What does a junior modeller do?
You Tao
Source; “Robotic process automation isometric composition with little human characters” https://www.magnific.com/
By reading this post, you will understand what a junior modeller does.
This post is intended for PhD students who study computational modelling to support drug discovery and development projects. If you just started your first modeller role in the industry (e.g. PK/PD modeller, pharmacometrician, data scientist), this is also relevant for you.
This post is written from my industrial experience. Looking back to my days in big pharmas, these are the main things my R&D colleagues needed me to do. When these are done well, trust is built and it lasts for a lifetime.
The intention is to provide guidance for junior modeller development. Of course, we need to first talk about what junior modellers mainly do before we think about what success looks like and how to do well.
What does a junior modeller do?
Naturally, this answer comes to your mind: junior modellers are expected to create new knowledge through developing and validating computational models. In the world of pharmacometrics and systems pharmacology, here are the typical questions you are expected to answer well:
1. Explain the mechanisms of action for efficacy and safety
Articulate how target engagement is related to treatment efficacy. In the case of toxicity, explain how toxicokinetics at the site of action engenders toxicity (a continuous quantity or a discrete, ordinal number such as the grade of the adverse events). This is the bread and butter and indeed the absolute minimum you are expected to provide from your model-based analysis. The dose-response relationship for toxicity is often relatively easy to define. For efficacy, it may be relatively trickier sometimes. Modelling forces you to think through all the explicit and implicit assumptions to explain efficacy.
2. Enable translations: in vitro <=> in vivo <=> clinical
Crucially, any discovery project needs clinical relevance. If it hasn’t, you shouldn’t do it. It is obvious we need to ensure the in vitro and in vivo results translate into the clinics. I draw these three entities in a sequential order for the following reasons:
Firstly, an in vivo model may exhibit less heterogeneity than the clinical disease. For example, a cell line-derived mouse xenograft model exhibits less variability than cancer which is known for its heterogeneity within a patient and between patients. To ensure a good, robust understanding of these systems, a project should focus on the in vitro <=> in vivo translation first. Otherwise, you might find correlation between in vitro and clinical data, but they might be epiphenomenal (i.e. happen by chance) and hard to explain. If you cannot explain it, you are at risk.
Secondly, researchers often seek to “translate” preclinical in vivo modelling results into the clinics. When you select the first-in-human dose, you need to think about the species difference. The validity of allometric scaling is challenged and we are not going into this here.
Lastly but not least, in vivo models needs to be qualified with clinical data (efficacy and safety). For a fair comparison, you need to use modelling to adjust for the differences in pharmacokinetics under both scenarios, and evaluate how in vivo models may respond to the clinical drug pharmacokinetics.
3. Predict efficacy of a combination drug treatment
A new cancer drug almost inevitably has to be combined with existing drugs in a clinical trial. This is also true to infectious disease. You need to evaluate whether the combination adds additional benefits or reduces efficacy, whether it is safe or put patients into greater danger.
In the real world, multimorbid patients (i.e. individuals who live with two or more co-occuring chronic conditions) need different drugs at the same time. Their physiology is different from those subjects enrolled into the clinical trials during drug development. From the pharmacogenomics perspective, you need to identify variants playing a role in metabolising multiple drugs administered at the same time. You need to adjust dose for patients who would be exposed to adversely high drug concentrations.
4. Optimise dose scheduling
A combination therapy can be tricky. Concurrent dosing may be intolerable due to safety. Gapped dosing may lose efficacy. You need to work out the dosing sequence (which drug goes first), dosing frequency and the time gap between the two drugs to minimise toxicity and to ensure efficacy.
How? Pharmacokinetics-Pharmacodynamics (PK/PD) modelling is often used. You may use an empirical model to capture the action of the drugs on their own, and then model the combined effects by introducing an interaction term. This is relatively easy to do, given the simplicity of the empirical model. On the flip side, this approach does not capture the mechanisms. Therefore, if you change a combination partner, you cannot predict how the choice of the new drug may affect efficacy and safety. In other words, the empirical approach helps you recapitulate the data, but not to lend itself to ab initio predictions.
It is desirable to use mechanistic models to incorporate biomarkers relevant to the drug actions. The same as before, you may construct a model for each drug first. As the two models share common biomarkers for target engagement, physiological changes or disease modulation, the opportunity here is you may combine the two models together to simulate the combination treatment. You can test the model predictions for the combined treatment. If positive, this helps your project team build confidence in your modelling! Of course, this requires more data to characterise how targets of the two drugs interact with each other. However, such understanding (and the investment) is needed to validate the therapeutic hypothesis. Otherwise, how would you know the combined effects are generated through the conceptual ways you anticipated? Without this answer, do the preclinical in vivo results translate at all? Apparently, without validating this point, a project carries risks.
5. Select the right patients
This may be approached from two different angles. You may use PK/PD modelling and time-to-event analysis to identify who responded favourably after a clinical trial is complete. This involves testing whether the inclusion of a covariate may significantly improve the goodness of fit to the data. This delivers a clear cut answer, but the answer is limited by the candidate covariate you measure in the first place. It is tedious and time consuming, often taking months at the moment. AI may reduce the time it needs.
Another way is to start from the genome-scale high throughput data for each patient. Use regression-based modelling methods (e.g. LASSO or ElasticNet) or a classification algorithm (e.g. decision trees, random forest, XGBoost) to identify patterns that are correlated with an outcome. However, this approach only generates educated hypotheses which need experimental proof. Test of the hypotheses (dubbed “signatures”) using data from an independent trial frequently fails.
However, an obvious task is often overlooked:
6. Impart knowledge (e.g. develop others).
At university, a PhD student should mentor and help undergraduate and master’s students. In a business R&D, a junior modeller needs to develop work relationship with different departments, and educate collaborators what are the right modelling question to ask when, and how to prioritise these tasks to come up with the shortest path to success. We all need to be an educator in our work and junior modellers are no exception. You R&D colleagues will be genuinely thankful to you if they gain a better understanding of how to work with a modeller. To them, you are a massive extension of their ability to analyse the data.
Modellers tend to be introverts: You appreciate lone time you need to think out a plan. You take pride in perfecting the execution of a plan. You are frustrated if others are not fully onboard the first time you “explain” (“tell” might be more accurate sometimes) to them your plan. To work well with a team, you need to see things from the perspective or perception of your teammates. Be interested in the people, as well as the science. Know their deadlines, and offer help to make others’ lives easier.
These are my thoughts of what junior modellers should do. Share your thoughts below!
