Teleoplexis · KNBi, Poznań University of Technology, Poland

Computational drug discovery, from docking to biosimilar similarity

A drug discovery lab in Poznań, Poland. We develop PROTO-NOOS, an open-source pipeline that generates small molecules de novo, docks and co-folds them with the target, simulates intracellular target engagement and ranks the result with a decomposable score. We are also building probe-based pharmacophore mapping of pockets and whole proteins, and similarity methods for biosimilar antibodies.

EcDHFR · PDB 6XG5 · trimethoprim · NADPH EcDHFR · PDB 6XG5 · illustration: probe clusters map the pocket, independent of the bound compound Illustration: human insulin B-chain against a lispro-type analogue

Work

Structure first

Docking · Boltz-2

Docking and co-folding

AutoDock Vina scoring, Boltz-2 co-folding.

PROTO-NOOS

Seven-stage pipeline

REINVENT4 generation to a decomposable score.

EcDHFR campaign

3.8 million generated

3.79Mgenerated
61,183scored
1,003measured

In development

Probe-based pharmacophore mapping

Small probes map the pharmacophore of a pocket or a whole protein. The map belongs to the target, not to one compound.

In development

Similarity of biosimilar antibodies to their reference

The sequence is fixed by definition. We predict where a different cell line and process introduce differences in the full antibody, and whether those differences matter functionally.

  1. ReferenceDossier of the originator product
  2. Structure3D model, antibody numbering
  3. AttributesWhich quality attributes are critical
  4. VariantsGlycans and chemical liabilities from process conditions
  5. DecisionSimilar or not, against quality ranges

Predictions are checked against measurements on a small panel of samples.

Audit

13

findings against our own results.

Open source

Code on GitHub

Research

Three posters

Evidence · EcDHFR campaign

Measured results

Ranking quality by endpoint

ROC AUC
Score v20
Affinity only0
ECFP4 + XGBoost0
Random ranking0

Enzymatic n = 1,003, 10-fold GroupKFold. Whole-cell n = 2,116, 67 actives. Full results

Whole-cell endpoint

0.483

ROC AUC of score v2. Random ranking: 0.554. Affinity alone: 0.653.

Compute

62%

of a 50-molecule run is Boltz-2; dynamics adds about 33%.

Seconds per molecule

Campaign scale

3,794,691 retained pre-selection samples 602 REINVENT4 batches
61,183 scored and ranked records after removing the secondary KD gate
1,003 records with a measurement the calibration cohort
15 top de novo compounds inspected 9 share a carboxylic acid motif

Target engagement

Cell target engagement scoring: distribution of pre-Boltz-2 cellTE scores, rank before versus after Boltz-2 K_D rescoring, and the top ten candidates
cellTE score before Boltz-2, rank change after KD rescoring, top ten.

Stage agreement

Heatmap of Spearman correlations between the scores of the pipeline stages
Spearman correlation between stage scores. Stage 6b and KD: ρ = 1.00.

Dynamics

Residue fluctuation along EcDHFR for trimethoprim and a ChEMBL analogue with the M20 loop highlighted
Residue fluctuation, M20 loop highlighted.

Library

85.4%
singleton scaffolds
91.8%
below Tanimoto 0.4 to known actives

Collaboration

What we need

Compute

GPU time

Boltz-2 and molecular dynamics are about 95% of a run.

Validation

Wet lab

Synthesis, enzymatic and whole-cell assays.

Review

Expert eyes

Dynamics, flux balance and experimental design.

Partnerships

Collaborate with us

Join

Build it with us