Old medicines, new questions.
Multi-drug resistant superbugs are evolving faster than our ability to create treatments, causing 1.27 million deaths globally every year.1Source: The Lancet, 2022 This model asks a faster question: could a medicine that is already approved for something else also act against a drug-resistant superbug?
Broader molecular dataset
Source · ChEMBL
20K
Molecular structures
Extracted from ChEMBL, an open database of bioactive molecules
- other compounds with laboratory records against the four pathogens
- 17,878
- approved medicines in the FDA screening library
- 1,761
- other ChEMBL drug entries outside that library
- 619
20,258 valid structures, each counted once. The 1,761 library medicines include 157 that also have laboratory records, so 18,035 structures have records in all. 138 that could not be read were dropped, not repaired.
Approved-medicine library
1,761approved medicines checked
1,019
repurposing candidates: not already antimicrobials, with AI-predicted activity ≥40% against at least one pathogen
39,548registered clinical studies linked to them
One capsule ≈ 100 medicines. Rounded.
What is drug repurposing?
Finding a new use for a medicine that is already approved. Its safety in people is already known, so testing a new use can be faster and cheaper than starting from nothing.
Aspirin
First used forPain and fever
Later also used forPreventing heart attacks and strokes
Sildenafil
First used forChest pain (angina)
Later also used forErectile dysfunction
Halicin
First studied forDiabetes, as the preclinical compound SU3327
Identified by deep learningAntimicrobial activity (2020)
Not an approved medicine; it was preclinical when identified.
How this site works
From molecule to evidence: five steps the project ran, and the one it did not.
01
Collect
Approved medicines + laboratory activity records
Lab measurement
1,761approved medicines · 35,232 labelled lab records
02
Represent
Convert molecular structures into machine-readable patterns
Molecular fingerprint
20,258valid molecular structures, each stored as a fingerprint
03
Predict
The model estimates activity from the project's laboratory activity data
AI-predicted activity
7,044predictions; 1,276 medicines reach ≥40% for at least one pathogen
Prediction04
Check fit
Dock every medicine against a protein from each pathogen
Computer prediction
3,533 / 7,044docking jobs completed (1,761 medicines × 4 targets)in progress
Structural hypothesis05
Check evidence
Search registered clinical studies and other evidence
Clinical record
1,761 / 1,761medicines checked for registered studies
Existing evidence06
Laboratory testing
Not performed in this project
No new patient or laboratory experiments were performed.
Four pathogens. Four different resistance problems.
Only these four can show a percentage. Any other condition shows documented evidence only.
MRSA
Staphylococcus aureus · Gram-positive
Resistance example
Altered PBP2a target
MRSA carries an altered target, PBP2a, so beta-lactam antimicrobials bind it poorly. The thick wall of a Gram-positive cell surrounds it.
- Repurposing candidates ≥40%
- 671
- Docking target used
- Dihydrofolate reductase (PDB 3FRE)
The models in this project provide AI-predicted activity only for these four species. They are not predictions for every pathogen species or every resistant strain: the models learn patterns from previous laboratory measurements, and they do not directly simulate a patient’s response. Diagrams are schematic, not to scale.
Start investigating
Search a condition to see the medicines already documented for it and the other medicines the model surfaces, or search a medicine to see all four of its predictions.