	Discontinued Phase 1
ABBV-011	SEZ6	Calicheamicin	Discontinued Phase 1
ABBV-787	SIGLEC3	bromodomain (BD) and extraterminal domain (BET) inhibitor	Discontinued Phase 1
ABBV-637	EGFR	Undisclosed	Discontinued Phase 1

This cluster represents one of the most concentrated ADC portfolio failures in the industry, stemming from AbbVie's 2016 acquisition of Stemcentrx.

The underlying commonality: all Stemcentrx ADCs relied on PBD-based payloads (Tesirine/SG3199/SC-DR003) at DAR 2 via Val-Ala linkers, targeting tumor-initiating cell antigens. Rovalpituzumab tesirine's Phase 3 failure in SCLC (due to PBD serosal toxicity, photosensitivity, and edema exceeding efficacy benefit) effectively invalidated the entire Stemcentrx portfolio approach.

Seagen "First-Generation MMAF and PBD" Cluster (~8+ programs halted)

ADC	Target	Payload
Vorsetuzumab mafodotin	CD70	MMAF
Denintuzumab mafodotin	CD19	MMAF
AGS16F	ENPP3	MMAF
MEDI-547	EphA2	MMAF
PF-06263507	5T4	MMAF
Lonigutamab Ugodotin	IGF-1R	MMAF
SGN-CD70A	CD70	PBD (SGD-1882)
SGN-CD19B	CD19	PBD (SGD-1882)

Seagen's first-generation maleimidocaproyl (MC)-linked MMAF ADCs (Vorsetuzumab, Denintuzumab, AGS16F, MEDI-547, PF-06263507) all shared the same fundamental design liability: MC linker retro-Michael exchange in circulation transferring MMAF to serum albumin and other plasma proteins, increasing systemic auristatin exposure relative to tumor delivery.

🔑 Commonality 7: The "Unconventional Payload" Experiment — Uniformly Unsuccessful

A striking cluster of ADCs involves programs that attempted entirely non-cytotoxic payloads:

ADC	Target	Payload Class	Status
Trastuzumab imbotolimod	HER-2	TLR7/8 agonist	Discontinued Phase 2
Pertuzumab zuvotolimod	HER-2	TLR8 agonist	Discontinued Phase 1/2
SBT6290	Nectin-4	TLR8 agonist	Discontinued Phase 1/2
NJH395	HER-2	TLR7 agonist	Discontinued Phase 1
ABBV-3373	TNF-α	Glucocorticoid Receptor	Discontinued Phase 2
Adalimumab fosimdesonide	TNF-α	Glucocorticoid Receptor	Discontinued Phase 2
Milatuzumab doxorubicin	CD74 (CLIP)	Doxorubicin	Discontinued Phase 1/2
ORM-5029	HER-2	GSPT1 degrader	Discontinued Phase 1
RG7861	Tfr1	Antibiotic (Rifalazil)	Discontinued Phase 1
TAK-500	CCR2	STING agonist	Discontinued Phase 1/2

Every program using a non-cytotoxic payload in this dataset was halted. The glucocorticoid-ADC approach for autoimmune diseases (Adalimumab fosimdesonide, ABBV-3373) failed to demonstrate sufficient superiority over systemic steroids at tolerable doses. The TLR agonist programs demonstrated that immune stimulation via ADC delivery produces complex and difficult-to-manage cytokine-driven toxicity without reproducible anti-tumor responses. The GSPT1 degrader ORM-5029 and the STING agonist TAK-500 represent even more novel mechanistic bets that also did not advance.

This is a highly significant negative signal for the field: It suggests that the ADC format — binding, internalization, lysosomal trafficking, and payload release — is specifically well-suited to cytotoxic small molecules but has not yet successfully enabled other pharmacological modalities at sufficient therapeutic index.

🔑 Commonality 8: Head-to-Head Displacement by Superior Competitors

Discontinuation of some ADCs was driven not by intrinsic drug failure but by the emergence of a superior competitor that made continued development commercially and clinically untenable:

Displaced ADC(s)	Winner That Displaced Them	Indication
BAT8001, TAA013, zanidatamab zovodotin, MEDI4276, ADCT-502	Trastuzumab deruxtecan (T-DXd)	HER2+ breast/gastric
Pinatuzumab vedotin, Coltuximab ravtansine, iladatuzumab vedotin, SGN-CD19B	Loncastuximab tesirine, polatuzumab vedotin	B-cell lymphoma
SYSA1801, elatatug vedotin, LM-102, SOT102, TQB2103	Zolbetuximab (non-ADC), improved CLDN18.2 ADCs	Gastric/GEJ cancer
MLN2704, ADCT-401, PSMA ADC	177Lu-PSMA-617 (radioligand), improved PSMA-ADC designs	Prostate cancer
Labetuzumab govitecan	Improved CEACAM5 ADC designs (Tusamitamab ravtansine, though itself also failed)	CEA+ solid tumors

This competitive displacement pattern illustrates a uniquely ADC-specific challenge: the field is evolving so rapidly that a program that was reasonable to advance 3–5 years ago may be rendered obsolete by a newer design before it completes development. This is distinct from most drug class failures and places a premium on speed to pivotal data and clear differentiation from the moment of program initiation.

📊 The Master Summary Table

#	Commonality	Data Signal	Core Mechanism
1	Phase 1 failure dominates (~80%)		DLTs prevent dose escalation to efficacious exposures
2	Auristatins + maytansinoids = 57% of failures		Insufficient TI at conventional DAR; off-tumor delivery of cytotoxic payloads
3	Legacy linker chemistry overrepresented	SPDB/SPP (~10), SMCC (~10), MC (~6), AcBut (~2)	Plasma instability, maleimide exchange, no bystander killing, premature release
4	Same targets fail repeatedly across companies	HER-2 (~9), EGFR (~6), CD70 (~5), SIGLEC3 (~4), IL-3R (~4)	Normal tissue target expression creates irreducible floor of off-tumor toxicity
5	Low DAR (≤2) programs fail despite potent payloads	~20 records at DAR ≤ 2	Insufficient absolute drug delivery to tumor; systemic toxicity from released payload persists
6	Portfolio-level failures (AbbVie/Stemcentrx, Seagen)	~17+ programs from 2 company clusters	PBD class toxicity + MC linker limitations + competitive displacement
7	Non-cytotoxic payloads universally fail	~10 records (TLR, glucocorticoid, degrader, antibiotic)	ADC format unsuited to non-cytotoxic modalities at current stage of development
8	Competitive displacement	HER-2 post-T-DXd, PSMA post-177Lu, CLDN18.2 post-zolbetuximab	Superior competitor makes continued development commercially inviable

🎯 The Unifying Principle

The dose required to kill tumor cells could not be separated from the dose that caused unacceptable toxicity to normal tissues.

This manifests differently depending on the specific failure mode — through the payload (too toxic at the doses delivered), through the linker (releasing drug before reaching the tumor), through the target (also expressed on critical normal cells), through the DAR (too low to deliver enough drug, too high causing aggregation/rapid clearance) — but the endpoint is always the same. The field's success stories (brentuximab vedotin, T-DXd, enfortumab vedotin, sacituzumab govitecan) all succeeded precisely by finding combinations where the tumor-to-normal tissue selectivity was genuinely achievable — either through highly tumor-restricted target expression, optimized linker-payload pairs enabling bystander killing without systemic toxicity, or both.

🔬 Commonalities Among Discontinued or Halted ADCs

📊 Discontinued & Halted ADCs (Selected Examples)

ADC	Target	Payload	Linker	Company	Status
Rovalpituzumab tesirine	DLL3	Tesirine (PBD)	Val-Ala	AbbVie/Stemcentrx	Discontinued (Phase 3)
Vadastuximab talirine	SIGLEC3 (CD33)	SGD-1882 (PBD)	Val-Ala	Seagen	Discontinued (Phase 3)
Depatuxizumab mafodotin	EGFR	MMAF (Auristatin)	MC	AbbVie/Seagen	Discontinued (Phase 3)
Tusamitamab ravtansine	CEACAM5	DM4 (Maytansine)	SPDB	Sanofi/ImmunoGen	Discontinued (Phase 3)
SYSA1801	CLDN18.2	MMAE (Auristatin)	Val-Cit	CSPC/Elevation Oncology	Discontinued (Phase 3)
Upinitatug rilsodotin	SLC34A2	Auristatin F-HPA	Fleximer Polymer	Mersana	Discontinued (Phase 3)
BAT8001	HER-2	Maytansine	3AA	Bio-Thera Solutions	Discontinued (Phase 3)
Trastuzumab duocarmazine	HER-2	Duocarmycin (DUBA)	Val-Cit	Byondis	Discontinued (Phase 3)
Vobramitamab duocarmazine	B7-H3	Duocarmycin (DUBA)	Val-Cit	MacroGenics/Byondis	Discontinued (Phase 2/3)
Glembatumumab vedotin	gpNMB	MMAE (Vedotin)	Val-Cit	Celldex	Discontinued (Phase 2)
Ladiratuzumab vedotin	LIV-1	MMAE (Vedotin)	Val-Cit	Pfizer/Seagen	Discontinued (Phase 2)
Cantuzumab ravtansine	CanAg	DM4 (Maytansine)	SPDB	ImmunoGen/AbbVie	Discontinued (Phase 2)
SAR566658	CA6	DM4 (Maytansine)	SPDB	Sanofi/ImmunoGen	Discontinued (Phase 2)
Indusatumab vedotin	GUCY2C	MMAE	Val-Cit	Takeda/Millennium	Discontinued (Phase 2)
Lorvotuzumab mertansine	CD56	DM1 (Maytansine)	SPP	ImmunoGen/AbbVie	Discontinued (Phase 2)
Enapotamab vedotin	Axl	MMAE (Vedotin)	Val-Cit	GenMab/Seagen	Discontinued (Phase 1/2)
Cofetuzumab pelidotin	PTK7	Auristatin (PF-06380101)	Val-Cit	AbbVie/Pfizer	Discontinued (Phase 1)
Bivatuzumab mertansine	CD44v6	DM1 (Maytansine)	SPP	Boehringer/ImmunoGen	Discontinued (Phase 1)
Mipasetamab uzoptirine	Axl	SG3199 (PBD)	Val-Ala	ADC Therapeutics	Discontinued (Phase 1)
ADCT-502	HER-2	SG3199 (PBD)	Val-Ala	ADC Therapeutics	Discontinued (Phase 1)
NJH395	HER-2	TLR7 agonist	Maleimide	Novartis	Discontinued (Phase 1)
Pertuzumab zuvotolimod	HER-2	TLR8 agonist	Val-Cit	Silverback/ARS Pharma	Discontinued (Phase 1/2)

Part 2.

🔑 Key Commonalities in Discontinued ADCs

1. 🧪 Payload-Related Issues — Maytansines (DM1/DM4) & Early-Generation Auristatins

One of the clearest patterns is the over-representation of maytansinoid (DM1/DM4) and first-generation auristatin (MMAE/MMAF) payloads among discontinued ADCs:

• DM1/DM4: Cantuzumab ravtansine, Cantuzumab mertansine, SAR566658, SAR428926, BIIB015, Lorvotuzumab mertansine, AVID100, Naratuximab emtansine, PCA062, AMG 595, BAT8001, BAT8003, LY3076226

• MMAE/MMAF: Glembatumumab vedotin, Ladiratuzumab vedotin, Depatuxizumab mafodotin, Cofetuzumab pelidotin, Enapotamab vedotin, Vadastuximab talirine, AGS16F, AGS67E, SYSA1801, and many others

Why? These payloads carry inherent toxicity risk — particularly peripheral neuropathy (MMAE/MMAF), hepatotoxicity (DM1/DM4 maytansine via liver uptake), and a relatively narrow therapeutic window. Their efficacy in later-phase trials frequently did not overcome these toxicity issues in unselected or heterogeneous tumor populations.

2. 🎯 Poorly Validated or Low-Priority Targets

Many discontinued ADCs targeted antigens that ultimately proved to be suboptimal for the ADC modality:

Target	ADCs Discontinued
CD70	Vorsetuzumab mafodotin, SGN-CD70A, AMG 172
Axl	Enapotamab vedotin, Mipasetamab uzoptirine
PTK7	Cofetuzumab pelidotin, PRO1107
gpNMB	Glembatumumab vedotin
CD37	AGS67E
CanAg/CA6/CA9	Cantuzumab ravtansine, SAR566658, BAY79-4620
FGFR2	Aprutumab ixadotin
5T4	PF-06263507, SYD1875, recaLRIp

Why? These targets were either heterogeneously expressed, showed low tumor specificity, or were expressed on normal tissue, all of which can lead to insufficient tumor delivery and/or unacceptable off-tumor toxicity.

3. ☠️ Dose-Limiting Toxicity / Unmanageable Safety Profile

A recurring theme across multiple programs:

• Pyrrolobenzodiazepine (PBD) payloads showed problematic toxicity — Rovalpituzumab tesirine (rova-T), Vadastuximab talirine, SGN-CD70A, SGN-CD19B, ADCT-502, and others all carry PBD warheads. PBD dimers are highly potent DNA crosslinkers, which translates to narrow safety margins and poor tolerability (liver toxicity, serosal effusions, rash).

• Several programs with duocarmycin payloads (Trastuzumab duocarmazine, MEDI4276) were stopped due to toxicity or insufficient efficacy.

• Calicheamicin-based ADCs (CMD-193, PF-06647263, SGN-15) were discontinued partly due to cumulative off-target damage.

  
  

4. 🔗 Linker Instability or Suboptimal Drug-to-Antibody Ratio (DAR)

Many older-generation ADCs used non-site-specific conjugation (e.g., lysine, random cysteine), resulting in:

• Heterogeneous DAR mixtures — some drug species were cleared too quickly, others were overly aggregated

• Linker instability — premature payload release in circulation causing systemic toxicity without tumor delivery

• High DAR formats with some technologies (XMT-1522 with DAR ~12, XMT-1592, Upinitatug rilsodotin with Fleximer Polymer) showed suboptimal pharmacokinetics

The Fleximer polymer platform (Mersana) appears repeatedly in discontinued/halted programs (XMT-1522, XMT-1592, Upinitatug rilsodotin), suggesting challenges with this conjugation format at scale.

5. 📉 Late-Phase Efficacy Failure in Phase II/III — Unselected or Over-Pretreated Patients

Many ADCs failed in Phase 3 despite promising Phase 1/2 signals:

• Rovalpituzumab tesirine (rova-T) — failed three Phase 3 trials in DLL3+ small cell lung cancer (SCLC), partly due to toxicity/narrow therapeutic index and modest efficacy in unselected populations.

• Tusamitamab ravtansine — CEACAM5-targeting ADC that failed Phase 3 endpoint in NSCLC.

• Depatuxizumab mafodotin — anti-EGFR ADC with MMAF that failed in glioblastoma despite promising Phase 1.

• BAT8001 and Trastuzumab duocarmazine — HER2 ADCs that could not compete with trastuzumab deruxtecan (T-DXd) and trastuzumab emtansine benchmarks.

  
  

6. 🏆 Competition From Superior ADCs or Emerging Therapies

One of the most consequential modern reasons for ADC discontinuation is being outcompeted by next-generation agents:

• HER2 space: Trastuzumab deruxtecan (T-DXd, GGFG/DXd platform) has dramatically raised the bar. Multiple older HER2 ADCs with DM1, MMAE, duocarmycin, or PBD payloads have been discontinued (BAT8001, GB251, DHES0815A, XMT-1522, ADCT-502, Trastuzumab duocarmazine, MEDI4276).

• TROP-2 space: Sacituzumab govitecan's success led to deprioritization of BAT8003 and PF-06664178.

• CD33 space: The discontinuation of Vadastuximab talirine was partially driven by safety concerns in AML, where less toxic alternatives became available.

  
  

7. 📋 Strategic/Commercial Deprioritization

Several ADCs were removed from clinical development not necessarily because of clinical failure, but due to:

• Portfolio reshuffling (e.g., Luveltamab tazevibulin deprioritized by Sutro Biopharma, Praluzatamab ravtansine deprioritized by CytomX)

• Business decisions — partnerships ended, companies pivoted, or assets were not commercially compelling

• Orphan or small-market indications — insufficient patient population to justify full development costs

  
  

🧬 Summary Table of Commonalities

Commonality	Examples
Maytansinoid/MMAE payload toxicity	DM1/DM4 ADCs, early auristatin ADCs
PBD payload narrow therapeutic index	Rova-T, Vadastuximab, SGN-CD70A, ADCT-502
Suboptimal target selection/expression	CD70, Axl, gpNMB, CanAg, PTK7, CD37
Linker instability or heterogeneous DAR	Fleximer-based ADCs (Mersana), early random-conjugation formats
Late-phase efficacy failure	Rova-T, Depatuxizumab, Tusamitamab ravtansine
Competition from superior ADCs	HER2 space (vs. T-DXd), TROP-2 space
Safety signals in Phase 1	Bivatuzumab (lethal skin toxicity), PBD-based ADCs
Strategic/commercial deprioritization	Luveltamab tazevibulin, Praluzatamab ravtansine
Immunostimulatory ADC challenges	NJH395, Pertuzumab zuvotolimod, TAK-500

💡 Key Takeaways

1. Payload evolution matters: The shift from DM1/DM4 and auristatins to topoisomerase-I inhibitors (DXd, exatecan, SN-38) has been driven by learning from earlier failures — the newer payloads offer a better bystander effect, broader therapeutic index, and novel mechanism of resistance.

2. Target selection is paramount: Heterogeneous, low-expression, or non-specific targets account for a large fraction of failures, underscoring the need for rigorous preclinical target validation and companion diagnostics.

3. PBD warheads remain high-risk — their exceptional potency is a double-edged sword, and most PBD-based ADCs have encountered safety-related discontinuations.

4. The landscape is increasingly competitive: A drug that was "good enough" five years ago may now be discontinued simply because a better option (higher DAR, better payload, more stable linker) from a competitor program has set a new efficacy benchmark.

5. Platform-level learning: Companies like ImmunoGen (DM1/DM4), Seagen (MMAE/MMAF), and ADC Therapeutics (PBD) have had multiple discontinued programs, but used that learning to advance next-generation designs.

The analysis provides a reasonable overview of why certain ADCs failed during clinical development. It also provides insights into specific payloads that have had significant challenges, such as PBDs, although one ADC with a PBD payload, Loncastuximab Tesirine (anti-CD19), is approved and TLR7/8 agonists.

The analysis isn't without flaws because It missed the recent failure of the beta6 integrin ADC, Sigvotatug vedotin. This could be due to the lack of advanced ADC against the beta6 integrin.

Overall, our chatbot can help prepare a comprehensive overview on certain questions but caution needs to be used because AI can and will hallucinate. To reduce this from occuring, we incorporated an anti-hallucination protocol and forced our chatbot to provide validated references to support the conclusions.