Polypharmacology

One of Alosenta’s key capabilities is the identification and prioritisation of promising drug pairs and triplets.

The proprietary AI/ML workflow evaluates potential combinations using an integrated, multi-layered approach. Rather than simply identifying drugs that may work against the same disease, the platform assesses the relationship between their mechanisms, potential target overlap, toxicity and complementarity.

The objective is to identify combinations with the greatest potential for synergy and selectivity, while reducing unnecessary overlap and cumulative toxicity. Structure-based high-throughput virtual screening is combined with off-target prediction and experimental evidence to provide a more systematic approach to prioritising drug combinations for further investigation.

Alosenta benchmarked its drug-combination model on unseen data, achieving:

• 84.4% R@5 — the correct combination was identified within the top five predictions in 84.4% of cases.
• 96.4% R@20 — the correct combination was identified within the top 20 predictions in 96.4% of cases.

The platform is designed not only to rank combinations, but also to provide the chemical and biological rationale behind its predictions. This is central to Alosenta’s approach to explainable AI: results are designed to be traceable to their underlying evidence rather than presented as unexplained scores.

The number of possible drug combinations grows rapidly as the number of candidate molecules increases. Alosenta’s approach uses multiple computational layers to narrow this search space, identifying combinations with complementary mechanisms while filtering for potential toxicity and off-target effects.

This creates a more systematic route from a large universe of possible combinations to a prioritised shortlist for experimental validation.

The platform’s combination capability has also been tested in specific disease settings, including multiple myeloma and EGFR+ lung cancer, where it independently recovered established clinical treatment combinations while also identifying novel combinations for further investigation.

Could AI rediscover decades of clinical knowledge—without being given any?

For this study, the platform was provided with just 16 myeloma drugs as chemical structures and a single disease code. It was not given treatment protocols, scientific literature or information about existing clinical combinations.

The platform evaluated 120 drug pairs and 286 drug triplets, generating ranked recommendations accompanied by biological rationale.

The Result

Among the highest-ranked combinations, the platform independently identified established multiple myeloma regimens, including:

• VTd — Dexamethasone + Bortezomib + Thalidomide
• VRd — Dexamethasone + Bortezomib + Lenalidomide
• Other clinically established combinations

These regimens emerged near the top of the ranked list despite the platform having no prior knowledge of current clinical practice.

Identifying what should not be combined

The platform also rejected 218 of 406 combinations, including combinations where drugs shared the same disease proteins and offered limited mechanistic complementarity.

For example:

• Lenalidomide + Pomalidomide shared key targets including CRBN, DDB1 and IL6, alongside 15 overlapping toxicity endpoints.
• Bortezomib + Carfilzomib showed substantial overlap across proteasome targets.

Beyond established treatments

The highest-ranked list also included novel combinations, such as Dexamethasone + Bortezomib + Iberdomide, highlighting opportunities for further investigation.

Every recommendation was accompanied by an explanation of the underlying biological mechanisms, rather than simply presenting a black-box score.

Identifying the right therapeutic architecture

The second case study explored whether the platform could identify meaningful combination strategies in lung cancer.

Using 12 drug structures, the platform evaluated 66 drug pairs and 84 drug triplets. It was not provided with treatment protocols or information about established therapeutic strategies.

The Result

The platform independently converged on the same fundamental therapeutic architecture used in clinical practice: combining a targeted therapy with a chemotherapy agent, rather than combining multiple targeted inhibitors with overlapping mechanisms.

The top-ranked combinations included:

1. Gefitinib + Docetaxel
2. Pemetrexed + Docetaxel
3. Crizotinib + Docetaxel
4. Imatinib + Docetaxel
5. Erlotinib + Docetaxel

The top-ranked triplet—Docetaxel + Pemetrexed + Gefitinib—engaged three distinct mechanisms involving Tubulin, Thymidylate Synthase and EGFR.

Knowing what to reject

The platform rejected 44 of 66 drug pairs, including all combinations involving two targeted inhibitors with overlapping target engagement.

For example, Osimertinib + Afatinib showed overlapping targets and 19 overlapping toxicity endpoints, offering limited mechanistic complementarity while potentially increasing toxicity burden.

The platform evaluated 120 drug pairs and 286 drug triplets, generating ranked recommendations accompanied by biological rationale.