Leandro Oliveira Bortot

Large language models (LLMs) can generate drug-like molecules and may be used locally in drug discovery projects where confidentiality and intellectual-property constraints preclude the use of external APIs. Weight quantization enables deployment of large open-source LLMs on limited hardware, but its impact on molecular generation and on the ability of models to follow reference chemical spaces remains poorly understood. Here, we evaluated Gemma-4-31B, GPT-OSS-120B, Mistral-Small-4-119B, and Qwen-3.6-35B-A3B at the BF16 (original precision), Q8, Q6, and Q4 quantization levels in two molecular generation tasks. The models were prompted either to generate small molecules without constraints or to generate drug-like molecules conditioned on 69 reference kinase inhibitors provided in-context. Quantization increased inference throughput in a model-dependent manner, with the largest gains, up to 24-fold increase in performance, occurring for Gemma-4-31B and Qwen-3.6-35B-A3B. In the task of unconstrained generation, model quantization did not affect molecular validity, uniqueness, drug-likeness, diversity and synthesizability. In the task of reference-guided generation, however, the models displayed markedly different levels of adherence to the provided chemical space. Gemma-4-31B was the only model that consistently generated small molecules similar to the reference compounds at all precision levels. Mistral-Small-4-119B and Qwen-3.6-35B-A3B failed to generate small molecules according to the reference chemical space. GPT-OSS-120B showed adherence to the reference molecules at BF16, Q8, and Q6, but lost the capacity to generate molecules in the desired chemical space at the Q4 quantization level. These results show that quantization can substantially improve the performance of local LLMs for molecular generation without compromising output quality, although model selection and validation is essential for guided exploration of the chemical space.

Chemical spaces remains poorly understood. Here, we evaluated Gemma-4-31B, GPT-OSS-120B, Mistral-Small-4-119B, and Qwen-3.6-35B-A3B at the BF16 (original precision), Q8, Q6, and Q4 quantization levels in two molecular generation tasks. The models were prompted either to generate small molecules without constraints or to generate drug-like molecules conditioned on 69 reference kinase inhibitors provided in-context. Quantization increased inference throughput in a modeldependent manner, with the largest gains, up to 24-fold increase in performance, occurring for Gemma-4-31B and Qwen-3.6-35B-A3B. In the task of unconstrained generation, model quantization did not affect molecular validity, uniqueness, drug-likeness, diversity and synthesizability. In the task of reference-guided generation, however, the models displayed markedly different levels of adherence to the provided chemical space. Gemma-4-31B was the only model that consistently generated small molecules similar to the reference compounds at all precision levels. Mistral-Small-4-119B and Qwen-3.6-35B-A3B failed to generate small molecules according to the reference chemical space. GPT-OSS-120B showed adherence to the reference molecules at BF16, Q8, and Q6, but lost the capacity to generate molecules in the desired chemical space at the Q4 quantization level. These results show that quantization can substantially improve the performance of local LLMs for molecular generation without compromising output quality, although model selection and validation is essential for guided exploration of the chemical space.

Prion disease is caused by the misfolding of the cellular prion protein, PrP ^C , into a self-templating conformer, PrP ^Sc . Nuclear magnetic resonance (NMR) and X-ray crystallography revealed the 3D structure of the globular domain of PrP ^C and the possibility of its dimerization via an interchain disulfide bridge that forms due to domain swap or by non-covalent association of two monomers. On the contrary, PrP ^Sc is composed by a complex and heterogeneous ensemble of poorly defined conformations and quaternary arrangements that are related to different patterns of neurotoxicity. Targeting PrP ^C with molecules that stabilize the native conformation of its globular domain emerged as a promising approach to develop anti-prion therapies. One of the advantages of this approach is employing structure-based drug discovery methods to PrP ^C . Thus, it is essential to expand our structural knowledge about PrP ^C as much as possible to aid such drug discovery efforts. In this work, we report a crystallographic structure of the globular domain of human PrP ^C that shows a novel dimeric form and a novel oligomeric arrangement. We use molecular dynamics simulations to explore its structural dynamics and stability and discuss potential implications of these new quaternary structures to the conversion process.

Biomolecular crowding affects the biophysical and biochemical behavior of macro- molecules when compared to the dilute environment in experiments on isolated proteins. Computational modeling and simulation are useful tools to study how crowding affects the structural dynamics and biological properties of macromolecules. As computational power increased, modeling and simulating large scale all-atom explicit solvent models of the prokaryote cytoplasm has now become possible. In this work, we build an atomistic model of the cytoplasm of Escherichia coli composed of 1.5 million atoms and submit it to a total of 3μs of molecular dynamics simulations. The model consists of 8 different proteins representing about 50% of the cytoplasmic proteins and one type of t-RNA molecule. Properties of biomolecules under crowding conditions are compared to those from simulations of the individual compounds under dilute conditions. The simulation model is found to be consistent with experimental data about the diffusion coefficient and stability of macromolecules under crowded conditions. In order to stimulate further work we provide a Python script and a set of files to enable other researchers to build their own E. coli cytoplasm models to address questions related to crowding.

The Zika virus (ZIKV) is an arthropod-borne virus that belongs to theFlaviviridaefamily. The ZIKV infection is usually asymptomatic or is associated with mild clinical manifestations; however, increased numbers of cases of microcephaly and birth defects have been recently reported. To date, neither a vaccine nor an antiviral treatment has become available to control ZIKV replication. Among the natural compounds recognized for their medical properties, flavonoids, which can be found in fruits and vegetables, have been found to possess biological activity against a variety of viruses. Here, we demonstrate that the citrus flavanone naringenin (NAR) prevented ZIKV infection in human A549 cells in a concentration-dependent and ZIKV-lineage independent manner. NAR antiviral activity was also observed when primary human monocyte-derived dendritic cells were infected by ZIKV. NAR displayed its antiviral activity when the cells were treated after infection, suggesting that NAR acts on the viral replication or assembly of viral particles. Moreover, a molecular docking analysis suggests a potential interaction between NAR and the protease domain of the NS2B-NS3 protein of ZIKV which could explain the anti-ZIKV activity of NAR. Finally, the results support the potential of NAR as a suitable candidate molecule for developing anti-ZIKV treatments.

Every two years groups worldwide participate in the Critical Assessment of Protein Structure Prediction (CASP) experiment to blindly test the strengths and weaknesses of their computational methods. CASP has significantly advanced the field but many hurdles still remain, which may require new ideas and collaborations. In 2012 a web-based effort called WeFold, was initiated to promote collaboration within the CASP community and attract researchers from other fields to contribute new ideas to CASP. Members of the WeFold coopetition (cooperation and competition) participated in CASP as individual teams, but also shared components of their methods to create hybrid pipelines and actively contributed to this effort. We assert that the scale and diversity of integrative prediction pipelines could not have been achieved by any individual lab or even by any collaboration among a few partners. The models contributed by the participating groups and generated by the pipelines are publicly available at the WeFold website providing a wealth of data that remains to be tapped. Here, we analyze the results of the 2014 and 2016 pipelines showing improvements according to the CASP assessment as well as areas that require further adjustments and research.

Biological membranes are continuously remodeled in the cell by specific membrane-shaping machineries to form, for example, tubes and vesicles. We examine fundamental mechanisms involved in the vesiculation processes induced by a cluster of envelope (E) and membrane (M) proteins of the dengue virus (DENV) using molecular dynamics simulations and a coarse-grained model. We show that an arrangement of three E-M heterotetramers (EM ₃ ) works as a bending unit and an ordered cluster of five such units generates a closed vesicle, reminiscent of the virus budding process. In silico mutagenesis of two charged residues of the anchor helices of the envelope proteins of DENV shows that Arg-471 and Arg-60 are fundamental to produce bending stress on the membrane. The fine-tuning between the size of the EM ₃ unit and its specific bending action suggests this protein unit is an important factor in determining the viral particle size.