BASSA model explorer
Overview Models
Models Recipes Your families Network Map Compounds
New here? Three steps to read this page 1 Read What drives the predictions : one sentence and a bar, what the model says makes a better compound.2 Click a point in the plot: the molecule and the tabs Why this prediction and How to make it better explain that compound.3 The line over the plot, How well it predicts new compounds , is the number to quote. Click a column of the table to look at one model.Got it
Rows by frequency goes with selection families ?
Links strongest all off ?
Pin
Show ensemble model the selected model Design Threshold analysis Export Save picture Copy model Predictions as CSV
Compare models full explorer ↗
How to read this page
?
31 18 1 2 3 A
B .61 1 2 3 4 5 6 7 8
What drives the predictions (over the table): one sentence and a bar, what the ensemble model says makes a better compound and how much each family counts. Point at a segment to find it in the table, click to see it on the molecule.A column is a model (or models that differ only within families). Its bar, and the number over it, is how often BASSA visited it, in % of its steps; the most visited stands first. Click one.A coloured block is a family : descriptors that stand in for one another. Its letter and colour are the same everywhere.A dot is a descriptor in the model. A hollow dot swaps another of the family in.The plot : every compound as BASSA predicted it while it was left out. The switch over it, ensemble model / the selected model , chooses what the plot, the molecule and the weights show; ranking puts the compounds in the model's order, to see whether it finds the best. Click a point to see that compound.Under the plot , for a column you clicked: how much each descriptor weighs and whether it raises ▲ or lowers ▼ the response, and its equation. For the ensemble model, the bar over the table says what drives it.How well it predicts new compounds (the line over the plot): the number to quote. "How it was checked" opens the deals of the folds (blue dots) and the same with the responses shuffled (grey).The molecule shows what the model's descriptors measure.Why this prediction (a tab beside the molecule): what each family adds for the compound in hand. How to make it better : the families that could gain most. Every small ? on the page opens a drawing of that number. The tab Your families lets you make the families yourself.
How this case is split
The same, in words
The table (Models). A column is a model BASSA visited; a row is a descriptor. Over each column is how often the sampler visited it, in % of its steps, which is also the height of its bar; the most visited stand first. That is BASSA's own ranking. Click a column to draw its model on the molecule; the arrow keys move to the next one. Under each column stand its Q² and R² as a fixed model: it was chosen on all compounds, so they flatter; the number to quote is the one over the plot.Families and ideas. A family is a group of descriptors that stand in for one another: their columns are correlated and the sampler does not put them in the same model more than chance. Every family, and every descriptor on its own, is one idea with a letter and a colour (A is the most used). A family is a coloured block; it opens by itself when the model in hand uses it, and its heading folds it. It shows its four most used descriptors, with the rest behind "more". The "?" at the end of a row says what the descriptor is and why it is in its family.The dots. One dot is one descriptor of the model. A hollow dot is a descriptor that another model of the same column uses in its place: click it to swap it in. A folded family carries its dot on the heading. The lettered squares under a column name the ideas the model combines, and a line joins its dots.A column is. "A combination of families" counts together the models that differ only in which descriptor of a family they use, and leaves out combinations in less than 1 % of sampled models. "A single model" shows every model on its own.The mark under a column (advanced; ! or ‼, none when no check flags the model) counts the checks that flag it: a fit that flatters (R² far above the leave-one-out Q²), a model that works as well with every descriptor cut to high or low, and a leading descriptor that drops out when single compounds are left out. Click the mark, or "show them" under the plot, for the list with its numbers. An orange pill beside the case name says when the whole case is not above chance.The blue box over the plot is BASSA's ensemble model, validated with the sampler rerun in every fold, so the choice of descriptors is repeated without the held-out compounds: the number to quote. Q² is the mean over 20 deals of the compounds into 5 folds; the blue dots are the deals, the grey dots the same procedure on shuffled responses (the chance line), and the line under them says how many deals beat every shuffle. Then the share of measured values inside the 90 % interval, and the compound furthest off with Q² without it. The plot shows the same predictions. Each number has a "?" that explains it.Your families (tab). Group the descriptors yourself. Pick one: the page lists the descriptors that look like it, by how their values move together (r) and by how the sampler treats the pair (in one model more than chance, or one or the other), draws the two against each other and on the molecule, and maps every pair. Each of your families says how alike its members are and where the page's rule differs. "Use my families in the table" makes the table and the recipes follow them; the validation does not change. Your families are kept in this browser; "Save as CSV" writes them out.The equation under the model's table is the model on normalized descriptors (the same in the descriptors' own units is folded under it). "Weight" is the change in the response for one standard deviation of the descriptor.Show: ensemble model / the column you clicked : the plot, the molecule and the weights follow this one switch. The line over the plot always gives the ensemble model's numbers, the ones to quote.On the molecule. The descriptors of the model are drawn with their values. With a dipole term, "dipole arrow ×" sets the length of the arrows; with a Sterimol term, "Sterimol drawing" adds the B1 plane, the angle θ, all widths, the end-on map or the measured atoms.Recipes show the same combinations as cards. Click a descriptor to draw the model of that card that contains it; the arrows step through the card's models. "Together" is the card's models averaged, weighted by visits. What "advanced" adds
Views. Network and Map draw which descriptors share models (green) or stand in for each other (red). Compounds shows how the ten most used descriptors move when each compound is left out and the sampler rerun.The table. Over each column its 5-fold Q² (blue) and R² (grey) as a fixed model; sorting the models by strength, Q² or R²; filters on size, Q², R² and visits; rows by frequency or by what goes with a selected descriptor; the links between descriptors as arcs; pins. Sorting and filtering by Q² use the leave-one-out value, not the 5-fold one over the bars.Beside each descriptor. In how many of the leave-one-out reruns it stays among the three most used, and the range of its share. On the bars: black dashes are what chance gives, the dark line the spread between chains, the blue line the range over those reruns. Family headings name the ideas they most often share a model with, and "on molecule" draws a family's descriptors together.Over the plot. The model in hand kept fixed: R², 5-fold Q², MAE, also when it is not plotted; Compare models keeps one model in the plot as hollow marks beside a second one; the checks on the model under the plot.Under the plot. Where each descriptor's share goes when single compounds are left out; Threshold analysis , which turns one descriptor into a yes-or-no rule; names on the points.An orange "!" beside a descriptor means one compound sits more than five standard deviations from all the others in that column.The plot : click a point to look at the compound, shift-click to compare, right-click to hold it out so that the model is fitted without it. Diamonds are external compounds, never fitted. Rerun BASSA without them runs the sampler again on the rest. A model picked by eye from this page has been selected on these compounds. Quote the ensemble model's number, not the bar of a column.
All cases
Cases validated over all the deals first, best first, by BASSA's ensemble model (cases with one deal, or none, follow), validated with the sampler rerun in every fold: Q² is the mean over the deals of the compounds into 5 folds, with its range and the typical miss. The strip shows every deal (blue) against the same procedure on shuffled responses (grey). Against chance: above when the mean beats every shuffle and at least 90 % of the deals do; weak when fewer deals do. The interval column is the share of held-out values inside the 90 % interval ("not shown" where the first-fold check failed). Click a row to open the case.
The links beside the descriptor names
Each arc joins two descriptors and says how they behave across all the models BASSA visited.
Green, solid: they appear together. The two are in the same model more often than two unrelated descriptors would be. They carry different information and BASSA wants both.Red, dashed: they replace each other. The two columns are strongly correlated and are rarely in the same model. They carry the same information under two names; a model takes one or the other.Thicker means a stronger effect. No arc means nothing either way.Strongest draws only the six clearest arcs of each kind, so the column stays readable. Point at a descriptor , or select it, and all of its links appear. All draws everything; off hides them.Click an arc to select the two descriptors it joins.
The same information is in the tint of the rows once a descriptor is selected, and as pictures under Recipes, Network and Map.
Rows by: three ways to order the descriptors
Frequency. The plain ranking. The descriptor BASSA used most is on top; the grey bar is the share of all visited models that contain it. Use it to see what matters at all.Goes with selection. Click a descriptor first. The others are then ordered by how they behave beside it : at the top, the descriptors that appear in the same models much more often than usual (its partners, tinted green); at the bottom, the ones that almost never share a model with it (tinted red, usually because they say the same thing). The small bar in the "beside" column runs from the descriptor's usual share (the tick) to its share next to your selection. Use it to answer "what does BASSA combine this with?".Families. A family is a group of descriptors that stand in for one another: by default the sampler decides: two descriptors are one family when BASSA clearly takes one or the other (they share a model far less often than unrelated pairs do) and their columns are correlated at 0.3 or more. The switch over the table (View) groups by correlation instead: 0.7 or more with the most used one, or 0.5 or more when the sampler takes one or the other. Either way the sampler must not put the two in the same model more than chance. A descriptor that is correlated with a family but used together with it is its own idea. This order puts each family together, marks it with a coloured stripe at the left, and draws a line between families. BASSA usually takes one member of a family per model, so a family is the unit to read, not the single descriptor: three descriptors at 20 % each in one family are one idea at about 60 %. Use it to see how many different ideas the models draw on.
Say what this column measures
It is Sterimol B1 on an axis Sterimol B5 on an axis Sterimol L on an axis loc B5 on an axis a Sterimol angle on an axis
NBO charge on an atom Hirshfeld charge on an atom CM5 charge on an atom a length, angle or dihedral (number the atoms) Atom numbers
New case from your own table
Table (CSV) Response column Run length quick look: 4,000 iterations, one chain full run: 20,000 iterations, three chains Structures folder
First column: the compound name. One file per compound in the folder, named like the compound (L1.feather, or a longer name that contains L1 as a word). Feathers and logs bring charges and the dipole; xyz brings the geometry only.
Descriptors are drawn on the molecule when their column names follow the DescriPyTor platform (B1_6-4, loc_B5_6-4, nbo_atom_5, mu_y[C4->N5|C6], 0_angle_[5, 6, 7]); other columns are used everywhere else.
The model
The clipboard is not available here; copy from the box.