Alex Abrudan

Alex Abrudan

Auto-generated profile, based on publicly available information

Alex Abrudan

InstaDeep Ltd

Research Focus & Metrics

Academic Career Length

3Years (2024-Present)

(Calculated once Scilit publication data is synced: current year − year of first publication + 1)

Representative Publications

View All Publications
22 November 2024 in Chemical Science
Verified
Journal Article
Open access
PDF file
JHJoren Van Herck
Joren Van Herck
MGM. Victoria Gil
M. Victoria Gil
KJKevin Maik Jablonka
Kevin Maik Jablonka
AAAlex Abrudan
Alex Abrudan
AAAndy S. Anker
Andy S. Anker
BSBerend Smit
Berend Smit
We studied the performance of fine-tuning open-source LLMs for a range of different chemical questions. We benchmark their performances against “traditional” machine learning models and find that, in most cases, the fine-tuning approach is superior.

20 August 2025 in bioRxiv
Verified
Preprint
Open access
BPBenjamin du Pont
Benjamin du Pont
AAAlex Abrudan
Alex Abrudan
RSRob Scrutton
Rob Scrutton
VRVladimir Radenkovic
Vladimir Radenkovic
HZHaowen Zhao
Haowen Zhao
Tuomas KnowlesTuomas Knowles
Tuomas Knowles
Macrocyclic peptides offer strong therapeutic potential due to their enhanced binding affinity and protease resistance, but their design remains a challenge due to limited structural data and tools that address only a narrow set of cyclization chemistries. Moreover, existing models are built to only consider ground state or mean conformations, rather than conformational ensembles that more accurately describes peptides. We introduce CycLOPS (a Cyclic Loss for the Optimization of Peptide Structures), a model-agnostic framework that conditions Boltzmann generators to sample valid cyclic conformations without retraining. To overcome the scarcity of cyclic peptide data, we reformulate the design problem in terms of conditional sampling over linear peptide structures via chemically informed loss functions. CycLOPS spans 18 possible inter-amino acid crosslinks enabled by 6 diverse chemical reactions, and is readily extensible to many more. It leverages tetrahedral geometry constraints, using six interatomic distances to define a kernel density-estimated joint distribution from MD simulations. We demonstrate CycLOPS's versatility via two distinct generative models: a modified Sequential Boltzmann Generator (SBG) and Equivariant Normalizing flow (ECNF). In both settings, CycLOPS successfully biases the Boltzmann distribution toward chemically plausible macrocycles.

11 September 2026 in bioRxiv
Verified
Preprint
Open access
PDF file
QTQuentin Trolliet
Quentin Trolliet
AAAlex Abrudan
Alex Abrudan
ABAryan Bhasin
Aryan Bhasin
YFYunguan Fu
Yunguan Fu
JEJoao Paulo Euko
Joao Paulo Euko
FSFrancesco Saccon
Francesco Saccon
Peptides occupy a valuable niche between small molecules and biologics, but the clinical translation of de novo peptide designs requires rigorous scoring to simultaneously optimise target binding affinity alongside multiple developability traits, including stability, membrane permeability, aggregation propensity, and non-fouling behaviour. Here, we evaluate two distinct approaches for scoring these candidates: classical chemical descriptors and modern deep learning representations derived from protein language and folding models. Assembling nine public datasets spanning five developability traits and four binding-affinity endpoints, we find sequence-derived chemical descriptors alone contain sufficient information to predict developability task labels effectively. Given their drastically lower computational cost and higher interpretability, classical machine learning models trained on these simple descriptors frequently match or approach the performance of complex deep learning architectures, emerging as a highly efficient and interpretable alternative for high-throughput scoring. Finally, for scoring binding affinity, we demonstrate that Boltz-2 pair representations capture the most information among the tested representations; however, the model's predictive power is confounded by a significant bias from the molecular weight of the peptides. Together, these results establish a comprehensive assessment of state-of-the-art methods for predicting both peptide developability and binding affinity, highlighting the enduring value of interpretable chemical descriptors alongside deep learning in the scoring and selection of de novo peptide designs.