Questions & explanations
1. Calculate the equilibrium potential for chloride if its inside concentration is 10 mM and outside is 110 mM at 37°C (use Nernst equation).
The Nernst equation for chloride at 37°C is E_Cl = -61.5 * log(inside/outside) mV. Inside is 10 mM, outside is 110 mM, so ratio = 10/110 = 0.0909. log(0.0909) = -1.041. Multiply by -61.5 gives +64 mV. So the equilibrium potential for chloride is about -64 mV (since chloride is negative, the sign is negative). Actually careful: E_Cl = -61.5 * log([Cl-]in/[Cl-]out) = -61.5 * log(10/110) = -61.5 * (-1.041) = +64 mV. But the standard convention is that the potential is inside relative to outside, so a positive value means inside is positive relative to outside. However, chloride is an anion, so its equilibrium potential is usually around -65 mV. Let's correct: For chloride, the Nernst equation is often written as E_Cl = 61.5 * log([Cl-]out/[Cl-]in) = 61.5 * log(110/10) = 61.5 * 1.041 = 64 mV, but the sign is negative because chloride is negative. Actually, the equilibrium potential is defined as the voltage inside minus outside. For chloride, if outside concentration is higher, chloride will move in, making inside more negative, so E_Cl is negative. Using the formula E_Cl = -61.5 * log([
2. What is a potential problem when using transfer learning across different types of neural signals, like from EEG to electrocorticography (ECoG, brain signals from electrodes on the brain surface)?
EEG and ECoG have different properties: EEG is recorded from the scalp and has lower resolution and more noise, while ECoG is directly on the brain and has higher frequency content. A model trained on EEG may not work well on ECoG because the signal characteristics differ. For example, EEG has strong low-frequency activity, while ECoG captures high-frequency bursts. Transfer learning might still help if you adapt the model to the new signal type. You could pre-train on a large EEG dataset and then fine-tune on a small ECoG dataset. However, the model's early layers might be tuned to EEG's noise patterns, which are different. So you may need to retrain some early layers as well. The key is to find common features, like the fact that both signals reflect neural activity, but the exact patterns vary.
3. Give an example of transfer learning where a model trained on motor imagery (thinking about moving) is used for a different task like detecting errors from brain signals.
A model trained on motor imagery learns to recognize patterns like event-related desynchronization (ERD) in the sensorimotor cortex. These patterns are changes in brain rhythms when you imagine moving. For error detection, the brain produces a different signal called error-related negativity (ERN) that occurs when you make a mistake. Even though the tasks are different, the basic features of EEG, like frequency bands and spatial patterns, are similar. You can take the pre-trained motor imagery model and remove its last classification layer. Then you add a new layer for error detection and train it on a small set of error trials. The model's earlier layers already know how to filter noise and extract relevant frequencies. This speeds up learning and can achieve good accuracy for error detection.
4. What is the difference between transfer learning and domain adaptation in neural signal analysis?
Transfer learning is a broad term where knowledge from one task is used for another. Domain adaptation is a specific type of transfer learning where the source and target tasks are the same, but the data distributions differ. For neural signals, the source domain could be data from one recording session, and the target domain could be data from a later session with different electrode positions. Domain adaptation methods try to align the feature distributions so that the model works across sessions. For example, they might transform the signals to have similar statistical properties. Transfer learning can also involve different tasks, like from motor imagery to P300 detection. Domain adaptation focuses on the same task but different conditions. Both are useful for making BCI systems robust.
5. What is a common challenge in compiling a regulatory submission for a biopharmaceutical, and how can it be addressed?
A common challenge is managing the huge amount of data required, especially for complex biologics. For example, a monoclonal antibody submission may include hundreds of stability studies, viral clearance validation, and process characterization reports. To address this, companies use electronic submission systems like eCTD (electronic Common Technical Document). They also create cross-referenced summaries to help reviewers find key information. Another challenge is keeping up with changing guidelines; ICH updates regularly. Companies employ regulatory affairs specialists who track these changes. Early and frequent communication with regulators (e.g., through pre-submission meetings) can clarify expectations. Proper planning and document management are essential.
6. Compare structured models with unstructured models for describing cell growth.
Unstructured models, like the Monod equation, treat the cell as a single black box that grows based on one limiting substrate. They are simple and require few parameters. Structured models, however, include multiple internal variables (e.g., RNA, protein content) and reactions. They can predict more complex behaviors, such as lag phases or metabolic shifts. For example, an unstructured model cannot capture how a cell switches from growth to product formation when nutrients run low. A structured model can simulate this by tracking a key regulator molecule. The downside is that structured models need more data and are harder to calibrate. For many industrial applications, unstructured models are sufficient, but structured models are better for research.
7. Compare how erosion-based and diffusion-based drug delivery systems differ in their mechanical design.
In erosion-based systems, the drug is mixed in a material that slowly dissolves or breaks down in the body. As the material erodes, the drug is released. The mechanical strength of the material determines how fast it erodes: stronger materials last longer. In diffusion-based systems, the drug is trapped in a stable matrix and escapes by moving through pores or the material itself. The matrix does not break down; instead, the drug concentration gradient drives release. The key mechanical design difference is that erosion systems need to be strong enough to hold shape until they erode, while diffusion systems need to maintain their structure throughout release. Both require careful control of material properties to achieve the desired release profile.
8. Compare the downstream processing challenges for viral vectors versus monoclonal antibodies.
Viral vectors are much larger (20-100 nm) and more fragile than antibodies. They can be inactivated by shear stress or harsh conditions. Purification often uses ultracentrifugation or affinity chromatography, but yields are low (10-30%). Monoclonal antibodies are robust and can be purified with Protein A chromatography with high yield (>90%). Viral vectors also have a wide size distribution (full vs empty capsids), which is hard to separate. Antibodies have uniform size. Additionally, viral vector titers are measured in vector genomes per mL, while antibodies are in mg/mL. The cost of goods for viral vectors is much higher. Both require viral clearance steps, but for vectors, the product itself is a virus, complicating safety.
9. What is the main goal of scale-up in biochemical engineering?
The main goal of scale-up is to reproduce at a larger (industrial) scale the same product quality and yield that were achieved in a small lab experiment. This is challenging because factors like mixing, heat transfer, and oxygen supply do not change linearly with size. Engineers use key parameters such as power per volume, impeller tip speed, and oxygen transfer coefficient (kLa) to compare scales. A scale-down model is a small system designed to mimic the conditions of a large tank, so that problems can be studied cheaply. For example, a small stirred tank with the same kLa as a production reactor can be used to test how cells respond. The ultimate test is whether the large-scale process matches the small-scale performance.
10. How can transfer learning be used to improve a brain-computer interface (BCI) that controls a wheelchair, so it works for a new user quickly?
A BCI for a wheelchair typically uses EEG to detect commands like 'move forward' or 'turn left'. Training a new user can take hours. With transfer learning, you start with a model that was trained on many previous users. This model already knows common EEG patterns for these commands. The new user then performs a short calibration, maybe 10 minutes, where they imagine moving in different directions. You use this data to fine-tune the model. The fine-tuning adjusts the model to the user's specific brain rhythms. This reduces calibration time significantly. The BCI can then decode the user's intentions with high accuracy. Transfer learning also helps if the user's signals change over time due to fatigue or electrode shifts.
11. Compare CFD with experimental methods for studying mixing in bioreactors.
Experimental methods like tracer studies or particle image velocimetry (PIV) measure mixing in real reactors. They give accurate data but are expensive and time-consuming. CFD, on the other hand, is a virtual experiment that can test many conditions quickly and cheaply. However, CFD results depend on assumptions like turbulence models, which may not be perfect. For example, CFD might predict a mixing time of 10 seconds, but the real value could be 12 seconds. Experimental methods are the ground truth, but they are limited to existing reactors. CFD can explore new designs that haven't been built yet. Often, the best approach is to validate CFD with a few experiments, then use CFD for further optimization.
12. Give an example of how CFD can be used to minimize shear stress on shear-sensitive cells (e.g., mammalian cells).
Mammalian cells are easily damaged by high shear stress from impellers. CFD can simulate the flow field and calculate shear stress values throughout the bioreactor. Engineers can test different impeller designs, like a marine propeller vs. a Rushton turbine, in the simulation. For example, a low-shear impeller like a pitched-blade turbine may produce lower shear stress while still mixing well. CFD also helps optimize impeller speed: too slow causes poor mixing, too fast damages cells. By finding the sweet spot, CFD ensures cells stay healthy. The simulation can also predict how adding a surfactant or using a different sparger reduces shear. This saves time and avoids costly experiments with live cells.