How AI Experts Can Drive Better Collaboration in Drug Discovery

Connect with us: Valence is constantly seeking talented individuals with diverse backgrounds and expertise to join our team. Explore open roles here.

Imagine a drug discovery team asks their ML group to build a generative model for new molecules. The medicinal chemists hand over a spec: molecules must have logP < 5, molecular weight < 500, no more than 10 hydrogen bond acceptors, and should maximize predicted binding affinity to a target protein. The ML team faithfully builds it and the model is good: it generates thousands of novel molecules that satisfy every constraint and have high predicted binding affinity.

The team presents their results. The chemists look through the output and are... underwhelmed. The molecules are clustered along the constraint boundaries: logP of 4.9, molecular weight of 498. They’re technically compliant but frankly unimaginative. The chemists had hoped for a bit more diversity (e.g. some molecules with logP of 2 or 3), not just solutions that live right against the boundary.

What happened here? The constraints were never really constraints. They were rough preferences, borrowed from textbook heuristics. By encoding them as hard walls, the model explored a much narrower chemical space than the chemists actually wanted. Nobody was wrong: the chemists stated their best understanding, the ML team implemented it faithfully. That is what makes this failure mode dangerous: faithful implementation can create false confidence, because everyone feels aligned around the written spec even when the written spec is only a lossy proxy for the real goal. The problem was in the translation from what the chemists wanted to what the model optimized.

This kind of failure is, in my experience, the most common way interdisciplinary AI projects go wrong. Not bad models, not insufficient data, but a mismatch between what domain experts want and what the AI system is actually set up to do. Between “what we want” and the AI’s “objective” sits a translation step where things can break in non-obvious ways.

This pattern is not at all unique to chemistry. A clinician might ask for a diagnostic model that maximizes sensitivity, when what they really need is a triage system with calibrated uncertainty and a safe escalation path. An operations team might ask for a model that minimizes average delivery time, when what they actually care about is avoiding rare but catastrophic delays. In each case, the stated metric is not wrong; it is a lossy proxy for the real goal.

Interdisciplinary collaboration is hard. Different vocabularies, priorities, and responsibilities means it’s normal when collaboration doesn’t feel entirely seamless.

Problem formulation- which I define as setting objective functions, evaluation metrics, loss functions, etc. - sits in a kind of “no man’s land” in the middle between AI experts and domain experts. On one hand, it is a very consequential decision that genuinely requires expertise and knowledge from both sides. On the other hand, extended discussions of problem formulation usually just feel exhausting. When discussing “which of these 10 objective functions should we use”, domain experts can often feel like this is an implementation detail (and therefore up to the AI experts to decide), while AI experts often feel like it’s part of the problem definition (and therefore up to the domain experts to decide).

Add on the pressure that everybody generally feels busy, and there is a strong temptation to “just pick something” and fix it if it doesn’t work. Everybody gets to walk out of the room and get back to work. Unfortunately, I think this is what leads AI teams to solve the wrong problem and get pathological outcomes.

The point of this post is to recommend a different approach. Although I think that domain experts are overall responsible for defining goals and direction, I believe AI experts should be wholly accountable for the problem formulation step (not the domain experts). The word “accountable” is key here- it doesn’t just mean “decide”, it means deciding and being responsible for the outcome of that decision. If AI experts take this accountability seriously, it means they really need to understand the problem well in order to formalize it properly. This means driving lots more conversations with domain experts, not just retreating to the whiteboard.

Let’s apply this idea to the logP example from earlier. The domain experts should still specify what kinds of molecules they want, but AI experts decide whether specifications like “logP < 5” get encoded as a hard constraint or something else. That is, AI experts should think about the spec and proactively ask the domain experts the question “would a bunch of molecules with logP=4.9 be acceptable”, and when they get the answer “no”, act accordingly.

I think most AI experts don’t see themselves as experts in problem formulation. Most AI papers simply state a problem formulation at the start and focus their attention on the methods they choose to solve that problem. It feels to many like a topic outside the purview of AI.

I think this is a mistaken view- AI research training actually gives most AI experts a really strong foundation in how to formulate problems well. Introductory courses in machine learning talk at length about differences between L1 and L2 norms and how this influences algorithm behaviour. Concepts like “reward hacking” and “adversarial examples” have countless research papers written about them. Within AI methods themselves, there can be embedded formulation steps (e.g. substituting accuracy with a differentiable objective like cross-entropy). I think this gives most AI experts a good “gut feeling” for what a good problem formulation looks like, a big library of problem formulations to draw from, and the instinct to hunt for misspecification when evaluating the suitability of a proposed formulation.

I think these are exactly the skills necessary to do problem formulation well in an applied setting. Even though AI experts won’t usually start out knowing the real-world goal, it’s fairly easy to get a rough understanding such that they can propose a couple of candidate problem formulations. From there, AI experts should be able to imagine what real-world consequences could emerge from the different formulations and know which questions to ask to reveal the desired outcome.

Let’s take the standard machine learning regression tasks as an example. Mean squared error and mean absolute error both seem like reasonable objectives to minimize: they are 0 if the predictions are perfect, and increase as predictions get worse. Some domain experts might think both objectives seem fine because of this. However, I think the superpower of an AI expert is first to see that we might not achieve zero error, second to see that the different objectives will “distribute” the errors differently, and finally to propose a question like “would you prefer a model that makes many small errors or a few big errors” that would more precisely elucidate the experts’ preference between the two.

The mean squared error vs mean absolute error is so widely known and documented that it’s unlikely to be an actual stumbling block in real interdisciplinary work. The examples I’ve come across in drug discovery are less obvious. Here are a few common cases:

Constraints that aren’t constraints. Don’t automatically assume things described as “constraints” are true constraints in the mathematical sense. Like in the opening example, the motivation behind constraints like “logP < 5” is often a more smoothly varying preference than a true hard cut-off. Penalties or multi-objective methods can sometimes be a better fit. Similarly though, don’t assume all constraints are really trade-offs: some constraints (e.g. legality, budget) are actually hard constraints.

Heuristics. Experts will often recommend heuristics which do provide useful signals, like logP < 5 for example. However, directly encoding these heuristics may not always be the best choice for AI systems for three reasons:

1. Sometimes these heuristics exist as a simple predictor of a more complex relationship. When that is the case, the value-add of AI is to predict these relationships directly and outperform the heuristics. Directly adopting heuristics can miss an opportunity!

2. Sometimes heuristics are based on correlations. For example, if we want to measure X but it is expensive, we may base a rule on a cheap proxy Y which is correlated with X. When Y does not perfectly predict X, this creates a small misalignment which AI systems may unintentionally amplify.

3. Sometimes a heuristic makes sense in isolation, but not when it is combined with many other heuristics. For example, a single structural filter may trigger you to discard a compound with a downstream toxicity risk, but 1000 filters stacked together might remove all your compounds.

Experts will happily untangle heuristics if you ask them, but might be so used to using them in the first place that it needs to be made an intentional point of conversation.

Workarounds for missing capabilities. Sometimes expert requests are tempered by assumed limitations of AI For example, “run model X, and then check the output by also running model Y” might really mean “we want to assess uncertainty in our predictions.” The expert may specify a workflow, but the underlying need is often a measurement, uncertainty estimate, ranking, or decision rule. AI experts should feel empowered to deviate from specifications when necessary in order to fulfill the underlying request.

In the opening example, the fix wasn’t a better model or more data. It was a conversation: asking the chemists whether their constraints were really constraints, and reformulating accordingly. That conversation probably would have taken a few hours. The alternative (building the wrong system, testing it, realizing the formulation was off, and revising) probably took months. Ultimately, my advice is to have more in-depth conversations about goals and objectives to ensure that AI models do what we want them to do.

But there’s a reason I framed the advice as “AI experts are accountable for the decision” rather than “have more conversations”. “Conversations” is a very open-ended goal- how do you know when you’ve had enough? And whose responsibility is it if we don’t have enough conversations? Plus, it’s very generic and forgettable (it sounds like the advice you’ll hear from countless other places on the internet). I think it’s actually easier to remember accountability (and view extended conversations as a sensible strategy) rather than remembering to have the conversations themselves.

There are also some misaligned incentives to call out. Some problem formulations are genuinely easier than others to work with- e.g. a differentiable objective might be easier to optimize than a discrete one. As AI experts, the temptation to choose an easy problem formulation will always be there. That’s why I think an accountability mindset is so important: we need to get AI systems to do the right thing, not the easy thing.

I think this kind of interdisciplinary work is one of the most exciting areas in ML today. As AI systems take on more complex, long-horizon scientific tasks, the cost of optimizing the wrong objective only gets larger.

At Valence, a lot of our work lives in exactly this translation layer: turning messy scientific goals into AI systems that optimize the right thing. If that kind of work sounds exciting, we’re hiring.

This post is part of “Inside Valence”, a series where you’ll get a behind-the-scenes look at our research, exploring new ways to predict, explain, and ultimately decode biology. If this resonates, consider subscribing!