On 27 April 2026, I defended my doctoral thesis, Multimodal redundancy-reduced deep learning for drug discovery, and completed my doctorate with summa cum laude. It marked the end of an important chapter and gave me a good reason to look back at what made these years so valuable.

From biochemistry into AI

I am a biochemist by training. My transition into computational research began during my master's thesis, when I first started applying machine learning to biochemical questions. I enjoyed working at the interface of both fields and decided to continue in AI for drug discovery full-time. My doctorate at the Sieber Research Group was therefore entirely computational.

The central theme of my work was how to make useful predictions when experimental data are limited. In TwinBooster, we combined molecular structures with written assay descriptions to predict activity for assays the model had never seen. BarlowDTI used molecular and protein sequences to predict drug-target interactions without requiring experimental 3D structures. Together with experimental colleagues, I also helped connect generative modelling, prediction, automated synthesis, and biological testing in the discovery of a new antibiotic lead active against Gram-positive and Gram-negative bacteria.

Further projects made me more critical about how computational methods are evaluated. We carried out a prospective experimental assessment of structure-based drug design and found that plausible-looking generated molecules rarely translated into reproducible biochemical activity. In a later Chemical Science paper, we examined widely used biomolecular benchmarks and showed how leakage, conflicting labels, and redundancy can distort model comparisons.

A good place to learn

I am very grateful for the freedom and trust I was given at the Technical University of Munich. I was taught and mentored by great people who were generous with their time and knowledge. At the same time, I was able to develop my own ideas, manage projects, work closely with collaborators from chemistry and biology, and take responsibility for parts of our research infrastructure. That combination of guidance and independence was one of the most valuable parts of the experience.

Conferences were an important part of that experience. At Generative AI in Molecule Discovery at Helmholtz Munich, I presented a poster and met researchers working on related questions. At the Gordon Research Seminar and Conference on Computer Aided Drug Design in Portland, I joined discussions on next-generation methods and presented our work in a short talk and poster. Our prospective study of structure-based drug design was later presented by Joshua Hesse at the ICLR 2026 GEM Workshop in Rio de Janeiro. These meetings gave me a perspective on the field that papers alone cannot provide.

I am thankful to everyone who taught me, worked with me, and made the difficult parts easier. The freedom, shared projects, and people around them made these years both valuable and enjoyable. I am grateful for the experience and curious about what comes next 👀

The full dissertation is available on mediaTUM.

Publication update, 22 July 2026: This work is now published open access in Chemical Science. I have updated this post to reflect the final, peer-reviewed paper.

A benchmark is a shared exam

In machine learning, a benchmark is a little like a shared exam. Researchers train their models on one part of a dataset and test them on a separate part. Because everyone answers the same questions, the resulting scores give us a convenient way to compare methods—or at least that is the idea.

Now imagine that the exam contains questions that already appeared in the homework. Or that the same question appears twice with two different correct answers. A student might score unusually well by recognising familiar material, or lose points because the answer key contradicts itself. Either way, the score no longer measures quite what we thought it did.

The same problem can occur in molecular machine learning. A test molecule may also be present in the training data. Nearly identical molecules can sit on both sides of the split. The same compound can carry conflicting measurements. In drug–target interaction datasets, the compounds may differ while the protein targets are identical or extremely similar. A model can then succeed through familiarity rather than by learning a rule that transfers to genuinely new chemistry or biology.

That possibility made us want to inspect the exam before designing yet another student.

What we audited

We built an open-source toolkit called BenchAudit and applied it to 51 dataset configurations from Polaris, Therapeutics Data Commons, MoleculeNet, and several widely used drug–target interaction benchmarks. Together, these collections have accumulated more than 10,000 citations, so the question is not academic: they underpin a great deal of work in the field.

The audit follows the data from basic hygiene to the final train–test split. It looks for duplicate molecules, contamination between splits, conflicting labels, highly similar structures, and sharp changes in activity between near-neighbours. For drug–target interaction tasks, it also checks whether proteins or complete drug–target pairs recur across splits and how similar the protein sequences are.

These checks are diagnostic, not automatic verdicts. A shared molecular scaffold can be scientifically meaningful. The same drug tested against two different proteins is not a duplicate experiment. Activity cliffs—similar molecules with very different activities—are often exactly the difficult cases a useful model should learn. The point of an audit is to make these properties visible, so that we can decide whether a split actually tests the kind of generalisation being claimed.

What we found

Problems were widespread, but they did not look the same everywhere. We found cross-split contamination, label conflicts, and substantial structural redundancy across many of the benchmarks. The drug–target interaction datasets deserve particular care: exact duplicate drug–target pairs were comparatively rare, yet the same or closely related proteins and ligands often appeared across training and test data. A model can therefore be evaluated in a very familiar neighbourhood even when the complete interaction pair is technically new.

We then tested whether these issues really move the numbers. In controlled experiments, we deliberately added different kinds of noise to 41 datasets. Conflicting or random labels made performance worse, while train–test contamination made it look better. The direction of the effect was consistent across random forests, gradient-boosted trees, and a neural network. In other words, these are not merely untidy rows in a spreadsheet; they can change the conclusion of an experiment.

The final paper adds another test that I find especially informative. We kept every model and every prediction fixed, but changed the composition of the test panel. The original leading model remained first on audit-clean panels in 23 of 28 datasets, so the result is not simply that every winner benefited from obvious leakage. Yet rankings became much less stable when test panels contained more near-training examples or conflicting labels. In the 14 datasets where a conflict-enriched panel could be constructed, the original leader’s chance of remaining first fell below 50% in 12 cases.

That is a more useful conclusion than saying that every leaderboard is wrong. A leading model may well be the best model for a particular benchmark. But its lead is conditional on what the test set contains, and small differences in a table should not automatically be treated as evidence of a generally superior method.

What I take from it

A benchmark score is never a property of the model alone. It belongs to the model, the dataset, the split, and the chemical or biological composition of the test set. If a paper makes a claim about new scaffolds, unseen targets, or real-world generalisation, the benchmark has to create that challenge explicitly.

For me, this moves dataset auditing from a cleanup step to part of the scientific method. We should report what is shared across splits, explain what kind of novelty the test set represents, and check whether a model ranking survives on clean or deliberately challenging subsets. Better documentation helps, but provenance alone does not guarantee quality; even a reproducible split can reproducibly test the wrong thing.

Benchmarks are still indispensable. They let us build on each other’s work and measure progress without repeating every experiment from scratch. We just need to stop treating them as neutral yardsticks. As Pat Walters has argued, the field needs to look beyond the bold numbers in a table. The data beneath those numbers are part of the result.

Read the paper

M. G. Schuh, A. Daniluk and S. A. Sieber, Chem. Sci., 2026, 17, 17455–17467.

The BenchAudit source code and documentation are openly available.

The problem

Antibiotic discovery has slowed down while resistant bacteria have continued to spread. One reason is that our current molecular libraries are biased toward chemistry we already know well. That matters even more for Gram-negative bacteria such as E. coli, because their outer membrane blocks many molecules before those molecules can do any useful work.

The idea

In this paper, we asked whether generative AI could help us search beyond that familiar chemical space. We did not want to stop at drawing new molecules on a screen. We wanted to build a workflow that could generate unfamiliar candidates, narrow them down intelligently, and then test them in the lab.

How it works

We trained a chemical language model on drug-like molecules and natural products, then refined it with a diverse set of antibiotic scaffolds. We filtered the resulting candidates with predictive modeling and expert curation so that we could prioritize molecules that looked both synthetically accessible and promising for antibiotic activity. We then synthesized and tested the shortlisted compounds, and we followed the best lead with an iterative round of optimization through 40 derivatives.

What the paper showed

That process produced a clear lead. We identified a candidate with potent activity against methicillin-resistant Staphylococcus aureus, and our follow-up chemistry generated a broader set of active compounds. In the paper, we report that 30 derivatives showed activity against S. aureus and 17 against Escherichia coli. One lead compound, D8, reached submicromolar potency against S. aureus and single-digit micromolar potency against E. coli. Our mechanistic experiments pointed to reductive generation of reactive species as the main mode of action.

Why it matters

What makes this project interesting to me is that we moved generative chemistry out of the purely virtual realm. We did not stop at “the model made some molecules.” We followed the harder path from generated structures to synthesis, biological testing, and lead refinement. That makes the argument more convincing: AI can help us explore new parts of chemical space rather than simply rearrange the parts we already search.

Limits

We are still in early-stage antibiotic discovery, not at the stage of a finished medicine. Much of the work lies in turning a huge virtual set into a tractable experimental program, and the outcome is a lead series, not an endpoint. But that is exactly why the paper matters: it shows what it takes to turn generative suggestions into experimental traction.

Read the paper

Kollen, M. F.; Schuh, M. G.; Kretschmer, R.; Hesse, J.; Schum, D.; Chen, J.; Bohne, A. I.; Halter, D. P.; Sieber, S. A. JACS Au 2025, 5 (9), 4249–4259. 10.1021/jacsau.5c00602

Update, March 2026: A follow-up paper, presented at the GEM Workshop at ICLR 2026, took this pipeline into the laboratory. See Experimental follow-up below.

The problem

Antibiotic resistance is not just a scientific challenge. It also challenges the way we organize discovery. We now have many computational tools that promise to help, from protein structure prediction to generative chemistry, but we still lack clear guidance on how to connect them into a discovery process that people can actually use.

The idea

In this project, we tackled that gap directly. Instead of presenting yet another isolated model, we asked what an end-to-end AI-guided antibiotic pipeline should look like if we want to move from target selection to a realistic set of candidate compounds.

How it works

We began by looking across predicted proteomes from multiple pathogens to identify targets that are conserved, essential, and not close human counterparts. We then compared six 3D-structure-aware generative models from different families, including diffusion, autoregressive, graph neural network, and language-model approaches. From there, we kept the workflow deliberately practical: we used post-processing filters and commercial analogue searches to reduce more than 100,000 generated compounds to a much smaller set that we could realistically synthesize.

What the paper showed

We found that DeepBlock and TamGen were the strongest overall performers across the criteria we examined, but we also found that no single model solves the whole problem. Some methods are more sophisticated but harder to use well. Others are easier to deploy but weaker on chemical or biological relevance. That makes this paper useful not just as a benchmark, but as a map of the trade-offs we face when we try to turn AI promises into decisions.

Why it matters

I think that broader view matters because early-stage antibiotic discovery often suffers from fragmentation. We can easily end up with many impressive components and very little guidance on how to combine them. In this work, we shifted the focus from isolated performance claims to pipeline design: what we should prioritize, where the bottlenecks are, and how we can move from a large virtual search space toward something that could matter in practice.

Limits

We are still describing an early-stage blueprint, not a finished discovery engine. Our conclusions depend on the quality of the target-selection logic, generative models, and filtering steps underneath the pipeline, and all of those components will keep evolving. But that is part of the point: in this paper, we tried to make the workflow itself easier to see and judge.

Experimental follow-up

A later paper, presented at the GEM Workshop at ICLR 2026, took this pipeline into the laboratory. Together with experimental colleagues, we ran six state-of-the-art structure-based generative models against three structurally selected bacterial targets and reduced more than 100,000 generated designs to a prioritized, purchasable set through standardized post-processing, medicinal-chemistry filtering, and commercial analogue mapping. For MurC, the most tractable target, we acquired and tested 47 compounds. Only one showed weak, measurable inhibition at the screening concentration. That outcome is sobering but informative: it marks a substantial gap between protein-conditioned designs that look plausible on paper and molecules that show activity under realistic discovery constraints, and it points to where the next generation of structure-based methods needs to improve.

Read the paper

Schuh, M. G.; Hesse, J.; Sieber, S. A. arXiv 2025. 10.48550/arXiv.2504.11091

Schuh, M. G.; Hesse, J.; Eckel, K.; Van, Q. C.; Halbwedl, N.; Weber, F.; Jäschke, A.; Zacharias, M.; Sieber, S. A. "A Prospective Experimental Assessment of Structure-Based Drug Design: Are We There Yet?" GEM Workshop, ICLR 2026. openreview.net/forum?id=k53jrNuEBT

The problem

For all the excitement around AI in drug discovery, many methods still depend on something we often do not have: high-quality 3D structural data. That creates a real bottleneck. If we want to test many drug-target pairs quickly, the most detailed approaches are often the hardest ones to use at scale.

The idea

In this project, we took a more pragmatic route. Instead of beginning with full 3D structures, we started from the information that is usually available from the beginning: a SMILES string for the molecule and an amino-acid sequence for the protein. These inputs are simple, but they are cheap, common, and fast to process. We wanted to know whether that simpler view of the problem could still produce genuinely useful predictions.

How it works

We turn the molecule into a compact fingerprint and the protein sequence into embeddings from a protein language model that carries some structural awareness. A Barlow Twins training step then learns a shared representation of drug-target pairs, and a gradient-boosted model makes the final prediction. That hybrid design sits at the center of the paper: we use deep learning where it is strongest, in learning good representations, and then we use a smaller, efficient model where robustness and speed matter most.

What the paper showed

Across several established drug-target benchmarks, we saw state-of-the-art performance while using only one-dimensional inputs. We also looked beyond leaderboard numbers. By tracing which training examples most influenced a prediction and comparing the results with co-crystal structures, we found that the model’s signal aligned with catalytically active and stabilizing residues. That matters because it suggests the model is not simply exploiting accidental patterns in the data. We then extended the approach to BarlowDTI XXL, trained on more than 3.6 million curated drug-target pairs for larger, more realistic applications such as virtual screening.

Why it matters

I think the appeal here is practical. If we can make strong predictions from sequences and molecular strings alone, more people can use these tools without waiting for expensive structural experiments or heavy computation. That makes the method useful not only as a research result, but also as a working tool for early-stage discovery.

Limits

Sequence-only methods still have limits. Some interactions depend on fine 3D geometry, and when reliable structural data exist, structure-based approaches can still have an advantage. More broadly, we can only trust a predictive system as much as we trust the data it learns from.

Read the paper

Schuh, M.G., Boldini, D., Bohne, A.I. et al. J Cheminform 17, 18 (2025). https://doi.org/10.1186/s13321-025-00952-2

Try it

We also provide a public web demo: bio.nat.tum.de/oc2/barlowdti

The problem

In drug discovery, we often want to know whether a molecule will be active in a biological test, or assay. The difficulty is that the most interesting assays are often the newest ones, and new assays usually arrive with very little data. That creates a familiar bottleneck: the question matters, but the measurements are not there yet.

The idea

TwinBooster begins with an overlooked fact. Assays are not just empty labels waiting for numbers. They usually come with words attached: a name, a target, a short description, sometimes a protocol. In this project, we asked whether we could treat that text as part of the scientific evidence rather than as background paperwork. If a model can read the assay description and connect it to the molecule in front of it, it may be able to make useful predictions before that assay has built up a large dataset of its own.

How it works

At a high level, we give the model two views of the same problem. One view is the chemical structure of the molecule. The other is the text describing the assay. A language model turns the assay text into a numerical representation, while a standard chemical representation does the same for the molecule. We then use a self-supervised training step to encourage those two sides to line up when a molecule is active in a given assay. After that, a lightweight gradient-boosted model makes the final prediction.

What the paper showed

On FS-Mol, a benchmark designed to test performance on new and low-data assays, we achieved state-of-the-art zero-shot performance. In plain English, we were able to make strong predictions for assays we had not trained on directly. We also looked at a more practical scenario: choosing compounds for a screening campaign. There too, we could help prioritize likely hits without collapsing chemical diversity into the same obvious choices.

Why it matters

I think that matters because the early stage of drug discovery is full of sparse information. If we can use assay descriptions as part of the signal, we do not have to wait for thousands of measurements before getting something useful from a model. We can turn the written context around an experiment into something computationally valuable.

Limits

The approach still depends on the text being informative. A vague assay description gives the model much less to work with. And zero-shot prediction does not replace real assay data; I see it as an early guide, not a final verdict.

Read the paper

Schuh, M. G.; Boldini, D.; Sieber, S. A. J. Chem. Inf. Model. 2024, 64 (12), 4640–4650. 10.1021/acs.jcim.4c00765

Research papers are usually written for specialists. That is necessary, but it also means the most interesting part of the work is often buried under compressed language, technical shorthand, and assumptions about what the reader already knows.

This blog is where I want to do the opposite. I want to write about research in plain language, with enough precision to stay honest and enough narrative to make the ideas readable.

Most posts here are short commentaries on projects I have worked on: what question the work asks, why that question matters, what the central idea is, and where the limits still are.