Eric J Ma's Website

Today, I had a chance to read the article, 4 Skills the Next Generation of Data Scientists Needs to Develop. (h/t Eric Yang for sharing this one on LinkedIn.) This article resonated massively with me, and I wanted to share how some of the ideas in here translate over to the DSAI Research team at Moderna.

1. Problem Spotting: Seeing the real issue

The first skill is distinguishing the real issues faced by our collaborators from the apparent problems. The real issues might be different from the issues they think they have.

Within biotech research, scientists come with differing levels of savviness with machine learning methods and the capabilities they bring to laboratory science. As such, our collaborators may come to us with requests for help that may need refining before they match what they need.

One example that I can recall was related to the hype of ChatGPT. A news and perspectives article in Nature wrote that

Language models similar to those behind ChatGPT have been used to improve antibody therapies against COVID-19, Ebola and other viruses.

And pretty soon, we were getting inquiries about whether we could use an internal implementation of GPT-4 to design protein sequences.

An inexperienced data scientist would say, "That sounds like a cool idea! We should try it out!"

But an experienced data scientist well-versed in the training of ML models and laboratory science would immediately see the flaws in that logic and instead ask what the real problems are -- why would one want to design proteins in the first place? Here are some examples of questions we might ask:

1. Is the protein you're trying to engineer particularly difficult to express, and is your goal to improve expression?
2. Is your protein chemically modified under manufacturing conditions, and do you need to re-design the protein to minimize chemical modifications while retaining function?
3. Is your protein degrading too quickly within the cell?

Notice how the questions we're asking aren't related to ChatGPT at all! Instead, they're all questions about the laboratory science itself. These questions are designed to address the lab scientist's fundamental scientific challenges. Doing so is paramount for our team to continue being valuable and innovative: uncovering the fundamental problems orients us to build Model Ts instead of breeding faster horses.

2. Problem Scoping: Gaining clarity and specificity

The second skill is asking probing questions that narrow down the space of problems and the possible solutions that our collaborators could use.

Let's continue the previous example by expanding on option #2, which is antibody-related.

Gaining clarity would involve asking follow-up questions that look like these:

1. What chemical modifications are you worried about?
2. How do you detect these chemical modifications? Is it mass spectrometry?
3. How is the mass spec data analyzed? What are the data transformations involved to get to the endpoint measurement?
4. How do you know your antibody continues binding to its target? What is the assay that you're using? Can you describe the experiment? What are the controls used?
5. What's your hypothesis behind using machine learning for your protein engineering problem? How will there be recurring benefits from our effort?
6. What's the experimental budget in terms of libraries that you can screen? Can you turn an arrayed screen into a pooled screen?

With answers to these and other questions of the same spirit, we can enhance our understanding of the problem space. We can also begin to see connections between what our collaborators need and what other collaborators may need and begin to design entire systems that can benefit more than one group at a time.

3. Problem Shepherding: Getting updates, gathering feedback

As mentioned in the article, it's tempting for a data scientist to dive into the problem for an extended period and come back up with a solution that may impress a collaborator.

Within a biotech context, this is junior-level thinking -- and an easy way to lose the trust of sharp and well-trained PhD-level scientists and leaders. Universities train PhDs to question everything they don't understand. No assumption will be left unturned.

A junior-thinking data scientist who keeps themselves disconnected to do a "big reveal" will likely not understand the questions the laboratory scientists will ask. They may build something they don't understand and are unwilling to adopt and use.

By contrast, the senior-thinking data scientist who makes the concerted effort to jointly co-create the solution with the laboratory scientist will cultivate trust in the solution, build their vocabulary within the scientific domain, and construct a library of misconceptions and assumptions they can anticipate and address in future builds.

4. Solution Translating: Speaking in the language of the audience

Following the original article, speaking the audience's language is paramount. As you might have guessed from the previous points, when developing solutions for laboratory science, we need to speak the language of laboratory science throughout the entire conversation. It's insufficient to stop at general questions that sound like:

• What's the problem that you're facing?
• What are possible solutions that you've tried to implement?

By contrast, the questioning strategy that garners trust involves progressively asking more and more specific questions. Notice the specificity level of the questions above in section 2. In my books, those are only the basic questions; our questions are best when they show that we truly understand the underlying biology, chemistry, immunology, analytical chemistry, organic chemistry, or other theory that backs a problem. When we also speak the vocabulary of the laboratory methods, shown for example, by actively questioning how experiments are designed and possessing an order of magnitude intuition about assay throughputs, that trust with laboratory scientists is solidified.