Category Archives: Machine Learning

Issues with graph neural networks: the cracks are where the light shines through

Deep convolutional neural networks have lead to astonishing breakthroughs in the area of computer vision in recent years. The reason for the extraordinary performance of convolutional architectures in the image domain is their strong ability to extract informative high-level features from visual data. For prediction tasks on images, this has lead to superhuman performance in a variety of applications and to an almost universal shift from classical feature engineering to differentiable feature learning.

Unfortunately, the picture is not quite as rosy yet in the area of molecular machine learning. Feature learning techniques which operate directly on raw molecular graphs without intermediate feature-engineering steps have only emerged in the last few years in the form of graph neural networks (GNNs). GNNs, however, still have not managed to definitively outcompete and replace more classical non-differentiable molecular representation methods such as extended-connectivity fingerprints (ECFPs). There is an increasing awareness in the computational chemistry community that GNNs have not quite lived up to the initial hype and still suffer from a number of technical limitations.

Continue reading →

How to interact with small molecules in Jupyter Notebooks

The combination of Python and the cheminformatics toolkit RDKit has opened up so many ways to explore chemistry on a computer. Jupyter — named for the three languages, Julia, Python, and R — ties interactivity and visualization together, creating wonderful environments (Notebooks and JupyterLab) to carry out, share and reproduce research, including:

“data cleaning and transformation, numerical simulation, statistical modeling, data visualization, machine learning, and much more.”

—https://jupyter.org

At this year’s annual RDKit UGM (User Group Meeting), Cédric Bouysset shared a tutorial explaining how to create a grid of molecules that you can interact with, using his “mols2grid“:

Continue reading →

AlphaFold 2 is here: what’s behind the structure prediction miracle

Nature has now released that AlphaFold 2 paper, after eight long months of waiting. The main text reports more or less what we have known for nearly a year, with some added tidbits, although it is accompanied by a painstaking description of the architecture in the supplementary information. Perhaps more importantly, the authors have released the entirety of the code, including all details to run the pipeline, on Github. And there is no small print this time: you can run inference on any protein (I’ve checked!).

Have you not heard the news? Let me refresh your memory. In November 2020, a team of AI scientists from Google DeepMind  indisputably won the 14th Critical Assessment of Structural Prediction competition, a biennial blind test where computational biologists try to predict the structure of several proteins whose structure has been determined experimentally but not publicly released. Their results were so astounding, and the problem so central to biology, that it took the entire world by surprise and left an entire discipline, computational biology, wondering what had just happened.

Continue reading →

Out-of-distribution generalisation and scaffold splitting in molecular property prediction

The ability to successfully apply previously acquired knowledge to novel and unfamiliar situations is one of the main hallmarks of successful learning and general intelligence. This capability to effectively generalise is amongst the most desirable properties a prediction model (or a mind, for that matter) can have.

In supervised machine learning, the standard way to evaluate the generalisation power of a prediction model for a given task is to randomly split the whole available data set X into two sets – a training set X_{\text{train}} and a test set X_{\text{test}}. The model is then subsequently trained on the examples in the training set X_{\text{train}} and afterwards its prediction abilities are measured on the untouched examples in the test set X_{\text{test}} via a suitable performance metric.

Since in this scenario the model has never seen any of the examples in X_{\text{test}} during training, its performance on X_{\text{test}} must be indicative of its performance on novel data X_{\text{new}} which it will encounter in the future. Right?

Continue reading →

CAML: Courses in Applied Machine Learning

*Shameless self-promotion klaxon!! Have a look at my new website!*

I’m excited to share a project I’ve been working on for the past few months! One of the biggest challenges of working on an interdisciplinary research project is getting to grips with the core principles of the disciplines which you don’t have much formal training in. For me, that means learning the basics of Medicinal Chemistry and Structural Biology so that when someone mentions pi-stacking I don’t think they’re talking about the logistics of managing a bakery; for people coming from Bio/Chem backgrounds it can mean understanding the Maths and Statistics necessary to make sense of the different algorithms which are central to their work.

Continue reading →

Can few-shot language models perform bioinformatics tasks?

In 2019, I tried my hand at using large language models, specifically GPT-2, for text generation. In that blogpost, I used Hansard files to fine-tune the public release of GPT-2 to generate speeches by several speakers in the House of Commons (link).

In 2020, OpenAI released GPT-3, their new and improved text generation model (paper), which uses a whopping 175 billion parameters (as opposed to its predecessor’s 1.5 billion) and not only proved to be capable of state of the art performance on common text prediction benchmarks, but also generated a considerable amount of interest in the news media:

Continue reading →

Bioinformatics Hackathon Reflection

A week ago I participated in Copenhagen Bioinformatics Hackathon 2021, a hackathon focusing on machine learning and proteins, as a mentor for a challenge proposed by our group. The whole experience was fun, but I am also sitting here contemplating over a lot of things I wish I had done differently. For this blog text, I therefore want to highlight two changes which I believe would have greatly improved my challenge and which can hopefully also work as an inspiration for others presenting a hackathon challenge. 

Going into this event I had some experience from a few hackathons I had previously attended. Based on this, I wanted to create a challenge containing two parts. First, a simple task which everyone would be able to create a solution for, and second, a more challenging addition to the first task for more experienced participants. I decided to go with the challenge of predicting which heavy and light chains can form a pair, where the additional challenge was to try to visualize which residues were relevant for this interaction. Together with OAS containing a really nice positive dataset of paired chains, I thought this was going to be an amazing challenge, but as soon as the event began I started seeing the flaws of the challenge.

Continue reading →

Is bigger better?

Recent work in Natural Language Processing (NLP) indicates that the bigger your model is, the better performance you will get. In a paper by Kaplan, Jared, et al., they show that loss scales as a power-law with model size, dataset size, and the amount of compute used for training.

Kaplan, Jared, et al. “Scaling laws for neural language models.” arXiv preprint arXiv:2001.08361 (2020).
Continue reading →

Better understanding of correlation

Although correlation is often used as the linear relationship between two sets of points, I will in the following text use it more broadly to mean any relationship between two sets of points.

You have tasked yourself with finding the correlation between the different features in your dataset. Your purpose could be to remove highly correlated features or just improve your understanding of your data. Nonetheless, calculating and using the Pearson Correlation Coefficient (PCC) or the Spearman’s rank Correlation Coefficient (SCC) to get an overview of the correlations might be the first thing that comes to your mind.

Unfortunately, both of these are limited to linear (PCC) or monotonic (SCC) relationships. In datasets with many and complex features, many of them will be highly correlated, just not linearly (or monotonic). Instead these correlations can be non-linear which, as seen in the third row in the below figure, does not get detected with PCC.

Figure: PCC of different sets of x and y points. https://en.wikipedia.org/wiki/Correlation_and_dependence
Continue reading →

The Coronavirus Antibody Database: 10 months on, 10x the data!

Back in May 2020, we released the Coronavirus Antibody Database (‘CoV-AbDab’) to capture molecular information on existing coronavirus-binding antibodies, and to track what we anticipated would be a boon of data on antibodies able to bind SARS-CoV-2. At the time, we had found around 300 relevant antibody sequences and a handful of solved crystal structures, most of which were characterised shortly after the SARS-CoV epidemic of 2003. We had no idea just how many SARS-CoV-2 binding antibody sequences would come to be released into the public domain…

10 months later (2nd March 2021), we now have tracked 2,673 coronavirus-binding antibodies, ~95% with full Fv sequence information and ~5% with solved structures. These datapoints originate from 100s of independent studies reported in either the academic literature or patent filings.

The entire contents CoV-AbDab database as of 2nd March 2021.
Continue reading →