You are here: Technology >> Positioning


Where Equinox Pharma has positioned itself


Technology and methodology behind INDDEx™


Advantages of using INDDEx™ over other technologies


Results of INDDEx™ application to various targets


Publications relating to Equinox Pharma technologies

INDDEx™ advantages

High predictive accuracy

INDDEx™ gives a 50% prediction rate

Our assessment of INDDEx™ in the Journal of Physical Chemistry B shows high enrichment factors and retrieval rates when tested on the Database of Useful Decoys benchmarking dataset, even for target datasets with high structural diversity.

Our internal validation has experimentally confirmed results on three targets in vitro, that resulted in identifying several high-quality compound hits that would be the basis for hit-to-lead and lead optimisation in each case. Our proof of concept shows that for receptor-based antagonists we can achieve a 50% rate of prediction of in vitro hits using INDDEx™. This compares very favourably with the low hit rate of high throughput screening (which is much more expensive and can only involve much smaller numbers of compounds), and with competing in silico approaches which provide between a 5–15% hit rate. We have also confirmed similar, excellent, results for many pharmaceutical and agrochemical companies.

Line graph of an enrichment curve
Strong enrichment: This graph shows the average enrichment factor (EF) achieved on all 40 targets of Database of Useful Decoys benchmarking dataset. EFs are calculated by dividing the fraction of active compounds found in the sample by the fraction of active compounds in the entire population. Here INDDEx™ is learning on 32 randomly selected ligands.

Scaffold hopping

75% of hits are new chemotypes

The strength of Equinox’s INDDEx™ technology is in the structural logic-based rules it learns. The number of rules depends on the number of molecules in the training dataset as well as the diversity of molecules in this set. The rules are generated on fragments of the overall molecular structure rather than the structure itself. This allows INDDEx™ to identify new chemotypes in which these fragments lie. This contrasts with the rules generated by many other systems that are based upon the complete structure and therefore tend to identify similar structures. We have shown that up to 75% of hits represent new chemotypes. Conventionally medicinal chemists consider a compound to represent a new chemotype is it has a Tanimoto coefficient (a measure of structural similarity) of greater than 70%. This shows the power of INDDEx™ technology for scaffold-hopping.

Our assessment of INDDEx™ in the Journal of Physical Chemistry B shows that enrichment factors remain high even when only measuring enrichment of molecules structurally dissimilar to any molecules in the training data.

Line graph of an enrichment curve for molecules that represent scaffold-hopping challenges
Strong enrichment for scaffold-hopping challenges: This graph shows the average enrichment factor (EF) achieved on all 40 targets of Database of Useful Decoys benchmarking dataset, but only measuring enrichment of molecules that have a less than 0.5 Tanimoto Maximum Common Substructure Similarity to any molecules used in the training data. EFs are calculated by dividing the fraction of active compounds found in the sample by the fraction of active compounds in the entire population. Here INDDEx™ is learning on only 8 randomly selected ligands.

Unique and original technology

Patented proprietary INDDEx™ method

INDDEx™ is based around the patented SV-ILP (Support Vector Inductive Logic Programming) method (Patent number 8126823). SV-ILP has been developed by Equinox Pharma staff and is proprietary to the company. INDDEx™ also incorporates a variety of in-house methods to get optimum results from your data.

Biochemist examining molecular structure

Cost effective hit finding

The clear advantage of using an in silico approach such as INDDEx™ is that large numbers of compounds can be screened rapidly and cheaply; for example, we are able to screen, identify, and rank the approximately 38 million compounds in the ZINC 12 database in just a few hours.

Rules can be easily understood by chemists

Understand activity at a glance with simple graphical representations and descriptions

Logic-based rules can be easily understood. Standard machine-learning approaches, such as support-vector machines and partial least squares, attempt to generate an equation incorporating the molecular descriptors that they are given. These rules make it difficult to understand the mechanism that governs the activity of the molecules on the targets. INDDEx™ uses a machine-learning engine which defines activity by forming logic-based rules in the form of pairwise distance relationships between chemically-relevant fragments of molecular structure. This makes it possible to identify the parts of the molecular structure that have an effect on activity, and gives guidelines for medicinal chemists to understand the mechanism of activity, and develop promising leads.

Rules from standard programs Activity = 0.45 LogP + 0.56667 LUMO + 1.65 V
Rules from the ILP method In an active molecule: Fragment A is from fragment B, which is bonded to fragment C

Two examples of the sort of rules generated by the ILP method are shown below. The structural diagrams overlayed with graphics are typical INDDEx™ outputs.

active(A):- C.ar_2ndLayer_C.ar_C.ar_C.ar_N.ar_H(A, B), Charge_0.0(A, C), distance(A, B, C, 5.7, 1.0).

Molecule is active if there is a nitrogen-containing aromatic ring system and a neutral charge centre (in this case a methyl group on the benzene ring) 5.7 ± 1.0Å apart. Note that the first fragment could also match various fused aromatic ring systems.
active(A):- positive(A, B), Nsp2(A, C), distance(A, B, C, 9.1, 1.0).

Molecule is active if there are two aromatic rings 9.1 ± 1.0Å apart.

Application in hit-to-lead and lead optimisation

INDDEx™ output rule data can guide the optimisation of compounds

The Quantitative Structure–Activity Relationship (QSAR) rules can be used to guide medicinal chemists in the hit-to-lead and lead optimisation programme by allowing ‘what if?’ questions to be posed regarding potential chemical substitutions and other structural modifications. Whilst this would not replace the need for synthetic medicinal chemistry, it will make the process both more efficient and more productive. Other rules based upon key ‘druggability’ properties such as solubility, metabolism, and distribution can be derived. These will also be used to guide lead selection and optimisation.

Protein with a ligand docked into its active site
Confirmation of predictions: A ligand predicted by INDDEx™ from training data is docked into a known active site. Being a ligand-based method, INDDEx™; only uses information about the structure of small drug molecules, so knowing the structure of the protein to be targeted is not necessary for prediction. However, as shown here, docking methods can be used to confirm the potential of INDDEx™ predictions to act as strongly binding ligands.

INDDEx™ does not require target structure data

No knowledge of crystal structures or binding sites is needed

INDDEx™ is a purely ligand-based virtural screening method, and can be applied to all drug targets where a Structure-Activity Relationship exists, and does not require knowledge of the 3D structure of the target. This is particularly relevant to GPCR’s where target-structure data is very limited.

INDDEx™ has other important advantages

Multiple features allow a broad spectrum of applications and in-depth analysis of data

  • Learns from both active and inactive compounds
  • Can manage large number of active compounds
  • Can manage diverse set of active compounds
  • Flexibility in defining features to generate the predictive Quantitative Structure–Activity Relationship (QSAR)
  • Does not require a global superposition
  • Can automatically identify more than one binding mode
  • Can create QSARs for compounds with polypharmacological activity
  • The strategy can readily be applied to a broad range of chemoinformatics problems including incorporating knowledge of the receptor