Interview Series – Dirk Tomandl, CTO & Co-Founder, InFocus Therapeutics

DT: I am a cheminformatician and AI scientist by training. Over the course of my career, I have built computational platforms and led teams across multiple drug discovery organizations, including Eli Lilly, Dow/Corteva, and Atomwise. My background spans cheminformatics, QSAR, ADME/Tox, virtual screening, molecular design, scientific computing, machine learning, and artificial intelligence.

At InFocus Therapeutics, my role as CTO and co-founder is to translate that experience into a computational discovery platform that helps our teams make better decisions. This platform, AIDE—our Artificial Intelligence Discovery Engine—is designed to guide which molecules to design, prioritize, and test next.

In practice, I lead platform architecture, computational strategy, and the integration of AI into day-to-day discovery workflows. A central objective is to ensure that the system remains tightly connected to medicinal chemistry, biology, and real program decisions, ultimately supporting meaningful medical outcomes.

DT: Technically, AIDE can be described as a layered decision engine structured around the design-make-test-analyze (DMTA) cycle.

The first layer is the scientific informatics layer, which captures compounds, experimental data, assay results, design history, and hypothesis tracking. It connects to electronic lab notebooks, CRO workflows, and visualization tools.

The second layer is the analysis and modeling layer, where predictive models such as AIDE Crystal, along with machine learning and computational chemistry approaches, are used to interpret structure-activity relationships and evaluate options.

The third layer is the design and prioritization layer, where generative and knowledge-based tools enumerate molecules and propose candidates. Systems such as AIDE Select then prioritize what to synthesize next, depending on the discovery stage.Finally, there is the execution and validation layer, which includes synthesis, biological testing, and the integration of experimental feedback into the next iteration. The key architectural principle is this closed feedback loop.

DT: We do not see those as competing philosophies. AI is very effective at predicting and correlating, but it does not decide.

Our approach is to combine statistical learning with mechanistic and structural reasoning. This includes using chemically aware and partially physics-based representations, maintaining strong involvement from medicinal chemists, and favoring models that are appropriately sized rather than unnecessarily complex.

The guiding principle is that models should inform scientific judgment, not replace it. Human experts remain the final decision-makers. In that sense, AI functions as a multiplier of scientific expertise rather than a substitute for it.

DT: The most tangible impact is in the design and analysis phases. AI allows us to explore much larger areas of chemical space and identify patterns in high-dimensional datasets.

However, the key benefit is not speed in isolation: it is decision quality and learning efficiency. If you can reduce low-value experiments, fail earlier on weak hypotheses, and prioritize stronger candidates sooner, you improve both cycle time and overall program efficiency.

This translates into fewer iterations and more informative cycles, which is where AI delivers its most concrete operational value.

DT: It is fundamentally a structured decision-support system, not a fully automated pipeline.

Scientists interact with property predictions, generative design tools, hypothesis tracking systems, and prioritization outputs. These outputs are always interpreted in context by medicinal chemists, biologists, and computational scientists.The platform is explicitly designed as human-in-the-loop. While we are exploring more agentic capabilities, domain experts retain authority over decisions. The objective is not to replace scientists, but to empower them with better tools and more structured insights.

DT: The key is to avoid reducing outcomes to binary labels. A failed experiment is only meaningful if it is captured with context – what hypothesis was being tested, under which assay conditions, how close analogs behaved, and what part of the cascade failed.

A negative result is not noise if it is reliable. It contributes to defining the local structure-activity landscape and helps narrow subsequent design decisions.

In fact, failed data can be highly informative, particularly when the underlying reasons for failure are well understood.

DT: Data scarcity is a fundamental constraint. We address it by designing systems specifically for low-data regimes, rather than assuming large-scale datasets.

This involves using chemically aware and partially physics-based embeddings, prioritizing generalizable models, and leveraging unlabeled or synthetic data where appropriate.

It also requires discipline in data generation. Not every experiment contributes equally, so the focus is on generating data that is maximally informative for the next decision step.

DT: The primary advantage is that you can design the entire operating model; data flows, decision systems, and team interactions, around AI from the outset. There is no need to retrofit into legacy infrastructure, which allows for tighter alignment between computational and experimental teams.

The downside is the lack of scale. You do not have the portfolio breadth, capital base, or redundancy of large pharmaceutical companies.

As a result, execution discipline becomes critical. You need strong alignment within a smaller team, clear prioritization, and early stop/go decisions, where each choice has a higher relative impact.

DT: Predicting the future precisely is difficult and has proven us wrong a lot of times, but the field is clearly moving through a convergence of several trends.

These include improvements in structure-aware modeling, particularly for complex targets such as RNA; advances in low-data and self-supervised learning; more capable but well-governed agentic systems; tighter integration between computation and experimentation; and the emergence of multimodal models that connect chemistry, biology, and translational constraints.

A key scientific challenge remains the ability of AI to truly understand and explain structure-activity relationships at a level comparable to experienced drug discovery scientists.Overall, the objective is not simply to generate more molecules, but to increase decision precision at each iteration of the discovery cycle.