# Biotech

## Senior Software Engineer - Machine Learning Platform

{% embed url="<https://www.freenome.com/careers/?gh_jid=4208229002>" %}

### **About the Role**

Freenome is looking for engineers to help us develop software to combat cancer and other age-related diseases. You will work as part of an interdisciplinary team of engineers and scientists building our internal machine learning platform. As an early team member, you’ll take the lead on major projects and collaborate actively with our world-class team of engineers, scientists, designers and product managers. You’ll design and build the systems used to power our Discovery Platform, the heart of Freenome’s experimental analyses. Since we’re a small team, you’ll also have an opportunity to determine the course of a broad range of projects and help shape the direction of the Engineering team at Freenome.

Freenome’s software systems provide the “nervous system” for the company by tracking sample analysis from start to finish, empowering and assisting lab technicians and scientists, and automating our growing collection of cancer-fighting robots. This nervous system is built using modern software development technologies and methodologies.

### **Responsibilities:**

* **Work closely with machine learning, bioinformatics, and product management teams to understand needs and then architect, roadmap, and lead development of the next phase** of Freenome’s discovery software platform
* Deep understanding of the role of the discovery platform for Freenome’s product development process and partnerships, and guidance of its purposeful evolution in support of these efforts
* **Own group charter and build a focused, collaborative engineering team**
* Develop and deploy reliable, maintainable, scalable, and fault-tolerant services
* **Guide and champion engineering hygiene and culture as a core part of the engineering backbone**

### **What We're Looking For:**

* **Ability to understand, plan, and develop** for key aspects of Freenome’s multi-analyze discovery analysis platform:
  * Heterogeneous data organization, accessibility, and modeling
  * Rapid, iterative, reproducible experimentation and analysis
  * Simple navigation to arbitrary states and checkpoints within the analysis tree
  * Clear interpretation and presentation of discovery insights in reports
* 5+ years experience as a part of a software development team successfully shipping a **machine learning, deep learning, data science, analytical, or similar platform**
* **Management or team lead experience**
* Knowledge of optimal methods for modern data storage systems, distributed systems, service architecture, and pipelining or workflow management.
* Track record of building distributed systems with service endpoints and distributed storage.
* **Understanding of, and practical experience with, statistical and machine learning methods**.
* Degree in computer science, mathematics, statistics, or related field or equivalent work experience
* Proficiency in a general-purpose programming language: **Python**, Java, **C, C++**, etc
* **Excellent written and verbal communication skills**
* **A mindful, transparent, and humane approach to your work and your interactions with others**

### **Nice to Haves:**

* Deep knowledge of **Python**
* PostgreSQL or similar relational database experience
* Experience with **Google Cloud Platform**, or another cloud computing service
* Domain-specific experience in computational biology, genomics or a related field
* Experience in **scientific parallel computing**
* Experience in **high-performance computing**, including SIMD or **GPU** performance optimization
* Experience with use of **automated regression testing, version control, and deployment systems**

## Onsite

1. A hands-on coding question. Writing some light classes, functions etc. No algorithms or anything tricky, just solving a problem with code.
2. An algorithmic whiteboard question. This requires no coding or coding knowledge. A computer science background will help but is not required. The problem is around the subject matter of bioinformatics analysis but does not require prior knowledge.
3. A infrastructure design whiteboard question. No coding. This is an exercise on the whiteboard to design the infrastructure and systems needed to create a web app that many users will use. This involves infrastructure, scaling, efficiency, and other considerations for a big user base.

## Problems we are looking to solve

### Pipelines

Tools: [Luigi](https://luigi.readthedocs.io/en/latest/), [Airflow](https://airflow.apache.org/index.html), [Ludwig](https://uber.github.io/ludwig/), [Kubeflow](https://github.com/kubeflow/), [Horizon](https://github.com/facebookresearch/Horizon)

{% embed url="<https://www.kubeflow.org/>" %}

{% embed url="<https://bionics.it/posts/luigi-tutorial>" %}

{% embed url="<https://airbnb.io/projects/airflow/>" %}

{% embed url="<https://eng.uber.com/introducing-ludwig/>" %}

{% embed url="<https://towardsdatascience.com/data-pipelines-luigi-airflow-everything-you-need-to-know-18dc741449b7>" %}

{% embed url="<https://static.googleusercontent.com/media/research.google.com/en//pubs/archive/35650.pdf>" %}

\
​<https://code.fb.com/core-data/introducing-fblearner-flow-facebook-s-ai-backbone/>​‌

​<https://code.fb.com/ml-applications/horizon/>​‌

​[https://www.reddit.com/r/bioinformatics/comments/5bu61o/anybody\_using\_luigi\_for\_their\_pipeline](https://www.reddit.com/r/bioinformatics/comments/5bu61o/anybody_using_luigi_for_their_pipelines/)

## Notes

> **Machine Learning** is a method of statistical learning where each instance in a dataset is **described** by a set of features or attributes. In contrast, the term “**Deep Learning**” is a method of statistical learning that **extracts** features or attributes from raw data. Deep Learning does this by utilising neural networks with many hidden layers, big data, and powerful computational resources. The terms seem somewhat interchangeable; however, with Deep Learning methods, the algorithm constructs representations of the data automatically. In contrast, **data representations are&#x20;*****hard-coded*****&#x20;as a set of features in&#x20;*****machine learning algorithms***, requiring further processes such as feature selection and extraction, (such as PCA).

### Toolkits

1. Airflow
2. Ludwig
3. Luigi
4. Horizon
5. Kubeflow

### Reading

{% embed url="<https://towardsdatascience.com/a-dummies-intro-to-bioinformatics-e8212ed7c09b>" %}

{% embed url="<https://www.sciencedaily.com/terms/bioinformatics.htm>" %}

{% embed url="<https://www.cancer.gov/research/nci-role/bioinformatics>" %}

{% embed url="<https://towardsdatascience.com/starting-off-in-bioinformatics-dna-nucleotides-and-strands-8c32515271a8>" %}

{% embed url="<https://towardsdatascience.com/dna-sequence-data-analysis-starting-off-in-bioinformatics-3dba4cea04f>" %}

> **Shotgun Sequencing**
>
> <img src="https://cdn-images-1.medium.com/max/800/1*7cs6g5s9gjNEygE9of8mtQ.png" alt="" data-size="original">
>
> Special machines, known as **sequencing machines** are used to extract short random DNA sequences from a particular genome we wish to determine (**target genome**). Current DNA sequencing technologies cannot read one whole genome at once. It reads small pieces of between 20 and 30000 bases, depending on the technology used. These short pieces are called **reads**. Special software are used to assemble these reads according to how they **overlap**, in order to generate continuous strings called **contigs**. These contigs can be the whole target genome itself, or parts of the genome (as shown in the above figure).
>
> The process of aligning and merging fragments from a longer DNA sequence, in order to reconstruct the original sequence is known as **Sequence Assembly**.
>
> In order to obtain the whole genome sequence, we may need to generate more and more random reads, until the contigs match to the target genome.

{% embed url="<https://towardsdatascience.com/pairwise-sequence-alignment-using-biopython-d1a9d0ba861f?source=user_profile---------5--------------------->" %}

{% embed url="<https://towardsdatascience.com/10-common-software-architectural-patterns-in-a-nutshell-a0b47a1e9013?source=user_profile---------4--------------------->" %}

{% embed url="<https://towardsdatascience.com/multiple-sequence-alignment-using-clustal-omega-and-t-coffee-3cc662b1ea82?source=user_profile---------2--------------------->" %}

{% embed url="<https://towardsdatascience.com/bioinformatics-workflow-management-systems-cc3edd97be79?source=user_profile---------1--------------------->" %}

{% embed url="<https://towardsdatascience.com/molecular-phylogenetics-using-bio-phylo-57ce27492ee9?source=user_profile---------0--------------------->" %}

### Pipeline(s)

#### **P**ipe-filter pattern <a href="#id-616b" id="id-616b"></a>

This pattern can be used to structure systems which produce and process a stream of data. Each processing step is enclosed within a **filter** component. Data to be processed is passed through **pipes**. These pipes can be used for buffering or for synchronization purposes.

**Usage**

* Compilers. The consecutive filters perform lexical analysis, parsing, semantic analysis, and code generation.
* Workflows in bioinformatics.

Pipe-filter pattern

![](https://cdn-images-1.medium.com/max/800/1*qikehZcDhhl_wWsqeI_nvg.png)

#### Blackboard pattern <a href="#ef1c" id="ef1c"></a>

This pattern is useful for problems for which no deterministic solution strategies are known. The blackboard pattern consists of 3 main components.

* **blackboard** — a structured global memory containing objects from the solution space
* **knowledge source** — specialized modules with their own representation
* **control component** — selects, configures and executes modules.

All the components have access to the blackboard. Components may produce new data objects that are added to the blackboard. Components look for particular kinds of data on the blackboard, and may find these by pattern matching with the existing knowledge source.

**Usage**

* Speech recognition
* Vehicle identification and tracking
* **`Protein structure identification`**
* Sonar signals interpretation.

Blackboard pattern

![](https://cdn-images-1.medium.com/max/800/1*ArbMx7A21I47llvwUTiSDg.png)

#### Comparison of Architectural Patterns <a href="#eaed" id="eaed"></a>

The table given below summarizes the pros and cons of each architectural pattern.

![](https://3501392451-files.gitbook.io/~/files/v0/b/gitbook-legacy-files/o/assets%2F-LLZ89zzVxrdnG1RG6CA%2F-Ldg5OHu5Xbev1u2VGsv%2F-LdgIOtUySMLzhZzZJYm%2Fimage.png?alt=media\&token=74f3d654-f071-428b-8f73-8eaa07ab5821)

### Books

1. ***Bioinformatics for Dummies*** by *Cedric Notredame and Jean-Michel Claverie*
2. ***Bioinformatics for Beginners: Genes, Genomes, Molecular Evolution, Databases and Analytical Tools*** by *Supratim Choudhuri*
3. ***Bioinformatics Programming in Python: A Practical Course for Beginners*** by *Ruediger-Marcus Flaig*
4. ***Bioinformatics Programming Using Python*** by *Mitchell L. Model*
5. **Python Programming for Biology: Bioinformatics and Beyond 1st Edition, Kindle Edition**

   by *Tim J. Stevens, Wayne Boucher*
6. [*http://biopython.org/DIST/docs/tutorial/Tutorial.html*](http://biopython.org/DIST/docs/tutorial/Tutorial.html)

### Software

{% embed url="<https://en.wikipedia.org/wiki/List_of_sequence_alignment_software>" %}

{% embed url="<http://www.bioinformatics.org/wiki/DNA_sequencing>" %}

{% embed url="<https://biopython.org/>" %}