Chou-Fasman

What is the Chou-Fasman method?

The Chou-Fasman method predicts protein secondary structure from amino acid sequence alone. Published by Peter Chou and Gerald Fasman in 1974, it was the first widely used algorithm to show that local amino acid composition carries enough information to predict helices, sheets, and turns.

The approach is statistical rather than physical. Chou and Fasman counted how often each amino acid appeared in helices, sheets, and turns of solved crystal structures, 29 proteins by 1978, and converted those frequencies into propensities. Glutamate has a helix propensity of 151 and a sheet propensity of 37: it turns up in helices half again as often as chance and in sheets about a third as often. A short set of rules then reads those numbers along the sequence.

The method sits well below modern predictors in accuracy: a 1983 benchmark found that no sequence-only method of its day got more than 56% of residues right in three states, and profile-based methods reached about 76% by 2001 (see the accuracy section below).

It remains useful as a teaching tool and as a fast, readable baseline. It needs no multiple sequence alignment, no neural network, and no GPU, and every call can be traced back to a few propensity values.

How to use Chou-Fasman online

Paste one or more sequences or upload a file. ProteinIQ applies the Chou-Fasman rules to each sequence and returns results within seconds.

Inputs

Output columns

The results table has one row per sequence.

The per-residue table has one row per residue, with the same numbers Fasman's own PREDICT program printed for each position.

The last three residues of a sequence start no tetrapeptide, so their averages and pt are blank. The per-residue table is left out when it would exceed 4 MB for one job, which is 30,000 to 60,000 residues depending on the length of the sequence names; the results table is always returned. Both tables can be downloaded as CSV or JSON, or copied.

How does Chou-Fasman work?

ProteinIQ uses the 1978 parameters for 29 proteins and the rules as Fasman summarized them in 1985, reprinted by Prevelige and Fasman in 1989. Chou and Fasman wrote these rules as guidelines for a person reading the numbers, not as a program, so an automatic run has to settle a few judgement calls. Those choices are stated below.

1. Helices

Four helix formers within six consecutive residues start a helix. The formers are the residues of class H or h: E, M, A, L, K, F, Q, W, I, and V. Asp and His are only weak formers and do not count. The helix then grows in both directions until it reaches a tetrapeptide with an average $P_\alpha$ below 100.

Proline may sit only in the first three residues of a helix. A helix is cut at any later proline, and the part from the proline onward is kept only if it contains a nucleus of its own. A helix survives when it is at least six residues long, its average $P_\alpha$ is above 103, and its average $P_\alpha$ is above its average $P_\beta$.

2. Sheets

Three sheet formers within five consecutive residues start a sheet: V, I, Y, F, W, L, C, T, Q, and M. The sheet grows until a tetrapeptide averages below 100 in $P_\beta$, and survives when its average $P_\beta$ is above 105 and above its average $P_\alpha$. Two regions that only touch end to end are judged separately, so a breaker tetrapeptide between two nuclei never ends up inside one long sheet.

3. Overlapping helices and sheets

Where a helix and a sheet that both survived overlap, the overlapping residues go to whichever structure has the higher average propensity over those residues; a tie goes to the helix. What remains of the losing region must pass its own rule again, so a helix cut below six residues becomes coil.

4. Turns

A tetrapeptide is a β-turn when three conditions hold. The product of its positional turn frequencies, $p_t = f_i \cdot f_{i+1} \cdot f_{i+2} \cdot f_{i+3}$, exceeds $0.75 \times 10^{-4}$. Its average $P_t$ exceeds 100. Its $P_t$ sum exceeds both its $P_\alpha$ and $P_\beta$ sums. When qualifying tetrapeptides overlap, the one with the higher $p_t$ is kept.

Turns are assigned even inside a predicted helix or sheet, as Prevelige and Fasman specify, so a turn can split a helix in two. Residues that end up in no helix, sheet, or turn are coil (C).

Understanding the results

The insulin B chain example shows what the rules get right and what they miss. For FVNQHLCGSHLVEALYLVCGERGFFYTPKT the prediction is EEEEEEETTTTCCCCCCCCTTTTEEETTTT. The turn at GERG (B20 to B23) and the short strand at FFY (B24 to B26) match the crystal structure. The helix from B9 to B19 is missed: that stretch nucleates both a helix and a sheet, but its average $P_\beta$ (116) beats its average $P_\alpha$ (108), so the helix is dropped. The sheet over B9 to B21 fails too, because its $P_\beta$ sum falls just short of its $P_\alpha$ sum (1,390 against 1,392).

The Structure string is the main result. Each character maps to one position in the input sequence:

• H (helix): the residue is in a predicted alpha-helix region
• E (sheet): the residue is in a predicted beta-sheet region
• T (turn): the residue is in a predicted β-turn
• C (coil): no regular structure predicted

If a prediction shows 0% for both helix and sheet, the sequence may be intrinsically disordered, or too short for any region to pass the rules: a helix needs at least 6 residues and a sheet at least 5, so shorter sequences can only return turns and coil.

The percentage and run-count columns are useful for comparing across sequences. A protein with 35% helix in 4 regions has a different architecture than one with 35% helix in 1 long region. When a call looks surprising, the per-residue table shows which propensities and tetrapeptide averages produced it.

Reading Pα, Pβ and Pt

A propensity of 100 means a residue type turns up in that state about as often as its overall frequency predicts. Glutamate (Pα 151) and methionine (145) favor helices, valine (Pβ 170) and isoleucine (160) favor sheets, and asparagine and glycine (both Pt 156) favor turns. The values come from the 29-protein table of Prevelige and Fasman (1989), so every residue of a given type gets the same three numbers.

The information is in how neighbors combine, not in any single row.

Helix class and Sheet class bin the same propensities. Only classes H and h count as formers when a nucleus is searched for; the other classes never start a region, and breakers pull down the window averages that decide where a region stops growing.

Seven residues (F, I, L, M, Q, V, W) are formers of both helices and sheets, which is why overlapping helix and sheet candidates are common.

Reading the tetrapeptide columns and turns

The rules never judge a single residue. ⟨Pα⟩ i to i+3 on a row is the mean Pα of that residue and the three after it, so the window starts at the row and is not centered on it. A region stops growing at the first window averaging below 100, then the whole region must clear 103 (helix) or 105 (sheet) and beat the other state. These cutoffs are the published rules (Prevelige and Fasman, 1989), not statistical significance levels.

Turn probability pt is the product of four position-specific frequencies, so it is a raw number near 0.0001 and not a probability between 0 and 1. A turn needs it above 0.000075, plus ⟨Pt⟩ above 100 and above both other averages. In the insulin B chain, the window GERG (B20 to B23) has ⟨Pα⟩ 90.75, ⟨Pβ⟩ 70.00, ⟨Pt⟩ 120.25 and pt of about 0.000092, so it passes. The window HLCG starting at B5 has ⟨Pt⟩ 107.25, above its other averages, but pt of about 0.000062 fails the product test, so it stays out of the turn calls.

Example: ubiquitin against its crystal structure

Human ubiquitin (76 residues, PDB 1UBQ) is a common test case. Its RCSB secondary-structure annotation has strands at residues 1 to 7, 10 to 17, 40 to 45, 48 to 50 and 64 to 72, a helix at 23 to 34 and a short 3-10 helix at 56 to 59. The tool returns this Structure string:

MQIFVKTLTGKTITLEVEPSDTIENVKAKIQDKEGIPPDQQRLIFAGKQLEDGRTLSDYNIQKESTLHLVLRLRGG
EEEEEEETTTTHHHHHHHTTTTEEEEHHHHHTTTTHTTTTHHHHHHHHHTTTTHCTTTTCCCCCEEEEEEEETTTT

The summary row reads Helix % 30.3, Sheet % 25.0, Turn % 36.8, Coil % 7.9, with Helix # 5, Sheet # 3 and Turn # 7.

The first strand (1 to 7) and the last (65 to 72) are found. The strand at 10 to 17 comes back as a helix at 12 to 18, the real helix starts as a strand call at 23 to 26, and residues 41 to 49, on two strands, are called helix. Counting turns as coil, 42 of 76 residues match the annotation (55.3%): 5 of 16 helix residues, 15 of 33 strand residues and 22 of 27 coil residues.

Helix avg is 4.6 here even though a helix needs six residues to survive. Turn calls are laid over helices afterward, so beside the two longer helices (7 and 9 residues) the string holds runs of 5, 1 and 1 residues. Treat one- or two-residue H or E runs as leftovers of that overlay, not as helices.

Common misreadings

Propensity is not structure. A high Pα says a residue type is common in helices in the 29-protein database, not that this residue sits in one. The helix called over ubiquitin residues 12 to 18 is locally helix-friendly sequence that the protein folds into a strand, because the tool sees only the sequence window.

T is not the same as C. Turns are assigned on top of helices and sheets and need only a four-residue window, so Turn % runs high (36.8% for ubiquitin), and a low Coil % does not mean the chain is ordered. To compare against a three-state reference such as the helix, strand, other grouping of DSSP, count T as coil, as in the 55.3% above.

The method has no tertiary context. It cannot see long-range contacts, other chains, membranes or ligands. Rost (2001) notes that stretches of identical residues can occupy different secondary-structure states depending on nonlocal conditions, which a window of four to six residues cannot capture.

How accurate is Chou-Fasman?

Accuracy is usually reported as Q3, the share of residues whose predicted state (helix, strand, other) matches the state assigned from the experimental structure. Kabsch and Sander (1983) tested the three most widely used sequence methods of their day, among them Chou-Fasman, on 62 proteins of known structure and found that none predicted more than 56% of residues correctly.

Methods using evolutionary information from related sequences later reached about 76% (Rost, 2001). The ubiquitin result above, 55.3% on one protein, is consistent with the 1983 ceiling, though one protein is an illustration and not a benchmark.

The original parameters were derived in the 1974 and 1978 papers from a small set of solved structures, and the 1978 table used here covers 29 proteins. A tool built on a few dozen structures from the 1970s should be read as a coarse local-propensity profile, not as a secondary-structure assignment.

When to use Chou-Fasman vs alternatives

Chou-Fasman is a sequence-only, statistics-only method. Several alternatives exist depending on what matters most:

• ESMFold and AlphaFold 2 predict full 3D structures from sequence. Running the resulting model through DSSP gives per-residue helix and strand states that use the whole chain as context. This is slower but far more informative, and the protein secondary structure prediction workflow compares it with Chou-Fasman. Judge such a model by its confidence scores first (see pLDDT).
• DSSP assigns secondary structure from a solved 3D structure (PDB file), not from sequence. If a crystal or cryo-EM structure exists, DSSP gives the standard assignment rather than a prediction.
• Protein Scale Profiler plots the Chou-Fasman Helix, Chou-Fasman Sheet and Chou-Fasman Turn scales as sliding-window profiles when the question is where propensity rises and falls, not which residues receive an H, E or T.

ProstT5 is not a drop-in alternative: ProteinIQ returns 3Di structural tokens from it, not H, E and C states.

Chou-Fasman's niche is speed and interpretability. Every prediction can be traced back to a specific propensity value and no black-box model is involved. This makes it well-suited for teaching the fundamentals of structure prediction, for quickly screening large sequence sets, and for situations where understanding why a prediction was made matters more than accuracy.

Sources

• Chou PY, Fasman GD. "Conformational parameters for amino acids in helical, β-sheet, and random coil regions calculated from proteins." Biochemistry 13, 211-222 (1974). doi:10.1021/bi00699a001
• Chou PY, Fasman GD. "Prediction of protein conformation." Biochemistry 13, 222-245 (1974). doi:10.1021/bi00699a002
• Chou PY, Fasman GD. "Prediction of the secondary structure of proteins from their amino acid sequence." Advances in Enzymology and Related Areas of Molecular Biology 47, 45-148 (1978). doi:10.1002/9780470122921.ch2
• Prevelige P Jr, Fasman GD. "Chou-Fasman prediction of the secondary structure of proteins." In Prediction of Protein Structure and the Principles of Protein Conformation, 391-416 (1989). doi:10.1007/978-1-4613-1571-1_9
• Kabsch W, Sander C. "How good are predictions of protein secondary structure?" FEBS Letters 155, 179-182 (1983). doi:10.1016/0014-5793(82)80597-8
• Rost B. "Review: protein secondary structure prediction continues to rise." Journal of Structural Biology 134, 204-218 (2001). doi:10.1006/jsbi.2001.4336
• Vijay-Kumar S, Bugg CE, Cook WJ. "Structure of ubiquitin refined at 1.8 Å resolution." Journal of Molecular Biology 194, 531-544 (1987). doi:10.1016/0022-2836(87)90679-6