PenicillinX

About

PenicillinX predicts which beta-lactam antibiotics are likely to cross-react, by measuring how chemically similar their side chains are. It is a companion to a chemoinformatics manuscript, and displays precomputed results — no analysis runs in your browser.

Why side chains

Every drug here shares the same beta-lactam core, so the core cannot explain why a patient reacts to one drug and tolerates another. Hypersensitivity is instead largely driven by the side chains hanging off that core: the R1 group, and for cephalosporins the R2 group. Each molecule is drawn with its beta-lactam ring shaded blue so you can see the shared anatomy, and matched side chain atoms circled in red (R1) or blue (R2).

Using the site

Compare answers "my patient reacted to these drugs — what do they share?". Pick up to three antibiotics and see them drawn side by side, with the prediction for every pair underneath.

Matrix shows all 51 antibiotics against each other at once. Cell colour is the published ground truth from Hutten et al. (2025); the dot on top is this model's prediction. The interesting cells are where the two disagree.

Matrix cell colours — Hutten et al. (2025)
	No cross-reactivity
	Cross-reactivity
	Discrepancy between source studies
	No data
	Same drug

Prediction dots — which side chain drove the match
	R1-driven
	R2-exclusive — R1 similarity fell below threshold
	R1 and R2 both fired independently

The R2-exclusive case is the novel finding of the accompanying work: pairs that look unrelated by their R1 groups, but share an R2 group closely enough to matter.

The ensemble

Seven algorithms vote, combined by logical OR — a pair is predicted to cross-react if any of them fires. Four compare whole-molecule fingerprints and return a single similarity score; three identify the specific atoms responsible and can therefore drive the highlighting.

Algorithm	Threshold	Highlights atoms
Direct Matching (DM)	substructure test	yes
Permissive SMARTS Direct Matching (PDM)	substructure test	yes
Maximum Common Substructure (MCS)	0.8998	yes
Morgan Dice (MD)	0.4798	no
Pharmacophore Dice	0.6714	no
MACCS Dice	0.7745	no
Atom Pairs Dice (AP)	0.5281	no

Thresholds were optimised by differential evolution against a specificity floor of 0.95. DM and PDM are unthresholded binary substructure tests.

Because the four fingerprint methods fire on aggregate similarity, they have no per-atom attribution. A pair can therefore report a high algorithm count with no atoms highlighted at all — the molecules are broadly similar without sharing one specific region. This is expected, not a fault.

Limitations

The model is unvalidated. It has not been tested prospectively, nor against patient outcomes.
Ground truth is itself imperfect. Of the 1,275 distinct drug pairs, 178 have no published data at all and a further 136 are marked as disagreements between the source studies Hutten et al. harmonised.
Chemical similarity is only one determinant of clinical cross-reactivity. Immunological and patient-specific factors are not modelled.
Predictions are symmetric; real-world reaction patterns need not be.

Results must not inform clinical decision making to any extent.

Source — AGPL v3.0