Biology subjects

Berman, D.

Publications and source records attributed to Berman, D..

6 recordsLinked to original sources

bioRxiv2025

Evaluating the Effectiveness of Parameter-Efficient Fine-Tuning in Genomic Classification Tasks

Foundation models are increasingly being leveraged for biological tasks. To address the high memory requirements of fine-tuning large pre-trained language models, parameter efficient fine-tuning (PEFT) methods are also being increasingly utilized. Previous studies have shown minimal, if any, loss in performance when using PEFT on binary classification tasks. However, the impact of using PEFT on tasks with large classification spaces has not been systemically evaluated. In this work, we apply PEFT to the problem of taxonomic classification using pre-trained genomic language models as the classification backbone. We explore various training strategies--including PEFT, full fine-tuning, and partial fine-tuning--for classifying sequences at the superkingdom, phylum, and genus levels. We find that PEFT-trained models significantly underperform compared to those trained via full fine-tuning or partial fine-tuning. Additionally, we demonstrate increased performance of pretrained models over those randomly initialized.

bioinformatics↗

bioRxiv2024

Protein Language Models in Directed Evolution

The dominant paradigms for integrating machine-learning into protein engineering are de novo protein design and guided directed evolution. Guiding directed evolution requires a model of protein fitness, but most models are only evaluated in silico on datasets comprising few mutations. Due to the limited number of mutations in these datasets, it is unclear how well these models can guide directed evolution efforts. We demonstrate in vitro how zero-shot and few-shot protein language models of fitness can be used to guide two rounds of directed evolution with simulated annealing. Our few-shot simulated annealing approach recommended enzyme variants with 1.62 x improved PET degradation over 72 h period, outperforming the top engineered variant from the literature, which was 1.40 x fitter than wild-type. In the second round, 240 in vitro examples were used for training, 32 homologous sequences were used for evolutionary context and 176 variants were evaluated for improved PET degradation, achieving a hit-rate of 39 % of variants fitter than wild-type.

bioinformatics↗

bioRxiv2023

Machine Learning Identifies Plasma Proteomic Signatures of Descending Thoracic Aortic Disease

BackgroundDescending thoracic aortic aneurysms and dissections can go undetected until severe and catastrophic, and few clinical indices exist to screen for aneurysms or predict risk of dissection. MethodsThis study generated a plasma proteomic dataset from 75 patients with descending type B dissection (Type B) and 62 patients with descending thoracic aortic aneurysm (DTAA). Standard statistical approaches were compared to supervised machine learning (ML) algorithms to distinguish Type B from DTAA cases. Quantitatively similar proteins were clustered based on linkage distance from hierarchical clustering and ML models were trained with uncorrelated protein lists across various linkage distances with hyperparameter optimization using 5-fold cross validation. Permutation importance (PI) was used for ranking the most important predictor proteins of ML classification between disease states and the proteins among the top 10 PI protein groups were submitted for pathway analysis. ResultsOf the 1,549 peptides and 198 proteins used in this study, no peptides and only one protein, hemopexin (HPX), were significantly different at an adjusted p-value <0.01 between Type B and DTAA cases. The highest performing model on the training set (Support Vector Classifier) and its corresponding linkage distance (0.5) were used for evaluation of the test set, yielding a precision-recall area under the curve of 0.7 to classify between Type B from DTAA cases. The five proteins with the highest PI scores were immunoglobulin heavy variable 6-1 (IGHV6-1), lecithin-cholesterol acyltransferase (LCAT), coagulation factor 12 (F12), HPX, and immunoglobulin heavy variable 4-4 (IGHV4-4). All proteins from the top 10 most important correlated groups generated the following significantly enriched pathways in the plasma of Type B versus DTAA patients: complement activation, humoral immune response, and blood coagulation. ConclusionsWe conclude that ML may be useful in differentiating the plasma proteome of highly similar disease states that would otherwise not be distinguishable using statistics, and, in such cases, ML may enable prioritizing important proteins for model prediction.

systems biology↗

bioRxiv2022

RNAi-mediated rheostat for dynamic control of AAV-delivered transgenes

Adeno-associated virus (AAV)-based gene therapy could be facilitated by the development of molecular switches to control the magnitude and timing of expression of therapeutic transgenes. RNA interference (RNAi)-based approaches hold unique potential as a clinically proven modality to pharmacologically regulate AAV gene dosage in a sequence-specific manner. We present a generalizable RNAi-based rheostat wherein AAV transgene expression is silenced using the clinically validated modality of chemically modified short interfering RNA (siRNA) conjugates or vectorized co-expression of short hairpin RNA (shRNA). For transgene induction, we employ REVERSIR technology, a synthetic high-affinity oligonucleotide complementary to the siRNA or shRNA guide strand to reverse RNAi activity and rapidly recover transgene expression. For potential clinical development, we report potent and specific siRNA sequences that may allow selective regulation of transgenes while minimizing unintended off-target effects. Our results establish a conceptual framework for RNAi-based regulatory switches with potential for infrequent dosing in clinical settings to dynamically modulate expression of virally-delivered gene therapies.

genomics↗

bioRxiv2022

Robust and durable prophylactic protection conferred by RNA interference in preclinical models of SARS-CoV-2

RNA interference is a natural antiviral mechanism that could be harnessed to combat SARS-CoV-2 infection by targeting and destroying the viral genome. We screened lipophilic small-interfering RNA (siRNA) conjugates targeting highly conserved regions of the SARS-CoV-2 genome and identified leads targeting outside of the spike-encoding region capable of achieving [≥]3-log viral reduction. Serial passaging studies demonstrated that a two-siRNA combination prevented development of resistance compared to a single-siRNA approach. A two-siRNA combination delivered intranasally protected Syrian hamsters from weight loss and lung pathology by viral infection upon prophylactic administration but not following onset of infection. Together, the data support potential utility of RNAi as a prophylactic approach to limit SARS-CoV-2 infection that may help combat emergent variants, complement existing interventions, or protect populations where vaccines are less effective. Most importantly, this strategy has implications for developing medicines that may be valuable in protecting against future coronavirus pandemics.

pharmacology and toxicology↗

bioRxiv2018

Sustained hovering, head stabilization and vision through the water surface in the Pied kingfisher (Ceryle rudis).

Pied kingfishers (Ceryle rudis) capture fish by plunge diving from hovering that may last several minutes. Hovering is the most energy-consuming mode of flight and depends on active wing flapping and facing headwind. The power for hovering is mass dependent increasing as the cube of the size, while aerodynamic forces increase only quadratically with size. Consequently, birds above a certain body mass can hover only with headwind and for very short durations. Hummingbirds are referred to as the only birds capable of hovering without wind (sustained hovering) due to their small size (ca. 2-20 gr), high wing-beat frequency and unique anatomy.\n\nWe studied the hovering characteristics of pied kingfishers in relation to wind and sun orientation, in 139 hovers. Furthermore, plunge diving necessitates the coping with the visual effects of light at the air/water interface. The kingfishers oriented their body axis towards the wind more than towards the sun. Hovering in little or no wind was common. With increased wind speed (a) orientation precision increased, (b) wing beat amplitude did not change, (c) wing beat frequency decreased and (d) body tilt became more horizontal. The head was highly stabilized and with orientations that indicated monocular viewing of prey.\n\nWe conclude that pied kingfishers achieve sustained hovering. This is despite their being significantly heavier than the theoretical maximum and showing ordinary kinematics and morphology. Head stabilization is a means of aiding viewing of submerged prey across the interface.

zoology↗