← Case studies
Drug Discovery

Identifying promising compounds for drug discovery

A drug discovery researcher uses Deep Netts networks to sort through thousands of screening measurements and find the compounds that change a protein the way a medicine should – in pure Java, from data the lab already has.

Pure JVM
no native libraries to install
No QSAR input
learns from properties already in the compound library
10,000s
of data points from one screening session
2024
joint paper with Deep Netts in Scientific Reports

At a glance

Who
Oscar Bastidas, Ph.D., a drug discovery researcher and Java developer in Minnesota – and, for the second project, Zoran Ševarac of Deep Netts.
The task
Find, among thousands of tested compounds, the few that change a protein the way a medicine should.
The problem
Screening produces tens of thousands of data points per session; common tools need native libraries or structural data many compounds lack.
The solution
Deep Netts feed-forward networks that learn from properties already in the compound library – in pure Java.

See it in motion

Both projects in one scene – the screening analysis with Deep Netts and the spike protein study. Play it, pause it, or click a step.

0
Frames
0
Top frequency
0
Screened
0
Hits
Step 1 / 5
Protein dynamics
Simulation

The study simulated the SARS-CoV-2 spike protein (PDB 6VSB) in water for 200 ns, one frame per nanosecond. From every frame the backbone angles φ and ψ of each residue are read – here φ of one residue.

200 ns
simulated, one frame per ns (PDB 6VSB)
23–63 MHz
favoured φ/ψ oscillations, top 42.97 MHz
Pure JVM
deep learning with no native libraries

Sources: O. H. Bastidas & Z. Sevarac, Time dependent dihedral angle oscillations of the spike protein of SARS-CoV-2, Scientific Reports (2024); O. Bastidas, Ph.D., Deep Learning in Drug Discovery (2021). The protein fragment, compounds and values drawn are illustrative.

Thousands of measurements, a few good compounds

Proteins do almost all the work in a living cell. Each one is a chain of amino acids folded into a 3D shape, and the shape decides what it does. Many medicines are small molecules that bind to a protein and change that shape.

Finding such molecules is slow and expensive: bringing a candidate to patients can take up to a decade, and up to 95% of candidates fail on the way. Lab work is one of the main bottlenecks – but it also produces a lot of data, and that is where machine learning helps.

High-throughput screening

Robots test a whole compound library – thousands of small molecules – against one protein. Each tiny well on a plate gets the protein, one compound and a readout that gives a signal.

In this screen, a signal inside a certain range means the compound changed the protein's shape in the desired way: it is a hit. One session produces tens of thousands of data points.

  • Good for machine learning: the data are already curated – controls and conditions are part of the design – and come in one consistent format.
  • Hard for people: labs with a purely experimental focus can feel overwhelmed by so much data.

negative control positive control darker = stronger signal in the target range = a hit (26 here)

A 384-well plate: one compound per well, control wells at the edges. Layout and values illustrative – the real screening data are not published.

Where Deep Netts fits in

The goal is to find which properties of a compound lead to a good signal – and point the lab to the molecules most likely to work. Deep Netts is the engine of that analysis.

  1. Compound libraryEach compound comes with physical and chemical properties.
  2. ScreeningThe lab measures a signal for every compound.
  3. Deep NettsTrainA Deep Netts network learns: properties → signal.
  4. Deep NettsTuneHyperparameters fitted to the data for the best accuracy.
  5. HitsWhich compounds land in the target range, and which properties drive it.

The data, as the network sees it

CompoundMol. weightlogPH-bond donorsPolar surfaceSignalHit
C-0412312.42.8164.10.71✓
C-0413198.20.9388.50.12–
C-0414451.64.3041.70.33–
C-0415287.32.1272.90.66✓
…………………
Illustrative rows. The article speaks of physical and chemical properties provided with the compound library; the real columns and values are not published.

Each row is one example: the properties are the inputs, the measured signal is the answer the network learns to predict. Once trained, it can estimate the signal of compounds it has never seen – and show which properties matter.

Sketch: a property-to-signal network in the Deep Netts API
TabularDataSet<MLDataItem> data =        DataSets.readCsv("screen.csv", 4, 1, true, ",");   // 4 properties → 1 signalTrainTestSplit split = data.trainTestSplit(0.8); FeedForwardNetwork model = FeedForwardNetwork.builder()        .addInputLayer(4)        .addFullyConnectedLayer(16, ActivationType.TANH)        .addFullyConnectedLayer(8, ActivationType.TANH)        .addOutputLayer(1, ActivationType.LINEAR)          // predicted signal        .lossFunction(LossType.MEAN_SQUARED_ERROR)        .build(); model.getTrainer()                                          // hyperparameters, tuned to the data        .setLearningRate(0.001f)        .setStopEpochs(1000);model.train(split.getTrainingSet()); float signal = model.predict(properties)[0];boolean hit = signal >= LOW && signal <= HIGH;

Written like the regression examples in the Deep Netts cookbook – not the researcher's code, which is not published.

No structural data needed

The usual computational approach, QSAR, predicts a molecule's activity from its chemical structure. That needs the structure, usually as a SMILES code – a line of text spelling out atoms and bonds. Aspirin, for example:

CC(=O)OC1=CC=CC=C1C(=O)O

For thousands of compounds in a real library, getting that structure can be very hard: many are not in public databases such as ChEMBL, or the lab lists them under in-house names.

Learning from what the lab already has.With Deep Netts the analysis learns from the physical and chemical properties that come with the compound library – no structural input. In the researcher's words, an invaluable supplement, or even an alternative, to QSAR.

He also points to an idea of his own: recording which hyperparameters work best for a given experimental method could itself say something about that method – for example, whether the best values change when the measurement physics changes but the protein and the compounds stay the same.

Why pure Java

  • No native libraries. Many deep learning tools rely on native code built for specific hardware and operating systems; on unusual machines they can be next to impossible to run. Deep Netts runs purely on the JVM, so the question disappears.
  • His own language. As a Java developer he could use it right away, instead of spending weeks or months learning another language just for a different toolkit.
  • Control over the model. Hyperparameters can be tuned to the nature of the data for the best prediction accuracy.
  • Tools and help. Models can be built and analysed in the Deep Netts visual development environment, and he points to prompt, knowledgeable support from the Deep Netts team – especially with the enterprise edition. A free community edition also exists.

A second project: how the spike protein moves

The spike protein sits on the surface of the SARS-CoV-2 virus and grabs human cells, which makes it one of the most important drug targets. Its structure is well known; how its backbone moves over time was not.

No neural network here.This is a physics-style analysis of a protein simulation, published by the researcher with Zoran Ševarac of Deep Netts and done in Java. It is shown here as their joint work, separate from the screening project above.

Wobbling angles

A protein's backbone repeats the same three atoms – N, Cα, C – for every amino acid. The chain can twist around two bonds per residue: the angle φ (phi) around N–Cα and ψ (psi) around Cα–C. Together they decide the local shape.

The motion came from the authors' earlier simulation (PLoS ONE, 2020): the full spike protein in water, starting from structure 6VSB, run for 200 nanoseconds with one snapshot per nanosecond – about 200 values for every angle.

Rside chainφψCNCαCNCαone residue
The protein backbone and its two twist angles per residue. Schematic.

From wobble to frequencies

A Fourier transform splits a wobbling line into the simple rhythms it is made of. The authors ran the fast Fourier transform of the JDSP library, in Java, on every φ and ψ series and counted how often each frequency appeared – per amino acid, per chain and for the whole protein.

One angle (φ) over 200 ns, one value per ns

Fourier
transform →

The rhythms it is made of (MHz)

favoured 23–63 MHz27.343.062.50 … 100 MHz
A messy wobble is a few simple rhythms added together. 43 MHz is one full back-and-forth about every 23 ns. Curve and bars illustrative, frequencies from the paper.

What they found

  • Favoured frequencies: the oscillations cluster between 23 and 63 MHz. The most common is 42.969 MHz, then 62.500 and 27.344 MHz – for both φ and ψ.
  • A pattern like music: the counts alternate high and low along the spectrum, like a harmonic series – the protein is more “in tune” with some frequencies than others.
  • Gaps: for rarer amino acids (histidine, methionine, tryptophan) some frequencies never appear in a chain, and the gaps differ between chains.
  • Known sites differ in time, not frequency: residues known as allosteric sites or druggable pockets show distinctive patterns over time – brief excursions, sudden shifts – but no special frequencies.
ResidueKnown forSource in the paper
Lys369, Phe377Allosteric site (luteolin binding)Alvarado et al.
Glu406, Asn439, Lys417, Asn501Allosteric communication centresVerkhivker et al.
Phe329, Phe515Motion of the receptor-binding domain when binding ACE2Xue et al.
Residues from the literature whose angle time courses the paper shows as distinctive.

From similar time profiles the authors propose further residues to study – for example Asn211, Tyr421 and Thr124 (φ) or Leu858, Lys444 and Gln493 (ψ) – as possible starting points for vaccine or drug design, and stress that they need confirmation.

Limits

  • One 200 ns simulation of the unmutated protein, with one chain of 6VSB in the “up” position.
  • What frequencies in the MHz range mean physically is not yet established.
  • The proposed residues need dedicated work to confirm a link to allostery or pockets, and proteins of other sizes must be studied.

The authors present the work as a reference for future studies, not a replacement for existing methods. The data are public on the Open Science Framework.

Sources

Glossary (18 terms)
Protein, residue
A protein is a chain of amino acids; one amino acid in the chain is a residue.
Conformational change
A change in a protein's 3D shape – what a medicine often aims to cause.
Compound library
A lab's collection of small molecules to test, each with listed properties.
High-throughput screening (HTS)
Robots testing thousands of compounds against a protein in one session.
Signal, hit
The measured readout per compound; a hit is a compound whose signal falls in the target range.
Controls
Wells with known outcomes (nothing / full effect) that calibrate a plate.
QSAR
Quantitative structure–activity relationship: predicting activity from chemical structure.
SMILES
A text code for a molecule's structure, e.g. aspirin = CC(=O)OC1=CC=CC=C1C(=O)O.
ChEMBL
A large public database of bioactive molecules.
Hyperparameters
Settings chosen before training: layer sizes, learning rate, number of epochs…
Native libraries
Machine code built for one operating system and processor – what pure Java avoids.
Spike (S) protein, 6VSB
The SARS-CoV-2 surface protein that binds human cells; 6VSB is its structure in the Protein Data Bank.
Backbone, φ and ψ
The repeating N–Cα–C chain; φ and ψ are its two twist angles per residue.
Molecular dynamics (MD)
A simulation that moves every atom step by step under physical forces.
ns, MHz
A nanosecond is a billionth of a second; 1 MHz is a million oscillations per second.
Fourier transform (FFT)
Maths that splits a signal into the frequencies it is made of; FFT is the fast algorithm.
Allosteric site
A place on a protein, away from its active site, where binding still changes how it works.
Druggable pocket
A cavity on a protein's surface where a small molecule can bind.

Keep reading