Date

23 Mar 2018

Attendees

Richard Norman (Unlicensed) (Pistoia Alliance)
Qing Chai (Deactivated) (Eli Lilly)
Craig Dickinson (Unlicensed) (Eli Lilly)
Guy Georges (Unlicensed) (Roche)
Francisco Fernandez (Deactivated) (Abvance Biotech)
David Hargreaves (AstraZeneca)
Helen Li (Unlicensed) (Eli Lilly)
Kieran Todd (Deactivated) (GSK)

Discussion items

Time

Item

Who

Notes

15:00 - 16:00 BST (60 mins)

>What internal efforts are ongoing in this space? (please can you find out or invite a relevant person)

>Agree the challenge(s): right data, right format, right quantity

>Solutions and approach

Richard, All

>Agree our approach to making a difference in the developability risk prediction space.

Minutes

Introduction – current position

Link to slides: https://drive.google.com/open?id=17kfs55dvCZVdX647QweIS2tsTd3lMcPP

Key points from discussion

Terminology:

Developability data/dataset = biophysical data and data capture methodology which the project aims to generate and is associated with a particular developability liability. Part of the scope should be to normalise and collate this data across the industry. It would be important to define the common rules and different weights for each (cf. Lipinski rule of 5 for SM) and consider the line of sight to final endpoint of developability risk readout.
Structure Descriptors = derived from Ab primary sequence and 3D structure and could act as surrogate for such. Including but not limited to; surface properties, global or local pI, charge and hydrophobic properties, dipole moments, etc. Lilly has developed such a set and are still developing them (through a combination of internal and external algorithms).
Experimental descriptors = Common (published) surrogate assays which provide intermediate Endpoint PK, aggregation, immunogenicity data.
PK data/dataset = monkey (cyno) and especially human exposure (AUC) unless otherwise stated.
Methods development = in silico model development based on developability data/dataset.

>Developability data for each Ab would ideally cover: sequence, 3D structure, biophysical data (esp. for viscosity and solubility) and immunogenicity. This would be for Abs with both good and poor PK. PK data prediction remains the holy grail.

>Perceived likely resistance from organisations to share sequence and 3D structure data for pre-clinical candidates.

>Alternative is to use structural descriptors although some scepticism as link to creating 3D models of Abs will be lost and will add an extra layer of complexity in terms of defining the right quality and quantity. Lilly use structural descriptors (still in development) to correlate to developability risk. Roche don’t use structural descriptors per se but look out for ‘easy win’ pointers to potential liabilities, e.g. glycosides or presence of Cys in CDR loops. The caveat is that these don’t always translate to risks. GSK don’t use structural descriptors.

Which data/format?

>Agreement that Abs should be the focus of the exercise i.e Ab-like molecules (bi-specifics, etc.) are out of scope.

>Either start from scratch or should likely focus on public (clinical) data to build an initial model with subsequent generation of novel data based on feedback/analysis.

>Defining the data standard and format could be a quick win. Post-meeting note: consider data FAIR principles.

>Need to consider data storage and access as this is key to facilitate mining.

Which Approach?

Some of these ideas/proposals are not mutually exclusive:

>Agree on a set of common data (share internal data if appropriate) and get a vendor to generate some additional data to fill key gaps.

>Collate public data from Abs into a dataset and conduct methods development to build a model. Share the model with industry which will test this internally on sequences/structures of interest. The outcome of testing is shared to build a better model and will not contain sensitive sequence/structure information.

>A variant of the above is to share proprietary data from Abs as well as PK data to understand if a correlation can be drawn. Issues might be that biophysical assay data is not comparable across companies and absence of sequence/3D structure would not allow for validation. Lilly could share internal assay protocol and ‘reference’ controls to standardise the assays and/or how they are different from e.g. Adimab paper assays. Roche conducted a validation exercise with Adimab by sharing molecules and internal assays to understand if there were differences in the way Adimab were running the assays.

>Focus on experimental predictors: use these to build models which would then allow you to predict final endpoint data and could be used to trace back to e.g. structural descriptors.

>Find common or agree on publically available clinical Ab molecules which have a variety of behaviours to normalise the experimental descriptors across the industry. This idea is a build on what was done by Adimab (Jain et al., 2017) but goes beyond the assays used and would include PK data. The added value would be that currently there are not many molecules with associated PK/viscosity/solubility data. Adimab paper use failure in clinic as a measure of lack of success but this may have been due to other factors e.g.PK/viscosity/solubility data.

Some outstanding questions

Q: what internal efforts are ongoing in this space?

Q: Can we validate how well the structure descriptors describe the sequence/3D structure?

Q: What would we not be able to do with confidence if using structure descriptors instead of sequence/3D structure data?

Q: Have we got the right people from each organisation?

----------

Richard Norman – 23 March 2018

Browser not supported