From Molecules to Decisions: My Takeaways from EFMC-ISMC 2026

Mon, 21 Sep 2026 00:00:00 +0000

The EFMC International Symposium on Medicinal Chemistry 2026 brought the medicinal chemistry community to Basel for five days of science spanning chemical design, computational modeling, pharmacology, structural biology, and drug development.

According to the closing figures shared at the meeting, nearly 1,100 participants attended, with 115 scientific presentations—including eight first-time disclosures—443 posters, and 60 flash talks.

Those numbers capture the scale of the conference, but not its central message.

The strongest theme was not simply that artificial intelligence is entering drug discovery. That transition is already well underway.

The more important shift is that drug discovery is becoming an integrated decision system in which medicinal chemistry, biology, DMPK, structural modeling, automation, and machine learning must operate together.

AI can generate molecules. Physics-based models can estimate how those molecules interact with proteins. Automated laboratories can synthesize and test them.

But none of these technologies removes the need to decide:

Which hypothesis is worth pursuing?

Which experiment will reduce uncertainty most efficiently?

Which molecule should advance—and which project should stop?

My main takeaway from EFMC-ISMC 2026 was therefore simple:

The future of drug discovery will depend less on producing more predictions and more on turning predictions into better decisions.

Drug discovery is becoming a connected system

For much of its history, drug discovery could be represented as a sequence of relatively distinct activities.

Chemists designed molecules. Biologists tested them. DMPK scientists investigated exposure and clearance. Computational chemists modeled binding and properties. Project teams periodically brought these results together to decide what to make next.

That sequence is becoming a continuously connected loop.

Figure 1. Modern drug discovery is less a chain of isolated disciplines than a connected decision system.

This integration sounds straightforward. In practice, it requires shared data infrastructure, compatible experimental and computational workflows, and teams willing to expose uncertainty rather than protect disciplinary boundaries.

That organizational challenge appeared repeatedly throughout the meeting.

Attrition is not only a molecular problem

In an opening keynote, Dr. Muneto Mogi of Novartis Biomedical Research placed the technological excitement against the stubborn reality of pharmaceutical R&D.

Developing a medicine can still require more than a decade.

Phase II remains a major point of attrition. The figures presented at the meeting placed Phase II success at approximately 28.9%, with insufficient efficacy accounting for a large fraction of failures and safety liabilities responsible for many others.

Another striking statistic concerned target concentration. Of approximately 13,600 investigated drug–target pairs presented in the analysis, roughly one quarter were concentrated around only 37 heavily studied targets, including EGFR, GLP-1, PD-1, CD19, HER2, and KRAS.

The exact numbers vary depending on the dataset and definition of success, but the strategic message is difficult to miss:

The industry continues to invest heavily in a comparatively narrow region of biological space.

Exploring less established targets could create new therapeutic opportunities. It also introduces greater uncertainty around target validation, disease relevance, biomarkers, translational models, and safety.

Better algorithms alone will not solve that uncertainty. They must be accompanied by better experimental design and better scientific judgment.

From models to mindset

The Novartis keynote framed the next generation of drug discovery around six complementary capabilities:

Scientific depth — maintaining strong foundations in chemistry, biology, pharmacology, pharmacokinetics, and toxicology.
Digital fluency — understanding data, model outputs, uncertainty, and the practical limitations of AI systems.
Curiosity — exploring unfamiliar chemical and biological spaces instead of repeatedly returning to familiar solutions.
Adaptability — learning new tools and adjusting scientific workflows as automation and modeling evolve.
Open-mindedness — combining human expertise with computational evidence and interdisciplinary perspectives.
Decision quality — using predictions to make accountable choices, including the difficult decision to terminate low-value work early.

I found the last capability particularly important.

Drug discovery does not fail because scientists lack ideas. Projects often fail because uncertainty is resolved too slowly, critical liabilities are discovered too late, or teams continue optimizing molecules that no longer have a credible path to a medicine.

The purpose of prediction should therefore not be to make a project look more certain.

It should be to determine which uncertainty matters next.

The GPS analogy for machine learning

One of the clearest analogies presented at the meeting compared machine learning with a navigation system.

A navigation system can evaluate many possible routes and respond to new information. It cannot decide why the journey matters or whether the destination is still appropriate.

In drug discovery:

The Target Product Profile defines the destination.
Property guidelines and project constraints define the rules of the road.
Experimental and computational KPIs indicate the current position.
Predictive models propose possible routes.
Human scientists remain responsible for steering.

Figure 2. Predictive models can recommend routes, but the destination and final decisions remain human responsibilities.

This analogy avoids two equally unhelpful extremes.

The first is the belief that machine learning will autonomously solve medicinal chemistry. The second is the belief that computational models are merely optional suggestions that can be ignored whenever they conflict with intuition.

A useful navigation system is neither an oracle nor a decoration. It is a decision-support system whose reliability must be understood.

Generative chemistry inside a closed DMTA loop

Generative chemistry featured prominently throughout the meeting, but the most convincing examples were not demonstrations of molecules appearing on a screen.

They were examples in which generation was connected to an operational Design–Make–Test–Analyze cycle.

A generative model can encode molecules in a continuous latent representation, search that space for structures predicted to satisfy multiple objectives, and decode promising points back into chemical structures.

The practical challenge is that drug discovery rarely has one objective. A molecule may need to balance target potency, selectivity, solubility, permeability, metabolic stability, exposure, synthetic accessibility, and safety-related properties such as hERG activity.

Improving one property may damage another. Generative chemistry is therefore most useful when embedded in multi-parameter optimization rather than optimized toward a single score.

Figure 3. Automation becomes most valuable when each experiment improves the next design cycle.

From 55,216 possibilities to nine compounds

One case study began with a combinatorial design space containing 203 head groups, two central cores, and 136 tail groups. Together, these components produced 55,216 virtual combinations.

A conventional medicinal chemistry strategy might explore only a small neighborhood of that space through matched molecular pairs, requiring the synthesis of perhaps 100–200 compounds over multiple cycles.

The presented workflow instead enumerated the complete accessible library, predicted relevant properties, applied multi-parameter scoring, and selected only nine compounds for experimental synthesis.

The important achievement was not simply numerical compression from 55,216 to nine.

The model created a defensible connection between the project objectives and the molecules that were synthesized. It also reduced the risk that promising regions of the enumerated space would remain unexplored merely because they were not obvious to the project team.

However, this efficiency depends on the reliability of the underlying property models. A molecule can only be optimized for liabilities represented in the scoring function.

An omitted objective does not disappear. It becomes an invisible risk.

Automation turns experimental data into a strategic asset

The MicroCycle platform provided a useful example of what a genuinely integrated discovery loop can look like.

MicroCycle combines microscale synthesis, automated purification, robotic handling, biological testing, and machine-learning-driven compound selection. Rather than treating experimental data as the final output of a project cycle, the platform feeds standardized results directly into the next round of decisions.

The deeper advantage is not only speed.

Automation can improve the consistency and completeness of the data generated during each cycle. Negative results become useful training information rather than disappearing into laboratory notebooks or project archives.

This changes the role of the experimental laboratory. It is no longer only a place where computational hypotheses are validated. It becomes an active data-generation engine that continuously improves the models guiding the project.

Can a model learn medicinal chemistry intuition?

A recurring problem in generative chemistry is that a molecule can satisfy formal property constraints while still looking unconvincing to an experienced medicinal chemist.

The objection may involve an unstable functional group, excessive synthetic complexity, an undesirable chemotype, suspicious three-dimensional geometry, a likely reactive metabolite, or a structural pattern associated with previous project failures.

Not all of this knowledge can be represented by a fixed molecular-weight cutoff or a single structural alert.

A preference-learning tool presented as MolSkill attempted to capture part of this tacit expertise by learning from human rankings of molecular designs.

The aim is not to replace the medicinal chemist. It is to make expert preference available earlier and more consistently during molecular generation.

This is promising, but it raises an important question:

Whose intuition is the model learning?

Medicinal chemistry preferences can contain valuable accumulated knowledge. They can also contain historical bias, organizational habits, or unnecessary aversion to unfamiliar chemistry.

Preference models therefore need the same scrutiny as property models. Their value depends on the diversity and consistency of the judgments used for training.

Predicting clearance—and admitting uncertainty

Metabolic stability remains one of the properties that can undermine an otherwise attractive lead series.

A multispecies modeling study discussed at the meeting used a multitask graph neural network to predict in vitro intrinsic clearance, written as CL_int, across six species simultaneously.

Multitask learning can be useful here because related assays and species contain shared information. The model can learn common structure–metabolism relationships while retaining species-specific outputs.

But the most valuable aspect of the work was not the graph architecture. It was the explicit treatment of uncertainty.

A clearance prediction without an estimate of confidence can encourage false precision. An uncertainty-aware model can distinguish between a prediction supported by related training compounds, a prediction based on sparse chemical precedent, and a molecule sufficiently unfamiliar that new experimental data are needed.

This distinction is crucial in prospective design.

The model does not need to be certain about every molecule. It needs to help the project recognize when it is uncertain.

Structural bioinformatics is moving beyond obvious pockets

Another strong theme came from the Cournia group, whose work focused on binding sites that conventional docking workflows may overlook.

Traditional structure-based design often begins with a well-defined pocket visible in an experimental structure. Many biologically important targets do not provide such a convenient starting point.

Potential intervention sites may be transient, allosteric, conformationally dependent, partially exposed to a membrane, or formed only in a specific protein state.

Two approaches—DREAMM and AlloPockets—illustrated how structural bioinformatics is expanding toward these more difficult environments.

DREAMM: targeting protein–membrane interfaces

Peripheral membrane proteins often depend on transient contact with a lipid bilayer. The protein–membrane interface can therefore contribute directly to localization, conformation, and function.

Yet these interfaces are difficult to study from a single soluble protein structure.

DREAMM—Drugging pRotein mEmbrAne Machine learning Method—uses an ensemble machine-learning model to predict membrane-penetrating residues. It then searches protein conformational ensembles for nearby pockets that could potentially support small-molecule binding.

The resulting workflow can be summarized as:

Protein structure → membrane-interface prediction → conformational ensemble → interfacial pocket detection → structure-based design

The important conceptual shift is that the membrane is no longer treated merely as part of the simulation environment. It becomes part of the binding-site definition.

DREAMM does not prove that a predicted pocket will bind a useful ligand. It prioritizes experimentally and computationally testable hypotheses in a region that standard pocket-detection workflows might ignore.

AlloPockets: identifying candidate allosteric sites

Allosteric modulation offers the possibility of controlling a protein through a site distinct from its primary orthosteric pocket.

Such sites can provide new routes to selectivity or allow modulation of targets whose functional sites are difficult to inhibit directly.

Finding them is challenging because allosteric behavior is not determined by geometry alone. A useful prediction may require information about pocket shape, physicochemical properties, solvent exposure, sequence conservation, protein flexibility, mechanical coupling, and the relationship between the pocket and functional motions.

The AlloPockets approach presented at the meeting combined structural pocket descriptors, sequence information, elastic-network-derived dynamics, and protein language model embeddings.

The EFMC-ISMC presentation reported a curated dataset containing more than 3,000 allosteric protein structures and encouraging classification performance for selected model configurations.

Because the software and benchmarks were still evolving in 2026, these values are best interpreted as conference-reported model results rather than a final universal estimate of performance.

The larger lesson is more durable:

A pocket is not allosteric simply because it exists away from the active site.

The important question is whether perturbing that pocket can influence function through the protein’s energetic and dynamical network.

Potency is only the beginning

Some of the most instructive project disclosures concerned molecules that looked highly attractive until a later liability changed the development decision.

These examples were useful reminders that medicinal chemistry is not an optimization contest for the lowest IC₅₀.

A successful drug candidate must operate as a complete system.

A USP1 inhibitor and the CYP3A4 problem

A presentation from Eikon Therapeutics described optimization of a USP1 inhibitor series.

The initial purine hit had challenging physicochemical properties, including a cLogD of 5.4 and ligand-efficiency metrics that left considerable room for improvement.

A scaffold-hopping campaign produced the imidazotriazine compound USP1i-11, with reported nanomolar activity:

USP1 IC₅₀: 58 nM
Ub-PCNA IC₅₀: 10 nM
reported oral bioavailability: approximately 70–141% across the studied conditions
tumor regression in an MDA-MB-436 xenograft model

On potency, exposure, and efficacy, the compound appeared highly promising.

The difficulty emerged in human-donor hepatocytes, where the compound induced expression of CYP2B6 and CYP3A4.

This mattered because USP1 inhibitors were intended for combination with PARP inhibitors, several of which depend strongly on CYP3A4 for their disposition.

Inducing CYP3A4 could reduce exposure to the combination partner and create a clinically important drug–drug interaction risk.

The project therefore illustrated a central principle of medicinal chemistry:

A molecule can succeed against its target and still fail as a medicine.

The CYP result was not a peripheral assay detail. It altered the therapeutic strategy.

New modalities still obey old constraints

The meeting also highlighted continued expansion beyond conventional occupancy-driven inhibitors.

Selective molecular glues

A compound presented as KFA118 was described as a selective molecular glue degrader targeting IKZF3 and IKZF4.

The relevant protein names are:

IKZF3: Aiolos
IKZF4: Eos

Helios is IKZF2, not IKZF3.

Molecular glues are appealing because relatively small molecules can create or stabilize interactions between a target protein and an E3 ubiquitin ligase, leading to target degradation.

But selectivity remains critical. Degrading several closely related transcription factors may produce a different biological and safety profile from selectively degrading one family member.

The modality changes the pharmacology. It does not remove the need for careful structure–activity relationships, exposure control, and translational biology.

One stereocenter can change the project

Another optimization campaign demonstrated the importance of stereochemistry when a single chiral methyl group separated desired antiviral activity from an unwanted mGlu2 negative-allosteric-modulator liability.

The optimized stereochemical configuration reportedly achieved:

antiviral EC₅₀: approximately 20 nM
mGlu2 NAM IC₅₀: greater than 50 μM

This is medicinal chemistry at its most precise.

A small three-dimensional change can reorganize binding interactions sufficiently to preserve one activity while eliminating another.

Machine learning can help prioritize such changes. Only synthesis and experiment can establish what each stereoisomer actually does.

What I took away from the meeting

After five days of presentations, several broader conclusions stayed with me.

1. AI is becoming infrastructure

Machine learning is moving from occasional project support into the operational core of design, compound prioritization, property prediction, and experimental planning.

The most successful examples were not standalone models. They were models connected to chemistry, assays, automation, and project decisions.

2. Multi-parameter optimization is the real problem

Generating potent molecules is rarely enough.

Drug discovery requires simultaneous movement through a landscape containing potency, selectivity, exposure, metabolism, safety, formulation, and synthesis constraints.

The goal is not the molecule with the best individual score. It is the molecule with the most credible overall path to a medicine.

3. Uncertainty must become a first-class output

A prediction without uncertainty encourages teams to confuse numerical precision with scientific confidence.

Models should help identify when a prediction is reliable, when it is extrapolative, and when an experiment is needed.

4. Structural biology is becoming more dynamic

Static structures remain enormously valuable, but the field is increasingly interested in transient pockets, conformational ensembles, membrane-associated states, and allosteric networks.

The target is no longer represented adequately by one structure. It is represented by an ensemble of states with different populations, functions, and druggabilities.

5. DMPK can overturn an otherwise successful design

The USP1 example showed how potency and in vivo efficacy can be insufficient when a compound creates a serious combination-therapy liability.

DMPK and safety are not final filters applied after molecular design. They are part of molecular design.

6. Human judgment remains accountable

Algorithms can enumerate more molecules than a chemist could draw. They can predict dozens of properties and rank enormous virtual libraries.

But models do not carry responsibility for the decision to synthesize a compound, expose an animal, advance a candidate, or begin a clinical trial.

Scientists do.

A changing European scientific landscape

The meeting also closed with announcements about the future of European medicinal chemistry and chemical biology conferences.

The inaugural EuroChemBio meeting will take place at Palazzo Lombardia in Milan from 5–8 October 2027.

EuroChemBio brings together two chemical-biology meeting series:

the European Chemical Biology Symposium, or ECBS;
and the EFMC International Symposium on Chemical Biology, or EFMC-ISCB.

This should not be confused with EFMC-ISMC, the medicinal chemistry symposium discussed in this article.

The EFMC medicinal chemistry meeting series will continue separately, with its next edition planned for Madrid in summer 2028.

Final perspective

EFMC-ISMC 2026 did not present one technology that will transform drug discovery by itself.

Instead, it showed several technologies beginning to converge:

generative chemistry proposes molecular ideas;
physics-based modeling evaluates molecular interactions;
machine learning predicts complex properties;
structural bioinformatics reveals unconventional pockets;
automation accelerates synthesis and testing;
DMPK identifies liabilities that potency alone cannot reveal;
and medicinal chemists integrate the evidence into decisions.

That final step remains the most difficult.

The future of drug discovery is unlikely to belong to AI alone, computational chemistry alone, or automation alone.

It will belong to teams that can connect these tools, understand their uncertainty, and make better decisions with the evidence they generate.

The frontier is not simply the ability to design more molecules.

It is the ability to learn more from every molecule we decide to make.

Notes on the conference disclosures

Compound-level results and numerical project details in this article reflect presentations and notes from EFMC-ISMC 2026. Some were described as first-time disclosures and may not yet have corresponding peer-reviewed publications.

Published references and public resources are provided below where available.

References and further reading

European Federation for Medicinal Chemistry and Chemical Biology. “EFMC International Symposium on Medicinal Chemistry 2026.” Basel, Switzerland, 6–10 September 2026.
Brocklehurst, C. E.; et al. “MicroCycle: An Integrated and Automated Platform to Accelerate Drug Discovery.” Journal of Medicinal Chemistry 2024, 67, 2118–2128.
Rodríguez-Pérez, R.; Trunzer, M.; Schneider, N.; Faller, B.; Gerebtzoff, G. “Multispecies Machine Learning Predictions of In Vitro Intrinsic Clearance with Uncertainty Quantification Analyses.” Molecular Pharmaceutics 2023, 20, 383–394.
Chatzigoulas, A.; Cournia, Z. “Predicting Protein–Membrane Interfaces of Peripheral Membrane Proteins Using Ensemble Machine Learning.” Briefings in Bioinformatics 2022, 23, bbab518.
Chatzigoulas, A.; Cournia, Z. “DREAMM: A Web-Based Server for Drugging Protein–Membrane Interfaces as a Novel Workflow for Targeted Drug Design.” Bioinformatics 2022, 38, 5449–5451.
AlloPockets software repository.
EuroChemBio 2027. Milan, Italy, 5–8 October 2027.

EFMC-ISMC | Yassir Boulaamane