How do you know if two cancer drugs can be combined safely?
Two drugs each work, but will combining them stack toxicities? How do I tell promising pairings from intolerable ones?
Two effective drugs don't make a safe combination if they share a toxicity, and most databases only list single-drug side effects, so combination risk is invisible. In a custom build on the PharosBio Clinical Trial Retriever API, we extracted efficacy (PFS, hazard ratio) and adverse-event profiles across breast-cancer subtypes, then scored which pairings stack toxicities - explore the interactive heatmaps below.
Who this is for: Translational, clinical-pharmacology, and safety scientists assessing whether two agents can be combined without stacking toxicities.
- most efficacious interventions, paired by mechanism
- top 30
- TNBC, PD-L1, BRCA1/2, HER2-low, other
- 5 subtypes
- every score auditable down to the event data
- raw AE counts
- API also available as a skill in Hydra
- custom build
most efficacious interventions, paired by mechanism
TNBC, PD-L1, BRCA1/2, HER2-low, other
every score auditable down to the event data
API also available as a skill in Hydra
The problem with judging combination safety
Two effective drugs don't automatically make an effective combination. If both carry an overlapping toxicity, combining them stacks that risk and the regimen becomes intolerable long before it becomes more effective, which is exactly why safety testing often restricts which combinations ever reach efficacy studies.
There's a data gap that makes this hard: most side-effect databases only contain adverse events for single drugs, not for combinations, so the overlapping-toxicity risk is invisible until the clinic. Reading every relevant trial for both response rates and adverse-event profiles, then reasoning about which toxicities compound, is the slow manual synthesis that gates combination strategy.
What teams in this space search for
- How do I know if two cancer drugs can be combined safely?
- How do you predict overlapping toxicity in drug combinations?
- Which drug combinations stack side effects?
How we built it
A custom project for a partner, built on the PharosBio Clinical Trial Retriever API. It fetches clinical endpoints (PFS, hazard ratios, adverse events) straight from trial records and scores drug similarity by target and mechanism of action. Here it paired the top-30 interventions and scored each pair on efficacy and combined adverse-event compatibility - with the reasoning kept transparent.
How the pipeline works
We pulled the breast-cancer clinical-trial cohorts, set the comparator to progression-free survival (PFS, in months), and extracted the relevant studies. Each intervention's efficacy was measured by hazard ratio, ranked, and the top 30 most efficacious were carried forward.
The pipeline then formed intervention pairs using cohort variables - complementary or synergistic mechanisms, non-redundant targets - and evaluated each pair's safety: would the adverse events reinforce each other, or stay manageable together? (Here we treat overlapping side effects as a less tolerable profile; the rule is tunable.) ADC receptor compatibility was estimated from TCGA subtype receptor-expression profiling.
Every pair was scored on efficacy and a combined adverse-event score, with the model's full reasoning attached - so the output is transparent and interpretable, right down to the raw event counts.
What it found
The scoring lands where the clinic does. A negative control, olaparib + talazoparib, scores low: both cause serious overlapping haematologic toxicity (anaemia, neutropenia), so combining them risks additive bone-marrow suppression.
A positive ADC pairing - sacituzumab govitecan + pembrolizumab in TNBC - scores well: overlapping but distinct primary toxicity profiles, manageable with dose modification. A novel pairing, trastuzumab emtansine + atezolizumab, also scores well, because atezolizumab's immune-related effects are distinct from T-DM1's haematologic and hepatic toxicities.
What we learned
Efficacy is only half the question - scoring has to read adverse events, not just response rates, and reason about combinations even when most databases only describe single drugs. Keeping the reasoning and raw counts visible is what makes a safety call defensible.
This was a custom build for a partner, but the engine underneath - the PharosBio Clinical Trial Retriever API - is reusable. The same endpoint that scored combination safety here also powers portfolio monitoring, in-vivo-to-clinical extrapolation, and repurposing, and it is available as a skill in Hydra.
Every pairing of the top-30 most efficacious interventions, scored for adverse-event compatibility per breast-cancer subtype. Switch subtypes and hover a cell to see the pair and its score - greener is more manageable together, redder is overlapping toxicity.
Worked examples - with the model’s reasoning
A negative control, a positive ADC pairing, and a novel ADC pairing - each transparent down to the raw event counts. (Illustrative: these specific agents are not all ideal conjugates; they show how the pipeline reasons.)
Olaparib + Talazoparib
Both agents cause serious overlapping haematologic toxicities - anaemia, neutropenia, infections - creating high risk of additive bone-marrow suppression.
Raw adverse-event data
- Anaemia - Olaparib 15/911 (17 events) vs placebo 1/904 (1)
- Neutrophil count decreased - Olaparib 148/911 (213) vs placebo 59/904 (82)
- Anaemia - Talazoparib 153/286 vs physician's choice 24/126
- WBC count decreased - Talazoparib 32/286 vs physician's choice 5/126
Sacituzumab govitecan + Pembrolizumab · TNBC
Overlapping haematologic (neutropenia, anaemia) and GI effects (nausea, diarrhoea), but distinct primary toxicity profiles - manageable in combination with dose modifications.
Raw adverse-event data
- Diarrhoea - Pembrolizumab + chemo 3/562 (4 events) ; sacituzumab govitecan 9/258
- Pneumonitis - Pembrolizumab + chemo 6/562 (8, 1%) ; sacituzumab govitecan 1/258 (0.4%)
Trastuzumab emtansine + Atezolizumab
Good compatibility - atezolizumab's immune-related effects are distinct from T-DM1's haematologic and hepatic toxicities. Both manageable in combination.
Raw adverse-event data
- Rash - T-DM1 + atezolizumab 2/47 (2) ; T-DM1 + placebo 3/49 (4) ; atezolizumab + chemo 33/293 (49)
- ALT increased - T-DM1 + atezolizumab 23/47 (52) ; T-DM1 + placebo 15/49 (25) ; atezolizumab + chemo 95/293 (160)
One API, many workflows
This was a custom build for a partner, powered by the PharosBio Clinical Trial Retriever API - which fetches clinical endpoints and scores drug similarity by target and mechanism of action. The same API can be repointed at:
- Safety scoring for combinations
- Extrapolating clinical outcome from in vivo data
- Portfolio monitoring
- Drug repurposing
It is also available as a skill inside Hydra
What you get
- Promising, low-overlap combinations flagged and scored, with the reasoning shown
- Additive-toxicity pairings (e.g. olaparib + talazoparib) ruled out before the clinic
- Efficacy and adverse-event compatibility weighed together, per breast-cancer subtype
- A transparent, tunable scoring framework - auditable down to raw event counts
Data sources used
- ClinicalTrials.gov (trial design & populations)
- Published trial outcomes (efficacy & adverse events)
- cBioPortal / TCGA (subtype context)
Figures reflect analyses PharosBio ran on public datasets and public benchmarks. Named competitors, collaborators, and logos are withheld at this stage; the methods and results shown are real and repointable to your own target.
Sources & methods
- Trial efficacy & adverse-event data: ClinicalTrials.gov; published trial outcomes
- Subtype context, cBioPortal / TCGA
Frequently asked questions
Why can't I just use a side-effect database?
Most adverse-event databases only list side effects for single drugs, not combinations, so the overlapping-toxicity risk that actually sinks a regimen isn't captured. The scoring has to extract per-agent profiles and reason about which toxicities compound.
Has the scoring ever matched real outcomes?
Yes. The negative control (olaparib + talazoparib) scores low for shared bone-marrow toxicity, while the positive pairing (sacituzumab govitecan + pembrolizumab in TNBC) scores well - both consistent with the clinical record. The example agents are illustrative of the method rather than ideal conjugates.
Was this built with Hydra?
No - this was a custom project for a partner, built on the PharosBio Clinical Trial Retriever API. That same API is available as a skill in Hydra, so the workflow can be reproduced there and repointed at monitoring, repurposing, or in-vivo-to-clinical extrapolation.
Run this analysis on your question
Hydra plans, executes, and validates, so you reach a defensible answer in hours, not weeks.