Table 1 Illustrative quotes

Data gaps

(1) The prediction and generation tasks we’ve outlined are data hungry… where are there mechanisms to gather quality training data?

Dataset coordination/data-sharing

(1) We need training datasets that merge chemical structure with biological and clinical outcomes, AI can’t learn robustly without that.

Data integration issues

(4) Public datasets are heterogeneous—ML models struggle because data from different labs aren’t standardised.

Conveying uncertainty

(3) ML models often give confident answers without conveying uncertainty… we need systems that know when they don’t know.

Possible solutions

(5) There are companies trying federated AI approaches… letting pharma share data without revealing proprietary info.

High-throughput screening

(4) With computers, we can now discover hundreds of thousands of antibiotic candidates in hours.

Combination therapies

(1) AI for predicting synergistic chemical combinations is ideal—but we lack the training data to do it reliably.

Drug repurposing

(1) If we train AI models across pharmacological properties, we can reuse what we’ve already done in other therapeutic areas.

Host factors

(7) AI could help predict which antibiotics will work across diverse human genotypes—but we need much better host-response data.

Vaccines

(3) Using smart statistical models to design vaccines is part of how AI could help us target high-risk resistant bugs.